Why Federated Learning Now?

In marketing, data is a competitive edge. The more audience signals, campaign performance data, and consumer behavior a Machine Learning (ML) model can learn from, the sharper its predictions and the greater its business impact. Across the marketing ecosystem — spanning WPP’s agencies, their clients, and external partners — the collective wealth of consumer and brand data is enormous. The potential to train models across these combined assets could unlock transformative capabilities: better audience targeting, smarter media spend, and faster creative optimization at a global scale.

But here’s the challenge. While WPP maintains centralized access to its own internal data assets, much of the most valuable complementary data resides with clients and partners and, for good reasons, it can never leave their walls. Client contracts, privacy regulations like GDPR, and the sheer sensitivity of consumer-level data make cross-organizational data pooling a non-starter. The result? Models are trained without the full picture, and transformative insights that could emerge from combining datasets across organizational boundaries remain permanently out of reach.

The traditional solution, centralized ML, pools raw data from multiple sources into a single cloud to train a global model. But uploading terabytes of sensitive data to a central server creates severe network latency and exposes collaborators to data breaches and potential violations of privacy regulations.

Distributed ML methods attempted to address this by splitting training across local worker nodes. While this reduces latency and avoids centralizing raw data, these architectures were designed for internal computing clusters, not secure collaboration between independent companies. Without cross-organization coordination, models remain isolated, difficult to align, and impossible to improve as a unified whole.

Problem: The Collaboration vs. Privacy Bottleneck

Organizations are effectively trapped between two undesirable choices: compromise sensitive data through centralization, or settle for underperforming, isolated models through distribution. Neither architecture allows multiple companies to collaboratively train a shared, high-performing model while keeping private data strictly localized.

Federated Learning (FL) offers a way out of this dilemma by bringing the model to the data, rather than the other way around. To understand why this shift matters, let’s look at how FL actually works under the hood.

How Federated Learning Works

Figure 1: Overview of the Federated Learning communication cycle between a central node and distributed client nodes.

Federated Learning (FL) enables multiple organizations to collaboratively train a shared model without ever centralizing raw data. Instead of moving data to the model, FL brings the model to the data. Training proceeds through iterative rounds:

1. A central server sends the current global model to all participating clients (Blue arrow).
2. Each client trains the model locally on its own private data (Green arrow).
3. Clients send back only their model updates, never the underlying data (Pink arrow).
4. The server aggregates these updates into an improved global model and starts the next round.

Throughout this process, raw data never leaves its source. Only learned model representations are exchanged across the network.

The Multimodal Challenge

While the above privacy-preserving framework is valuable in its own right, modern marketing data adds another layer of complexity. Organizations do not just work with spreadsheets and numbers. They work with images, video, text, audio, and structured data, often all at once. A single campaign might involve visual brand assets, ad copy, audience segments, and performance metrics across channels. Training models that can reason across these different data types, known as multimodal learning, is already one of the most demanding challenges in ML.

Now combine that with the constraints of federated learning. Each client may hold different combinations of modalities, in different formats and volumes. One partner might contribute rich visual data, another mostly text and tabular records. Coordinating a single global model that learns effectively from this fragmented, heterogeneous landscape, without ever seeing the raw data, pushes the problem to a new level of complexity.

This is precisely what makes the intersection of FL and multimodal learning so important, and so hard. If it can be made to work, it unlocks collaborative intelligence across organizations at a scale that neither approach could achieve alone.

Our Objective: Can Federated Learning Deliver?

The promise of FL is compelling, but before investing in real-world deployment, we need to answer a fundamental question:

Does federated learning actually work well enough on multimodal marketing data to justify the tradeoff?

Centralized training will always have an inherent advantage, it sees all the data at once. The question is not whether FL can beat centralized performance, but whether it can get close enough to make the privacy and collaboration benefits worthwhile. And beyond raw performance, we need to understand how FL behaves under realistic stress conditions: more partners joining, noisy data, and complex cross-modal relationships.

To answer this, we designed a series of experiments around four key questions:

Experiment 1 — Centralized vs. Federated Performance

• How close can FL get to centralized performance? In a centralized setup, the model sees all the data at once, the ideal scenario for learning. FL, by design, fragments this data across clients. The first question is whether this tradeoff costs us meaningful accuracy, or whether FL can match centralized results despite never accessing the full dataset.
• What happens as more clients join? In practice, a federated network might involve a handful of partners or dozens. As the number of participants grows, each client holds a smaller, potentially less representative slice of the overall data. We tested how model performance scales as we increase the number of clients.

Experiment 2 — Resilience to Noisy Data

• How robust are centralized and federated models to noisy data? Real-world datasets are messy, labels can be wrongly defined, and data quality varies across partners. We deliberately introduced noise into the multimodal dataset to simulate these imperfections and measure how much degradation the model can tolerate before performance breaks down.

Experiment 3 — Cross-Modal Relationships

• How sensitive centralized and federated models to underlying cross-modal patterns? Multimodal models learn by connections between different types of data. For example, a luxury brand might target a high-income audience through a premium creative tone on a specific platform. Some of these connections appear frequently in the data, while others are rare. We tested whether emphasizing the most frequent cross-modal patterns in our synthetic data improves performance compared to emphasizing the least frequent ones, helping us understand how much the model benefits from common, naturally occurring relationships versus rare, atypical ones.

The Data

For our experiments, we used a multimodal synthetic dataset generated by our own well-tested synthetic data generator, designed to mirror real-world marketing dynamics. The generator allows us to customize various elements of the data and design targeted datasets that stress-test our model architecture under controlled conditions, giving us full visibility into the factors that drive campaign performance.

Each campaign in the dataset is described using five key modalities:

• Audience – the consumer segment being targeted
• Brand – the positioning and perception of the brand
• Creative – the tone and message of the campaign
• Platform – where the campaign runs
• Geography – the markets being targeted

Each dataset’s sample is assigned a target label, Positive (Over performing), Negative (Under performing), or Average (Average performance), indicating whether that particular combination of modalities would lead to a successful, underperforming, or average campaign outcome.

Experimental Results

All federated experiments are implemented using Flower, a widely adopted open-source framework for federated learning research and deployment. Flower allows us to simulate multi-client federated setups in a controlled environment, making it possible to rigorously test different configurations before moving to a fully distributed architecture.

To ensure a fair comparison between centralized and federated setups, we kept the playing field level. Both setups use the exact same model architecture, so any performance differences come from how the model is trained, not what is being trained. In the federated setup, data is split equally across clients so that each partner sees a representative sample. This way, when we increase the number of clients, any change in performance can be attributed to the scaling itself, not to differences in what each client’s data looks like.

Experiment 1

Question: How does FL performance compare to centralized training? What happens as more clients join?

Figure 2: Impact of increasing client fragmentation on Federated Learning performance. Performance clearly degrade as the number of clients increases from 5 to 15, compared to the baseline centralized model version.

The centralized model sets the performance ceiling at 79.67%. This is expected, when a single model has direct access to all the data at once, it has the best possible conditions to learn. No information is lost to partitioning, and no coordination overhead is introduced. It’s the ideal scenario, and the benchmark everything else is measured against.

The federated results tell a clear story: as we add more clients, performance gradually declines. With 5 clients, the model reaches 76.23%, a modest drop from the centralized baseline. But as we scale to 10 and then 15 clients, scores fall to 70.29% and 67.65% respectively. The same pattern holds across all metrics, with the sharpest drops in the model’s ability to correctly identify both positive and negative cases.

Why does this happen? As more clients join, the total dataset gets divided into smaller slices. Each client sees less data, which means each client’s local training produces a less reliable picture of the overall patterns. When the server combines these local updates, the differences between them make it harder to converge on a strong global model, an effect we call the “aggregation penalty.”

Lesson Learned: FL with 5 clients comes remarkably close to centralized performance, showing that federated collaboration is viable with minimal accuracy loss. However, as the number of clients grows, makes it progressively harder for the global model to match centralized results.

Experiment 2

Question: How robust are centralized and federated models to noisy data?

In practice, marketing data is never perfectly clean. Campaign outcomes don’t fall neatly into “this worked” or “this didn’t.” Was a campaign that slightly exceeded expectations truly a success, or just average? Was a modest underperformance a failure, or noise in the measurement? Different teams may label the same outcome differently, tracking systems introduce inconsistencies, and the line between a “positive” and “average” campaign is often blurry.

To simulate this reality, we deliberately introduced noise into our synthetic dataset by blurring the boundaries between performance classes. With no noise, the labels are clean — positive, negative, and neutral outcomes are clearly separated. As we increase the noise level from low, to medium, and then to high, the boundaries between these classes increasingly overlap, making it harder for the model to tell them apart. Think of it like gradually turning up the fog: the underlying patterns are still there, but they become harder to see. The federated learning simulation for this experiment was configured with 5 participating clients, consistent with the best-performing federated setup identified in Experiment 1.

Figure 3: Performance comparison of centralized (left) and federated learning (right) configurations across increasing noise levels. Both paradigms degrade gradually, with Positive F1 and Negative F1 most affected, while the performance gap between the two remains approximately constant across all conditions.

As expected, both models perform best on clean data and gradually decline as noise increases. At high noise:

• The centralized model’s score drops from 82.05% to 78.11%
• The FL model’s score drops from 80.74% to 76.09%

The good news: neither model collapses. Even at the highest noise level both models still perform reasonably well. The overall accuracy dips, and the models struggle most with distinguishing clearly positive or clearly negative campaigns, which makes sense, since those are exactly the boundaries we blurred. However, their ability to capture general patterns across the dataset remains stable throughout.

As in Experiment 1, the centralized model maintains a consistent edge over the federated setup at every noise level, but the gap between them stays roughly the same. This means that FL doesn’t become more fragile in noisy conditions; it handles data messiness about as well as its centralized counterpart.

Lesson Learned: Real-world data is inherently noisy, and any viable model must be able to handle that. Both centralized and FL models show strong resilience — performance declines gradually rather than breaking down, even when the data is heavily corrupted. Importantly, FL’s relative performance holds steady across noise levels, suggesting it is no more vulnerable to messy data than centralized training.

Experiment 3:

Question: How sensitive centralized and federated models to underlying cross-modal patterns?

Our synthetic data generator creates campaign data based on a graph of relationships between five key factors: Audience, Brand, Creative, Platform, and Geography. Each relationship captures whether a particular combination of these factors tends to drive strong or weak campaign performance. Some of these relationships are common and obvious — they show up frequently and reflect well-known marketing dynamics. Others are rare and subtle — unusual combinations that don’t appear often but may carry uniquely valuable signal about what makes a campaign succeed or fail.

Understanding how these different types of patterns affect learning is important for both training paradigms. If the nature of the underlying data patterns matters, we need to know whether centralized and federated models respond to them in the same way — or whether one setup handles certain patterns better than the other. To investigate this, we generated three versions of our dataset, keeping everything else the same:

• Common-first: The generator focuses on the most frequently occurring combinations and downplays the rarest ones. This gives us a dataset dominated by typical, familiar marketing patterns.
• Rare-first: The opposite — the generator prioritizes the rarest combinations and downplays the most common. This fills the dataset with unusual, less obvious patterns.
• Middle-ground: The generator focuses on combinations that fall in the middle of the frequency spectrum, neither the most common nor the rarest.

As in Experiment 2, the federated learning simulation was run with 5 participating clients, and performance was compared against the centralized baseline across all three dataset versions.

Figure 4: Impact of cross-modal relationships on model performance. Prioritizing rare feature combinations (Rare-first) substantially improves accuracy compared with focusing on common patterns, showing that atypical relationships provide a stronger learning signal for both centralized and federated learning paradigms.

The results were striking. The Rare-first configuration dramatically outperformed the other two, achieving peak scores of 94.41% (Centralized) and 93.43% (FL), compared to scores in the 86–88% range for the Common-first and Middle-ground setups.

This tells us something counterintuitive: the model learns far more from unusual feature combinations than from common ones. The typical, frequently seen patterns are in some sense “easy”, they don’t give the model much new information. But rare combinations force the model to learn more nuanced and distinctive boundaries between what makes a campaign succeed or fail.

As in previous experiments, the centralized model maintains a small edge over FL, but the ranking between dataset strategies stays the same in both setups. Whether training centrally or federally, prioritizing rare patterns is the winning strategy.

Lesson Learned: Not all data is equally valuable. Prioritizing on rare, atypical feature combinations produces significantly better models than focusing mostly on common patterns. This has direct implications for how we design synthetic datasets: rather than mimicking the most typical marketing dynamics, we should deliberately include uncommon combinations to give the model a richer and more discriminative learning signal.

The Impact and Looking ahead

This work is just the initial spark for our federated learning efforts. Verifying that the centralized ML model performance our company provides is slightly degraded under a reasonable number of users opens the discussion about delivering ML solutions that address shared challenges among clients who are reluctant to share data to tackle a common industry problem. The FL approach allows companies to securely train a shared global model on their own datasets without the risk of data leakage throughout the training process.

Although Federated Learning has been an established collaborative learning method since 2017, it remains a highly active research domain in academia and a strategic priority for industrial implementation. The findings from WPP Lab’s initial FL research establish the foundation for further exploration, specifically focusing on the following directions:

1. Privacy Constraints in Malicious FL Environments

While FL provides a relatively secure framework for multiple organizations to collaboratively train a global model, a critical question remains: how safe is it to exchange local and global model updates during each communication round? Extensive literature shows that FL networks are vulnerable to attacks from malicious clients or a compromised central server. Consequently, there is a pressing need for robust defense mechanisms that enable honest participants to verify the integrity and security of the collaborative learning process.

2. Evaluation of Advanced and Realistic FL Scenarios

While simulating collaborative training with evenly distributed data provides a valuable baseline for foundational FL research, it does not fully capture the complexities of real-world implementations. Our next objective is to build on our preliminary investigations into data heterogeneity, drawing on the noise-injection experiments previously conducted on synthetic datasets. We will also assess the efficacy of using a shared synthetic dataset on the central server as a benchmark to evaluate the integrity of incoming model updates and detect potential malicious activity. Finally, we plan to move from the simulated FL environment currently facilitated by the Flower framework to a fully distributed architecture. By deploying distinct nodes to represent separate corporate entities, we will empirically investigate and address the communication bottlenecks inherent in practical FL deployments.

Disclaimer: This content was created with AI assistance. All research and conclusions are the work of the WPP AI Lab team.

Meet Your New Data Guardian

1. The High Cost of “Dirty” Data

In the modern enterprise, data is the raw ingredient behind every strategic decision. Think of it like a premier restaurant: the Data Engineer is the sous-chef, meticulously sourcing and preparing ingredients, while the Data Scientist is the executive chef, transforming them into the predictive models and insights that drive the business forward. If the ingredients are spoiled or mislabelled, the final dish fails, no matter how talented the chef.

Across several of our AI initiatives at WPP, we uncovered a pattern that was quietly draining velocity from our most ambitious projects. Our “sous-chefs”, skilled data engineers responsible for pipeline integrity, were spending up to one full day per week on tedious, largely manual Quality Assurance (QA) of data flowing into BigQuery. Row by row, column by column, they checked for missing values, logical contradictions, and phantom duplicates, work that was essential but deeply repetitive.

This wasn’t just an inconvenience. It was a strategic bottleneck: it slowed the delivery of every downstream AI application, consumed senior engineering talent on janitorial tasks, and most dangerously created risk. When a human eye is the only safeguard between raw data and a production model, errors don’t just slip through occasionally. They slip through systematically, at exactly the moments when the data is most complex and the engineer is most fatigued.

We asked ourselves a different question: What if, instead of building another dashboard or writing another validation script, we built an intelligent agent, one that could reason about data quality the way an experienced engineer does, learn from every audit it performs, and get better over time?

This article describes how we built that agent, what makes it fundamentally different from traditional automation, and what happened when we put it to the test.

2. The Problem: Why Data Quality Demands More Than Scripts

The Data & The Modeling Ecosystem

The agent operates on digital marketing campaign performance data hosted in BigQuery, massive tables that track how advertising campaigns perform on a daily basis across major ad networks like Meta (Facebook and Instagram). Each row represents a highly granular intersection of a specific campaign, audience segment, platform, device, and creative asset. This data captures everything from broad identifiers—like the parent brand and geographical targeting—down to precise performance metrics, including impressions, clicks, daily spend, conversions, leads, and app installs.

This foundational data is the lifeblood of two critical machine learning systems:

1. The Prediction Model: A classification system designed to predict whether a planned campaign will yield a negative, neutral, or positive outcome.
2. The Recommendation System: A highly flexible advisory engine capable of handling any combination of “missing modalities.” For example, if a media planner inputs a specific Brand, Target Audience, and Location, the system dynamically recommends the optimal missing parameters, such as the best platform to use and the most effective creative asset to deploy.

For more background on the broader modeling context, see From Guesswork to Glimpse: How AI is Predicting the Future of Marketing Campaigns.

The Silent Threat of Data Corruption

Because these models directly inform real-world media spend and strategic campaign planning, their accuracy is paramount. The underlying data is regularly refreshed directly from the advertising platforms to keep the models up to date. However, this automated refresh process frequently introduces subtle corruption and systemic inconsistencies.

For instance, while metrics like engagement and clicks generally remain stable, downstream pipeline issues frequently render conversions and awareness metrics unreliable (“not high quality”). At the individual row level, these anomalies are often entirely invisible. But at scale, they are devastating. If left unchecked, these untrustworthy data points bleed into the training sets, silently degrading the prediction model’s accuracy and causing the recommendation engine to suggest sub-optimal, expensive campaign configurations. This makes rigorous, automated data quality validation not just a nice-to-have, but an absolute necessity for the ecosystem to function.

The Failure Modes

The scale and velocity of data flowing into BigQuery mean that errors don’t announce themselves. They hide. Through our manual QA process, we catalogued six recurring failure modes, each one capable of silently degrading every model built on top of the data:

A static validation script can catch some of these — the easy ones, the ones you’ve already seen. But scripts are brittle: they encode yesterday’s assumptions and break on tomorrow’s edge case. They cannot reason about why a pattern looks wrong, weigh it against historical context, or decide whether a recurring anomaly is a genuine error or a known artifact of a data source.

That requires judgment. And judgment is what we built the agent to provide.

3. Our Approach: An Agent That Reasons, Remembers, and Improves

We designed the Data Quality Assurance Agent — not as a script, not as a dashboard, but as a reasoning entity capable of planning an audit strategy, querying data, forming hypotheses about its health, testing those hypotheses, and learning from the results. The distinction matters. A script checks what you tell it to check. An agent decides what to check, based on what it knows and it has the tools to act on that decision end-to-end.

Architecture: One Agent, Specialized Tools

The agent is powered by a single reasoning core that plans, decides, and acts. What gives it breadth is its toolkit, a set of specialized capabilities it can invoke as needed, selecting the right tool for each step of the audit:

• Database Tool: enables the agent to query BigQuery directly, fetching schemas, row counts, column statistics, and raw data samples.
• Auditing Tool: the agent’s analytical engine. It formulates hypotheses about potential quality issues, runs targeted checks, and compiles structured findings. This tool reads from and writes to the Memory Bank.
• Analytics Tool: generates visualizations using Python — charts, distributions, and plots that make audit findings immediately legible to stakeholders.
• Artifact Tool: packages the final audit report, charts, and evidence into downloadable artifacts stored in Google Cloud.

The agent orchestrates these tools autonomously. When a user asks it to audit a table, the agent formulates a plan, queries the data, runs its checks, generates visualizations where useful, and compiles a structured report, all without the user needing to specify which tool to use or in what order.

The Key Innovation: Long-Term Memory

Most AI tools are stateless. When the session ends, everything the system learned disappears. The next audit starts from zero. This is the fundamental limitation we set out to break. The agent maintains a persistent Memory Bank, a long-term knowledge store that survives across sessions and accumulates institutional intelligence over time. This memory captures three categories of knowledge:

1. Historical Explanations When a data engineer confirms that a recurring anomaly is caused by a known tracking limitation or data source quirk, the agent records that explanation. The next time it encounters the same pattern, it doesn’t waste time flagging it as a new issue, it references the known cause, notes it in the report, and moves on to genuinely novel problems.
2. Business Context Over successive audits, the agent absorbs the specific rhythms and patterns of our marketing data, seasonal spikes, platform-specific reporting delays, expected variance ranges for different campaign types. This contextual awareness allows it to distinguish between a real anomaly and normal business variation.
3. Evolutionary Learning With every audit, the agent’s knowledge base deepens. Instead of repeating the same blind checks, it refines its hypotheses based on what it has seen before — which columns tend to have issues, which tables are most prone to duplication, which logical inconsistencies recur. The agent doesn’t just run. It compounds.

This is what separates an agent from a script. A script executes the same logic every time, regardless of history. The agent carries forward everything it has learned and every audit it performs makes the next one sharper.

The Tech Stack

To ensure the agent was enterprise-grade, we built on the full Google Cloud AI ecosystem:

4. Proving It Works: Synthetic Error Injection

We didn’t hope the agent worked. We proved it using a controlled methodology we call Synthetic Error Injection. The premise is straightforward: take a perfectly clean dataset, intentionally corrupt it in specific, measurable ways, and then challenge the agent to find every error we planted. If the agent can detect artificially injected errors, whose exact type, location, and severity we control, we can be confident it will handle real-world data corruption, which is typically far less extreme.

Step 1: Preparing the Test Data

Before injecting errors, we prepare the data for safe, controlled experimentation:

• Anonymization — Real brand and advertiser names are replaced with generic identifiers ("Brand 1", "Company A"). Sensitive business information never enters the test environment.
• Corruption— The dataset then receives a different severity level of corruption. This allows us to map the agent’s detection accuracy as a function of error density, from subtle (5%) to extreme (40%).

Step 2: Injecting Controlled Errors

Using purpose-built scripts, we introduce precisely calibrated corruptions into a table, 4 types of Structural and 7 types of Logical errors:

Step 3: Synthetic Ground Truth Dataset

We keep track of the errors we introduce in a table and produce a ground truth dataset that looks like:

5. Evaluation Pipeline, Experiments and Results

To evaluate our Agent we follow the pipeline below:

The 4 Experiments and Results

Each experiment isolates a single variable to understand what affects the auditor agent’s detection quality.

Experiment 1: Prompt Comparison

Question: Does giving the agent a more detailed prompt improve error detection?

Runs the agent 3 times on the same table, each time with a different user query style:

Key insight from results:  Only the complex prompt successfully detected the injected spend_with_zero_impressions error (139 rows, 1.82%), while both the simple and medium prompts missed it entirely — confirming that more detailed, forensic-style instructions are critical for the agent to test nuanced logical relationships rather than just surface-level checks.

Experiment 2: Table Sweep

Question: How well does the agent detect different types of errors?

Experiment 2 stress-tests the Auditor agent (using the complex prompt) on 11 synthetic BigQuery tables with progressively stacked error combinations, ranging from a single logical violation to the full set of 7 logical plus 4 structural error types (11 total). The agent achieved perfect detection on 8 of 11 tables (72.7% with F1 = 1.0) and delivered strong overall performance (precision 0.905, recall 0.864, F1 0.883).

Experiment 3: Consistency Check

Question: If we run the exact same audit N times, do we get the same results?

Runs the identical configuration multiple times (default: 3).

The Consistency Experiment demonstrated that the auditor agent achieves perfect 100% reproducibility, consistently detecting both injected errors and real data quality issues with the exact same metrics and explanations across all three runs. This deterministic performance confirms that the agent’s complex prompt configuration is highly reliable effectively eliminating the need for redundant audits.

Experiment 4: Memory Ablation

Question: Does loading past audit memories help the agent perform better?

Runs the agent twice on the same table with the complex prompt:

Key insight from results:  Memory doesn’t unlock new capabilities when tables contain simple structural errors or when the user prompt leads to a very detailed auditing but acts as a critical consistency enforcer, stabilizing the audit scope to guarantee a 100% success rate. However, in cases that there are many different logical errors in a table and the prompt is simple, when past memories are loaded into the agent’s present context performance becomes optimal.

6. The Agent in Action

Below is an actual interaction with the deployed agent, demonstrating its reasoning in practice. The agent discovers available tables, loads its accumulated memory from prior audits, and then, guided by that memory, conducts a targeted audit.

Discovery

Human: “What tables do you have access to?”

Agent: “I have access to the following tables in the dataset:

• campaign_performance_table_1
• campaign_performance_table_2
• campaign_performance_table_3
• campaign_performance_table_4
• campaign_performance_table_5

Memory Recall

Human: “Load your memories.”

Agent: “I have loaded the following memories:

1. A general instruction to always check for null values in tables.
2. A specific data quality issue in campaign_performance_table_2 where 188 rows (18.8%) were duplicates based on event_date, campaign_id, and ad_creative_id.
3. A preference for reports to be formatted in Markdown and include an executive summary.”

The agent doesn’t start from scratch. It arrives with institutional knowledge — prior findings, user preferences, and standing instructions — all retrieved from its persistent Memory Bank.

Hypothesis-Driven Audit

Human: “Based on these memories, audit campaign_performance_table_3.”

The agent formulates a plan, queries the table using its database tool, and — informed by its memories — produces a structured audit report:

Executive Summary: This audit revealed a consistent pattern of missing values across key columns, affecting exactly 5% of total rows. Specifically, event_date, campaign_id, impressions, clicks, and conversions each had 50 null entries out of 1,000 rows. No exact duplicate rows were found. This suggests a systemic issue with data completeness rather than random omissions.

The agent then provides detailed findings per column — quantifying impact, assessing severity, and recommending next steps — before offering to proceed with outlier detection, categorical validation, and logical consistency checks.

Persistent Learning

Human: “Save the results for this table.”

Agent: “The audit results have been saved to memory.”

The next time the agent audits any table, it will carry forward what it learned here — the 5% missing-value pattern, the columns affected, the systemic nature of the issue. Every audit makes the next one faster, more targeted, and more informed.

7. Impact and What Comes Next

Operational Impact

The deployment of the Data Quality Assurance Agent has delivered measurable results:

• Reclaimed engineering capacity. By automating the bulk of manual QA, data engineers have shifted from repetitive inspection work to higher-value activities.
• Earlier error detection. Quality issues are now caught at the point of ingestion, before they propagate into downstream models. This “shift-left” approach reduces the blast radius of bad data from hours to minutes.
• Higher model reliability. Marketing agents, analytics pipelines, and machine learning models now operate on data that has been systematically validated, reducing the risk of predictions and recommendations built on flawed foundations.

The Bigger Picture

This agent is more than a tool. It is a blueprint for autonomous data governance, a pattern that can be replicated across any data pipeline where quality, scale, and velocity collide.

We are currently extending the agent along three axes:

• Cross-table auditing: enabling the agent to detect inconsistencies across related datasets, not just within a single table. Many of the most damaging data quality issues manifest as contradictions between tables that individually look clean.
• Event-driven execution: triggering the agent automatically whenever a BigQuery table is updated, transforming data quality monitoring from a scheduled chore into a continuous, always-on safeguard.
• Adversarial stress-testing: today, our synthetic error injection is script-based and manually configured. We are building a dedicated adversarial agent whose sole purpose is to generate increasingly complex, realistic data corruptions, subtle logical contradictions, plausible-looking outliers, correlated missing-value patterns, specifically designed to challenge the QA agent’s detection capabilities. By putting one agent against the other in a continuous red-team / blue-team loop, both improve: the adversarial agent learns to craft harder-to-detect errors, and the QA agent learns to catch them, driving each other toward sharper, more robust performance over time.

Together, these extensions move us toward a future where data quality monitoring is not a task that consumes an engineer’s day. It is a capability the agent handles continuously and intelligently, surfacing only the issues that require human judgment and decision-making.

TL;DR Most tools that claim to predict post performance tell you what’s in a post — they can’t tell you why it would resonate. The Prediction Optimization Agent closes that gap by converting the image, caption, and influencer metadata into a single natural-language description that captures tone, credibility, and visual cohesion the way a strategist would. It then forecasts engagement, measures its errors, and autonomously rewrites its own instructions to produce better descriptions next time — with no human prompt engineering required. The result is a system that doesn’t just predict what will perform, but allows creative teams to draw insights on what makes a post tick — and what doesn’t.

Introduction

Influencer marketing is a multi-billion industry, yet many spending decisions still rely heavily on intuition. Brands pick creators by follower count, approve content by gut feel, and learn what worked only after the budget is gone. The cost of a wrong bet is not just wasted spend. It is missed revenue, lost momentum, and a creative team back at square one.

Everyone in the industry wants the same thing: a way to know, before publishing, whether a post will perform. But the tools built to answer that question often disappoint. They extract surface features like hashtag counts, color palettes, and face detection, then output a score no one can interpret or trust. They tell you what is in a post but nothing about why it would resonate.

The gap is not technical. It is perceptual. What makes a post work is something a skilled marketer can often articulate after the fact: the influencer’s tone felt effortless, the product placement didn’t interrupt the narrative, the caption hit a cultural nerve. These judgments require understanding context, intent, and audience, not just pixels and metadata.

This is the problem the Prediction Optimization Agent was designed to close. Rather than extracting features and producing an opaque score, the agent reads a post the way a strategist would. It examines the image, the caption, and the influencer’s history, then writes a structured natural-language description of everything that contributed to the post’s performance. A creative director can then read this description, challenge it, and act on it without ever needing to trust a black box.

The intuition is simple: the description that best predicts performance is, by definition, the description that best explains it. The agent iteratively refines this description by diagnosing its own errors, identifying what the previous descriptions failed to capture, and rewriting its own instructions to self-improve. Over successive iterations it converges on the specific qualities that actually drive engagement, not because anyone told it what to look for, but because it is guided by the minimization of its own predictive error.

The Anatomy of a Viral Post

Predicting the performance of an ad or a social media post before publishing — and certainly before spending money to boost it — Predicting the performance of an ad or a social media post before publishing — and certainly before spending money to boost it — remains a primary objective for marketers and influencers alike. But how do you distill something as complex and subjective as a social media post into a single prediction?

Consider a typical Instagram post. It is never just a picture. It’s a complex combination of different data types working together simultaneously. Take the influencer post shown in Figure 1. To truly understand why this post succeeds or fails, you need to consider:

• The image itself — composition, lighting, color palette, subjects, products, and setting.
• The caption — where the influencer might share a discount code, crack a joke, or strike an emotional chord.
• The influencer’s identity — their bio, follower count, niche credibility, and historical performance.
• The metadata — the time of day, geographic location, hashtags, and platform-specific context.

Each of these dimensions carries signal. None of them tells the full story alone. The magic — and the difficulty — lies in how they interact.

Figure 1: A typical influencer post. Traditional analytics struggle to measure the combined impact of the visual aesthetic, the caption’s tone, and the underlying metadata. To accurately predict engagement for a post like this, our system analyzes the image, caption and influencer statistics together as a single cohesive unit.

Why Traditional Approaches Fall Short

Historically, the industry has tried to solve this problem by breaking a post into its component parts and analyzing each one in isolation. Engineering teams would string together separate, disjointed machine-learning models — each responsible for a single modality — and attempt to fuse their outputs at the end:

• Standard Computer Vision Models: Isolated image-recognition algorithms scan the visual to detect objects, people, or products. Separate models handle face detection and emotion recognition. The output is a list of labels: “person detected,” “beverage detected,” “outdoor setting.”
• Text Analyzers & OCR: NLP tools parse the caption — counting hashtags, flagging emojis, scoring sentiment. Meanwhile, optical character recognition (OCR) software reads any text visible within the image itself.
• Tabular Metadata Algorithms: A statistical algorithm ingests structured fields — follower count, posting time, engagement history — and produces its own independent prediction.

Engineers then attempt to fuse these isolated outputs into a single forecast. But this approach has a fundamental, structural flaw: it cannot capture meaning that emerges from the interaction between modalities.

Consider a concrete example. Imagine a fitness influencer posts a photo of herself laughing mid-sip from an energy drink, with the caption: “My face when someone says they don’t need pre-workout 😂.”

A traditional computer vision model would tag this as: “person detected,” “beverage detected,” “outdoor setting,” “positive facial expression.” A separate text analyzer would count the hashtags and flag the emoji. But neither system — nor any crude fusion of their outputs — would understand the joke: that the humor comes from the caption reframing the laugh as a reaction, turning a standard product shot into a relatable meme.

It would also completely miss that this influencer is a certified nutritionist — meaning her credentials paired with an energy drink carry implicit credibility that a fashion influencer holding the same product would not.

Our Approach: Unifying Multimodal Data Through Semantic Translation

To address this limitation, we developed the Prediction Optimization Agent, a self-improving AI agent that moves beyond traditional approaches by unifying all available data into a single format it can reason about: natural language.

The agent’s core mechanism is deceptively simple but profoundly effective. It takes complex, disparate multimodal data — numerical metrics, images, video, and text captions — and converts everything into a single natural-language paragraph that holistically describes the post’s content, aesthetic, tone, and context. By projecting all of these distinct formats into readable text, heterogeneous data is normalized into a structure that a language model can process as a unified whole.

Instead of treating image and text as separate inputs, the agent uses a single prompt to digest all available information at once. Multimodal LLMs serve as one of the agent’s tools, acting as universal feature extractors that capture the abstract, human-centric concepts that traditional pipelines structurally cannot.

But the agent does not simply produce any description and hope it is useful. It is driven by a feedback loop grounded in predictive error: the descriptions it generates are used to forecast engagement, those forecasts are compared against real outcomes, and the resulting errors tell the agent exactly how much predictive value its current descriptions are capturing — and how much they are missing. Through successive rounds of this loop, the agent autonomously rewrites the instructions that govern how descriptions are composed, converging on the paragraph structure that maximizes predictive accuracy.

This error-driven process has a profound consequence for explainability. The description the agent converges on is not a generic summary — it is the description that the agent has discovered, through empirical optimization, to be the most predictive of real engagement outcomes. In other words, the features highlighted in the final description are there because they matter — because including them reduced prediction error. When the optimized description of a high-performing post calls out “candid humor,” “golden-hour lighting,” and “influencer credibility,” those aren’t arbitrary observations. They are the factors the agent learned to pay attention to because they measurably improved its ability to predict what performs well.

How the Prediction Optimization Agent Works

The Prediction Optimization Agent orchestrates three internal stages in a continuous feedback loop: it observes a post, describes it, predicts its performance, measures how far off it was, and then rewrites its own instructions to produce better descriptions next time — closing the loop and getting measurably better with every iteration, without any human intervention.

Figure 2: The Prediction Optimization Agent architecture. Raw media, metadata, and an initial prompt are fed into Stage 1 (Semantic Translation), which produces a natural-language description of the post. Stage 2 (Engagement Predictor) reads that description and predicts engagement. Prediction errors are then passed to Stage 3 (Self-Optimizer), which autonomously analyzes what went wrong and rewrites the Stage 1 prompt — closing the feedback loop and improving the system’s accuracy with every iteration.

Stage 1: Semantic Translation

The agent begins by ingesting the raw post — the image or video file, the caption text, and all available metadata (follower count, posting time, influencer bio, etc.). Using a multimodal LLM as its translation tool, and guided by a detailed set of internal instructions (its prompt), it produces a single, rich natural-language paragraph that captures not just what is in the post, but what the post means: the visual mood, the emotional tone, the relationship between caption and image, and the brand alignment.

The quality and focus of this description is entirely governed by the prompt — and as we will see, it is the prompt that the agent learns to optimize.

Stage 2: The Engagement Predictor

The agent passes the semantic paragraph to its prediction tool — a model that evaluates the post’s potential performance based entirely on the natural-language description from Stage 1.

The predictor can be any machine learning model with the ability to understand text paragraphs. It can be based on trees, deep learning, or any other compatible architecture. It can even be a fine-tuned LLM, upskilled for predictions in a specific domain. Our Agent is compatible with all these options.

In all cases, the quality of the prediction is only as good as the quality of the semantic description. A richer, more targeted paragraph gives the predictor more signal to work with — which is precisely why improving the description is the single most important action the agent can take.

Stage 3: The Self-Optimizer

This is the core of the Prediction Optimization Agent’s intelligence — and what makes it a true agent rather than a static model.

After predictions are generated for a batch of posts and compared against actual engagement outcomes, the agent enters a self-evaluation cycle. The agent is presented with:

1. The prompt it used to generate descriptions in the current round.
2. The per-post prediction errors — for each post, the predicted engagement versus the actual engagement, and the magnitude and direction of the miss.
3. The aggregate error metrics — overall accuracy statistics such as mean absolute error, broken down by content category, audience size, or post type, revealing systematic blind spots.
4. The complete history of prior prompts and their corresponding error metrics — so the agent can see not just how it performed this round, but how each previous prompt version performed, which changes helped, and which didn’t.

Armed with this structured feedback, the agent doesn’t guess at improvements — it reasons about them. It asks: Which types of posts am I consistently getting wrong? What do those posts have in common? What dimensions of the content is my current prompt failing to capture?

Based on this analysis, the agent autonomously rewrites the prompt instructions used in Stage 1. The rewritten prompt is not a minor tweak — it can add entirely new analytical dimensions, remove instructions that proved unhelpful, or restructure the description format to foreground the features most correlated with engagement.

The critical insight is that the predictive error is the signal that drives the entire system. Every improvement the agent makes to its prompt is motivated by a concrete, quantifiable gap between what it predicted and what actually happened. The agent isn’t optimizing descriptions to be “better” in some abstract sense — it is optimizing them to be more predictive, and the error metrics tell it exactly where it is falling short.

The Agent in Action: Watching the Prompt Evolve

To make this concrete, let’s follow the agent through three optimization rounds on a single post — the Hype Energy ad featuring two young models in blue athletic wear, photographed from a dramatic low angle against a vivid blue sky lined with palm trees.

Figure 3: Example Instagram post used for semantic translation.

Round 1: The Baseline

The agent starts with a deliberately simple, generic prompt:

Prompt (Round 1): “Analyze this Instagram post and the provided metadata. Describe the visual elements of the image, the style, and the mood. Then, explain how the image relates to the brand or influencer described in the metadata.”

The agent’s description:

“This Instagram post is a high-energy lifestyle advertisement for Hype Energy, utilizing a vibrant, ‘streetwear-meets-summer’ aesthetic to target a young, active demographic. The image uses a low-angle shot, which makes the subjects appear dominant and ‘larger than life.’ Two young models are featured in athletic/leisure wear. The female subject in the foreground holds the product directly toward the lens. The male subject in the background wears reflective ‘matrix-style’ sunglasses and a blue Jordan hoodie. The post aligns perfectly with the brand’s identity as described in the metadata. The visual style is polished and fashionable, positioning the drink as a lifestyle accessory rather than just a functional caffeine source. The caption emphasizes ‘Zero Sugar’ and staying ‘sharp’ and ‘unstoppable.’”

This is already more nuanced than what a traditional CV model would produce — but the predictor, working from this description, significantly underestimates the post’s engagement. The error gets logged. Across the full batch, the agent notices a pattern: it is systematically underperforming on highly produced brand-partnership posts.

Round 2: The Agent Adapts

The self-optimizer reviews the Round 1 errors and identifies the gap: the prompt never asked the agent to evaluate visual cohesion between talent styling and brand identity, or the psychological impact of compositional techniques like camera angle. It rewrites the prompt:

Prompt (Round 2): “Analyze this Instagram post and the provided metadata. Describe the visual elements, style, mood, and composition. Assess the visual cohesion between the talent’s styling (wardrobe, accessories, posture) and the brand’s visual identity. Evaluate how compositional choices — such as camera angle, framing, and subject placement — contribute to the aspirational or emotional tone of the post. Explain how the image, caption, and influencer profile work together to create a unified message.”

The agent’s description now reads:

“This Instagram post is a high-energy lifestyle advertisement for Hype Energy that demonstrates strong visual-brand cohesion. The coordinated blue wardrobe across both models — the female subject’s athletic top and the male subject’s Jordan hoodie — creates a seamless extension of the brand’s blue-and-white visual identity, embedding the product within the scene rather than presenting it as a foreign element. The dramatic low-angle shot elevates the subjects into an aspirational, dominant position, making the viewer look up at them — a compositional choice that reinforces the caption’s message of being ‘unstoppable.’ The female subject holds the product at the camera’s focal point with deliberate confidence, while the male subject’s reflective sunglasses and relaxed posture add an effortless, streetwear credibility. Together, the image, caption, and influencer positioning frame the energy drink not as a beverage but as a lifestyle symbol for a young, fashion-conscious audience.”

The prediction for this post improves noticeably. Across the batch, the error on brand-partnership posts drops. But the agent notices it is still missing on posts where influencer credibility plays a key role — posts where who the influencer is matters as much as what’s in the frame.

Round 3: The Agent Sharpens Further

The self-optimizer reviews Rounds 1 and 2 side by side — seeing that the Round 2 prompt improved brand-cohesion analysis but still lacks attention to influencer authority and audience trust signals. It rewrites again:

Prompt (Round 3): “Analyze this Instagram post and the provided metadata. Describe the visual elements, style, mood, and composition. Assess the visual cohesion between the talent’s styling and the brand’s visual identity. Evaluate how compositional choices contribute to the aspirational or emotional tone. Analyze the influencer’s credibility and authority relative to the product category — consider their professional background, niche expertise, audience demographics, and how these factors amplify or undermine the authenticity of the endorsement. Assess the emotional resonance of the caption: does it use humor, aspiration, relatability, or urgency, and how does this tone interact with the visual content? Explain how all elements — image, caption, influencer identity, and audience context — work together to create a unified, persuasive message.”

The agent’s description:

“This Instagram post is a high-energy lifestyle advertisement for Hype Energy that achieves strong visual-brand cohesion and leverages influencer-audience alignment to maximize credibility. The coordinated blue wardrobe across both models creates a seamless extension of the brand’s visual identity. The dramatic low-angle shot places the subjects in an aspirational, dominant position, reinforcing the caption’s message of being ‘unstoppable.’ The female subject’s direct-to-camera product hold at the image’s focal point conveys deliberate confidence, while the male subject’s reflective sunglasses and relaxed posture add streetwear authenticity. Critically, the influencer’s established presence in the fitness and lifestyle niche — with a predominantly young, health-conscious following — makes the ‘Zero Sugar’ energy drink endorsement feel credible rather than transactional. The caption’s tone (‘Stay sharp. Stay unstoppable.’) uses aspirational brevity that mirrors the visual intensity, creating a tight loop between what the viewer sees and what they read. The overall effect positions the product as an identity marker for an active, fashion-forward audience, rather than a commodity — a framing that this influencer’s audience is primed to respond to.”

This description now captures visual strategy, brand cohesion, compositional psychology, influencer credibility, audience alignment, and caption-image interaction — all in a single readable paragraph. The prediction error drops significantly. And a marketing team reading this description can immediately see why the post works: the credibility is authentic, the visual identity is seamless, and the caption reinforces the image rather than fighting it.

The agent arrived at this level of descriptive richness not because someone hand-engineered the prompt, but because the predictive errors from Rounds 1 and 2 told it exactly what it was missing.

Real-World Application: Lessons and Implications

To validate the Prediction Optimization Agent in a real-world scenario, it was tested using an Instagram Influencer Dataset containing approximately 10.18 million posts spanning a diverse range of content categories and audience sizes. The results revealed key insights about both the agent’s learning dynamics and the practical implications for marketing teams.

The Agent Learns What Matters — Autonomously

By processing its own historical error rates, the Prediction Optimization Agent autonomously learned to rewrite its prompts — producing richer, more targeted post descriptions with every iteration, which in turn drove increasingly accurate predictions.

Figure 4 — Autonomous Learning: The chart tracks the agent’s predictive performance (y-axis) across successive optimization rounds (x-axis). Each point represents a full cycle of the agent’s loop: describe → predict → evaluate → rewrite. The trend demonstrates that as the agent iteratively refined its own prompt — guided by quantitative error metrics from prior rounds — forecast accuracy improved consistently and autonomously, without any human prompt engineering.

The agent’s optimization works by feeding it the complete history of prior prompts alongside rigorous, quantitative error breakdowns from every previous round. Armed with this granular self-knowledge, the agent identifies precisely which content dimensions it has been under-analyzing — production quality, humor style, credibility signals, visual-brand cohesion — and surgically updates its prompt to capture those features in subsequent iterations. Each round’s prompt is built on the empirical lessons of every round before it, producing a compounding improvement curve where the descriptions become progressively more predictive and, as a direct consequence, more explanatory.

Conclusions

The Prediction Optimization Agent demonstrates something that extends well beyond social media: natural-language prompts can be treated as tunable parameters, optimized autonomously by the AI itself. By allowing the agent to refine its own instructions through predictive error, the system progressively discovers what drives human engagement and expresses that knowledge in plain language.

For marketing teams, this is a significant departure from opaque AI tools that output a score and leave everyone guessing. When a team wants to understand why one campaign outperforms another, they don’t need to interpret a dashboard of abstract model coefficients. They can compare the text profiles of a high-performing post and a low-performing one, side by side, and immediately see the differences the AI picked up on: one might highlight “authentic, candid composition with humor-driven caption and strong influencer-niche alignment,” while the other notes “generic studio shot with formulaic promotional language and weak audience-brand fit.” The patterns reveal themselves in plain English, and they are the right patterns, because the agent discovered them by optimizing for predictive accuracy.

In practice, this means teams can run draft campaign concepts through the system before committing production and media budgets, getting a readable assessment of how the AI interprets the creative. Designers and copywriters can test variations of a post and compare descriptions side by side to see, in their own language, which direction resonates more strongly. And by normalizing visual and written media into a unified, readable format, brands can pair creative intuition with precise forecasting, treating creative assets as predictable drivers of revenue rather than expensive gambles.

The same architectural pattern, semantic translation, error-driven prediction, and autonomous self-optimization, is not limited to social media. Any domain where success depends on understanding the interplay of qualitative and quantitative signals, from political messaging to product design to entertainment, stands to benefit from systems that can read, reason, reflect, and improve on their own. The question is no longer whether AI can predict what resonates with people. It is whether we are ready to let it teach itself how.

Introduction

In the digital age, a brand is no longer defined solely by what a company communicates about itself. Instead, it is shaped by what millions of people collectively believe it to be. These beliefs are expressed continuously through social media posts, product reviews, news coverage, and countless other digital interactions.

For organizations attempting to understand brand perception, this environment presents a fundamental challenge. The volume of available data is unprecedented, yet the signals it produces are often inconsistent and contradictory. Traditional research tools such as surveys and focus groups remain essential for measuring brand equity, but they capture perception only at specific points in time and cannot fully reflect the fast-moving nature of digital conversation.

At the same time, social media offers a continuous stream of public commentary, revealing how brands are discussed, interpreted, and compared in everyday discourse. However, these signals are noisy and difficult to interpret in isolation.

This tension motivated the development of a Brand Perception Atlas, designed to integrate diverse perception signals into a unified analytical framework. By combining social media data, public knowledge sources, Large Language Model (LLM) summaries, and established brand equity research, the atlas aims to provide a more comprehensive understanding of how brands are perceived across the digital ecosystem.

The first iteration of this project analyzed perception signals for more than 200 brands and over 4,000 individual data points, enabling the construction of a visual representation of brand perception that reveals relationships, consistencies, and divergences across multiple sources.

The Brand Perception Atlas

The Brand Perception Atlas functions as a navigational system for brand strategy, like how a satellite navigation map helps a pilot understand the terrain below. A pilot does not see individual roads or buildings from the cockpit. Instead, they rely on a map that reveals patterns, landmarks, and relationships across the entire landscape.

Brand leaders face a similar challenge. Individual data points, social posts, reviews, survey responses, are like isolated coordinates. On their own, they reveal very little. But when thousands of signals are plotted together, patterns begin to emerge.

The Brand Perception Atlas converts these signals into a shared perceptual map, allowing leaders to see how brands cluster, where they compete, and which meanings they occupy in the public imagination.

To achieve this, the atlas synthesizes perception signals from several sources:

• Content from official brand accounts on platforms such as TikTok and Instagram
• Public narratives reflected in sources like Wikipedia
• AI-generated summaries, using Gemini, describing how brands are perceived in digital discourse
• To anchor the analysis in long-term brand perception, the system incorporates survey-based brand equity data from the WPP Brand Asset Valuator® (BAV).

Figure 1 illustrates the Brand Perception Atlas and the different data sources used. By integrating these sources, the atlas creates a more holistic representation of brand perception than any individual data source can provide alone.

Figure 1: Data sources flowing into the Brand Perception Atlas

The actual data from our sensors for two of the world’s most recognizable brands, Target and Amazon, are outlined in Table 1. By presenting the actual reporting generated by each source, we can see how the Brand Perception Atlas bridges the gap between decades of structured consumer research and the lived experience of these brands in digital spaces.

Table 1: Sample brand perception reports, for Target and Amazon, from the Brand Perception Atlas dataset

Developed by WPP, BAV is one of the largest and longest-running brand equity research programs in the world, with data collected over more than three decades across thousands of brands and markets. Unlike social media and digital sources, which provide indirect signals by analyzing what people post or what algorithms curate, BAV is survey-based and taps directly into human sentiment. By asking consumers directly, it captures deep-seated beliefs through structured imagery attributes such as innovative, reliable, and trustworthy.

Within the Brand Perception Atlas, BAV serves as the foundational benchmark. It represents the long-term equity and established identity of the brand, providing a stable baseline against which the more volatile, “everyday” signals of social media can be compared. This distinction is crucial: while social media shows what people are talking about today, BAV reveals what they fundamentally believe about the brand.

To provide a granular view of brand personality, the BAV system evaluates each brand against a standardized set of 48 Imagery Attributes. These attributes are the “building blocks” of perception, ranging from functional traits like Reliable, High Quality, and Simple to emotional and aspirational markers such as Charming, Daring, and Sensuous. By measuring a brand against this consistent list, the system can pinpoint exactly where its personality “peaks”, whether it is anchored in the “sturdiness” of a traditional leader or the “energy” of a modern disruptor.

In the Brand Perception Atlas, these 48 attributes act as a high-resolution lens. Rather than simply stating a brand is “liked,” the data reveals the specific nuances of why it is liked. For example, a brand may score highly on Intelligent and Visionary while scoring low on Friendly or Down to Earth. This multidimensional data allows the Atlas to plot brands with extreme precision, identifying “perceptual neighbors” who share the same underlying DNA across these 48 distinct dimensions of human sentiment.

Mapping Brand Perception

At the center of the Brand Perception Atlas lies the Perception Map, a visual representation of how brands relate to one another in terms of public perception.

Each perception signal is converted into a numerical representation of semantic meaning, known as an embedding. It is helpful to think of semantic meaning as the “vibe” or “essence” of a word, rather than just the word itself. In a traditional search, if you looked for “Luxury,” a computer might ignore a result for “High-end” because the letters don’t match. However, the Brand Perception Atlas uses semantic embeddings to understand that “Luxury,” “Premium,” and “Prestigious” all mean roughly the same thing. Because their meanings are similar, the system places brands described with these words in the same neighborhood on the map. It maps the ideas behind the words, not just the text.

This semantic translation approach is shared across the Satalia AI Research Lab. The Ad Performance Pod applies a complementary technique, using multimodal LLMs to convert raw ad creatives, images, video, and metadata, into structured text descriptions, and then predicting post-level engagement from those descriptions. Where the Atlas asks “what does this brand mean?”, the Ad Performance Pod asks “what makes this content perform?” — both relying on the principle that rich semantic embeddings can unlock patterns invisible to traditional feature engineering.

These embeddings are then plotted within a shared space. In this space, the distance between points reflects the similarity of the associations surrounding different brands. Brands with similar perceptions appear close together, while those associated with different ideas occupy distinct regions.

The resulting map resembles a landscape of brand meaning, allowing brand leaders to observe clusters of brands that share common associations and identify outliers that occupy unique perceptual positions.

Figure 2: Perception map for Nike across all platforms (US, 2025-26)

When a brand is selected within the Perception Map interface, as shown in Figure 2, the system reveals a rich, multi-layered profile that visualizes how that brand lives in the public imagination. The interface acts as a command center, surfacing the distinct signals captured by each sensor to provide a 360-degree view of the brand’s identity.

Once a brand like Nike is selected, the map highlights its specific coordinates across all platforms, while the sidebar populates with synthesized insights including the brand’s industry, country, and a “Cluster Label”, for example “Hope Innovation Compassion”, which summarizes the overarching sentiment. Users can navigate the Perception Signal Summaries to see the unique reports for each platform, from the foundational BAV sentiment to real-time narratives on TikTok and Instagram, allowing for an immediate comparison of how the brand “sounds” in social discourse versus how it is “valued” in long-term surveys.

Insights from the Unified Perception Map

Once the perception spaces were aligned, several patterns became visible:

1. Omnichannel Consistency (Unified Brands): These brands have the most “clustered” perception signals, meaning they look the same regardless of which sensor you use vs Heterogeneous Perception (Diverse Brands): These brands are “dispersed”, they present different facets of their identity depending on where you look.
2. Shared Equity, Different Vibe (Close on BAV, Far on Socials): These brands share deep-seated consumer beliefs but look different on TikTok/Instagram.
3. Different Equity, Shared Vibe (Far on BAV, Close on Socials): These brands come from different worlds but converge into a single “vibe” in the public imagination.
4. Invisible Infrastructure Brands: These foundational entities occupy a shared emotional territory of essentiality and high-stakes reliability, despite having radically different levels of public visibility.

These insights demonstrate the value of examining brand perception across multiple data sources rather than relying on any single indicator.

Omnichannel Consistency (Unified Brands) vs Heterogeneous Perception (Diverse Brands)

One of the most useful insights derived from the perception map is a metric we refer to as Omnichannel Consistency. This measure evaluates how closely aligned a brand’s perception is across different information sources. If the signals derived from social media, surveys, and public narratives cluster tightly together, the brand is communicating a coherent identity. Conversely, if these signals are widely dispersed, the brand’s presence is more diverse.

Analysis of the dataset revealed several brands with particularly strong consistency across channels, including John Deere,Caterpillar, and Oscar Health, with over 99% omnichannel consistency.

John Deere serves as the definitive benchmark for Omnichannel Consistency, maintaining one of the most stable identities in the Brand Perception Atlas. While many brands act like chameleons, shifting their personality to mirror the fleeting trends of each platform, John Deere operates as a perceptual anchor, where its core essence of “rugged reliability” remains unchanged whether viewed through a 30-year longitudinal BAV survey or a viral TikTok demonstration. Figure 3 demonstrates this pattern, showing all of the John Deere data points close together on the map.

Figure 3: Perception map showing high omnichannel consistency for John Deere on all platforms (US, 2025-26)

This “Nothing Moves Like a Deere” factor is so powerful that even complex public debates, such as the “Right to Repair” controversy, appear with remarkable alignment across all five sensors, reinforcing the brand’s status as an essential, high-stakes infrastructure provider for its community. For leadership, this proves that a sufficiently clear perceptual identity pulls every digital channel toward a single, unified truth.

As shown in Figure 4, the brand’s actual digital content reinforces this stability by bridging the gap between historical heritage and modern utility. The Brand Perception Report for this specific Instagram post highlights a reputation for “practical innovation,” where powerful, reliable machinery transforms arduous tasks like snow removal into efficient, even enjoyable endeavors. The data reveals a narrative centered on problem-solving and the ultimate reward of reclaimed time, positioning the brand as a premium, aspirational solution. While the content emphasizes the ease of machine efficiency over manual labor, it also subtly reflects the “significant investment” required for such equipment, a sentiment that aligns perfectly with the “rugged yet refined” and “worth more” attributes captured by the BAV survey sensors.

Figure 4: Screenshot of a John Deere Instagram video showing snow removal equipment

In contrast, several brands demonstrated much greater variation in perception across sources. These included Marriott, Southern Living, and American Airlines, whose signals appeared more widely distributed across the perception map.

Marriott provides a contrasting example of a multi-faceted brand that successfully occupies distinct perceptual regions depending on the sensor used. The BAV sensor positions the brand within a territory of reliability, superior quality, and prestigious appeal, where consumers fundamentally view Marriott as an intelligent, high-performing, and traditional leader. However, the social sensors (Instagram and TikTok) pull the brand into a separate region defined by aspirational luxury and curated escapism. In these digital spaces, Marriott’s intended portrayal is human-centric and inclusive, with official reports describing an organization that “champions individuals” and demonstrates “deep cultural understanding.” This is brought to life through actual content featuring stories of people, such as the one shown in Figure 5, an Instagram post highlighting Chef Kashif Khan using sign language to communicate at the St. Regis Mumbai. This mapping is diagnostic for leadership, revealing that Marriott is not a monolith; it is a reliable corporate giant that strategically uses social media to humanize its identity through stories of diversity, sustainability, and global connection.

Figure 5: Screenshot of the Instagram video of Chef Kashif Khan using sign language

Interestingly, multiple travel-related brands appeared in this lower-consistency group. The high volume of online discussion surrounding travel experiences, ranging from positive stories to customer complaints, may contribute to a more fragmented perception environment for brands in this sector.

Perhaps the most critical lesson our dashboard taught us is that online buzz doesn’t automatically change what people fundamentally believe about a brand. Social media is a powerful megaphone, but it’s not always a mind-changer. It is easy to be distracted by conversation volume and the sheer amount of noise, visibility, and viral activity a brand generates on platforms like TikTok or Instagram. However, the Atlas looks past this “digital footprint” to uncover a brand’s equity: the deep-seated associations that define how a brand is understood by the public. When we look at these deeper signals, brands that seem worlds apart on social media often turn out to be close neighbors in the mind of the consumer.

The perfect case study for this is Palo Alto Networks, a titan in the cybersecurity industry.

If you were to judge them solely by their Instagram presence, you’d see a flawless picture of a market champion. The conversation is filled with words like “Reliable, Innovative, Leader.” It’s a polished, professional, and highly positive image, the brand broadcasting its strengths and amplifying its message of leadership.

But our BAV data, the “ground truth” of long-term brand equity, built from years of customer experiences and market reputation, tells a more complicated story. The BAV data agrees on their high status, describing them as “Intelligent” and “Prestigious.” However, it adds a critical, and very human, third dimension: “Arrogant.” This single word is the key. The perception of “arrogance” isn’t something a few Instagram posts can erase. It’s a feeling built from countless interactions: sales calls, pricing negotiations, customer support tickets, and industry chatter. It’s a core part of their brand’s DNA, for better or worse.

What the map shows is that the high volume of positive conversation on Instagram is successfully amplifying the “Leader” and “Innovative” message. However, it hasn’t fundamentally altered the deep-seated perception of arrogance. The social media buzz is happening on the surface, while the core brand equity, the good and the bad, remains firmly in place. It’s a powerful reminder that you can’t just tweet your way to a new reputation.

In other words, social media can amplify a reputation, but it rarely rewrites it.

For leadership, this mapping is diagnostic, offering a clear view of whether this dual identity represents a successful multi-segment strategy or an unintended drift from the brand’s core promise.

It is important to note that omnichannel consistency is not inherently positive or negative. Some brands benefit from a tightly aligned identity across all platforms, while others thrive precisely because they present different facets of themselves in different contexts. For example, entertainment, fashion, and travel brands often generate highly diverse narratives online, reflecting the variety of experiences consumers have with them. In these cases, perceptual fragmentation may actually be a sign of adaptability and cultural relevance.

The value of the metric is therefore diagnostic rather than prescriptive. It reveals where a brand sits on the spectrum between a unified or multi-faceted perception ecosystem, allowing leaders to determine whether that specific pattern aligns with their strategic intent. For some, a tightly clustered, unified identity is the goal; for others, a more distributed, multi-faceted presence across different platforms is a deliberate strategy to reach diverse audiences.

Shared Equity, Different Vibe (Close on BAV, Far on Socials)

The Atlas also reveals unexpected relationships between brands that operate in entirely different industries. The most striking examples of the “Perception Paradox” occur when two brands share an identical equity backbone but present entirely different faces to the digital world.

Consider 3M, (originally the Minnesota Mining and Manufacturing Company) a global diversified technology conglomerate that applies material science to develop over 60,000 products for industrial, worker safety, healthcare, and consumer markets. At its core, 3M functions as a science-driven “innovation engine.” The company is famous for its “15% Rule,” which allows employees to spend a portion of their time on passion projects, a culture that famously led to the invention of the Post-it Note.

Then consider Marriott, a premier global hospitality company that operates, franchises, and licenses a vast and diverse portfolio of hotels and lodging facilities.

On social media, these brands live in completely different worlds. 3M’s presence is defined by science and industrial innovation, while Marriott trends through a lens of “aspirational luxury” and global travel. If you judge only by social footprint, you would assume these brands occupy different cultural spaces; indeed, on social sensors, they move to opposite ends of the map, as 3M focuses on the mechanics of research and Marriott on the spirit of leisure.

Figure 6 illustrates this social-layer distinction for 3M. This video is just one example of the content driving an overall Instagram vibe characterized as Innovative, Responsible, and Sustainable. The brand perception report for this post captures an aspirational narrative of “innovation, education, and community.” By highlighting initiatives like the “3M Young Scientist Challenge,” the brand is perceived as a champion of scientific exploration, fostering a “fun” and “inspiring” platform for young people to solve real-world problems. While this specific content positions 3M as a driver of STEM education, the Atlas reminds us that beneath this vibrant digital layer lies the same sturdy, reliable identity that makes it a perceptual neighbor to a hospitality giant like Marriott.

However, when we look at the BAV sensor, they appear as surprisingly close neighbors. The reason becomes clear when you look at the deeper associations: both brands are anchored by Reliability and Leadership. In the foundational mind of the consumer, 3M is an “innovation backbone,” while Marriott is a “service backbone.” While their social streams look nothing alike, their underlying brand equity is rooted in being an essential, dependable institution.

Figure 6: Screenshot of 3M Instagram video highlighting the Young Scientist Challenge

Now, consider Costco, a warehouse retailer known for bulk purchasing, and JetBlue, a major passenger airline. At first glance, these brands share little in common; however, when we look at the BAV sensor, they emerge as close neighbors within a shared “Consumer Champion” territory.

On BAV imagery attributes, both brands are anchored by a core identity of being Friendly and Reliable. Costco is fundamentally perceived through a lens of Simplicity, while JetBlue is defined by its Value. In the foundational mind of the consumer, both brands occupy a similar emotional space: they are the “trusted alternatives” that provide high-quality essential services without the exploitative friction often associated with their respective industries.

As shown in Figure 8, the interface allows us to see how these signals diverge as we move into the “Experience Layer” of social media. On Instagram, Costco remains tightly aligned with its BAV profile, with its digital presence reinforcing a narrative of affordability and value for money. JetBlue, however, shifts into a more aspirational territory on social platforms, emphasizing convenience and the lifestyle of travel.

Figure 7: Screenshot of the Brand Perception Atlas showing Costco and JetBlue, with Costco selected as the focus brand

This distinction is diagnostic for brand leaders. It shows that while two brands may share an identical “equity backbone” as consumer champions, their social expressions can vary significantly based on platform context, Costco doubling down on its “no-frills” value while JetBlue leans into the aspirational allure of the journey.

The result is that two brands from entirely different industries converge within the same emotional territory: brands that make consumers feel smart, respected, and well served.

Different Equity, Shared Vibe (Far on BAV, Close on Socials)

The Brand Perception Atlas also reveals an inverse pattern: brands that come from entirely different industries and share almost no long-term equity, yet converge into a single, unified “vibe” on social media. In these cases, the social layer acts as a cultural blender, pulling disparate corporate giants into the same neighborhood.

Fox is a major American media conglomerate primarily known for its television broadcasting, news, and sports programming. It operates the Fox Broadcasting Company, Fox News, and Fox Sports. Its identity is built on high-stakes live events, character-driven drama, and real-time news reporting—a business that naturally scores high on “Dynamic” and “Charming” attributes but often generates polarizing sentiment.

Kimberly-Clark is a global giant in personal care and hygiene products. It manufactures essential household brands such as Huggies diapers, Kleenex tissues, Cottonelle toilet paper, and Kotex feminine hygiene products. Its business is rooted in high-volume manufacturing and daily utility, leading to a long-term reputation for “Reliability” and “Practicality” that is very different from the world of mass media.

When we look at the BAV sensor, Fox and Kimberly-Clark are worlds apart, with a distance. Fox is perceived through a lens of high-stakes entertainment and sophistication, scoring highly on attributes like Charming, Sensuous, and Dynamic. Kimberly-Clark, meanwhile, is anchored in the “Practical Utility” of household essentials, defined by Reliability, Quality, and Trustworthiness.

Figure 8: Screenshot of the Brand Perception Atlas showing Fox and Kimberly-Clark, with Fox selected as the focus brand

To the average consumer, a polarizing media giant and a dependable tissue manufacturer should share no common ground. However, as shown in Figure 10, the Social Sensors reveal a convergence. On Instagram and TikTok, the distinction between a news broadcast and a household utility fades as both brands pivot toward lifestyle-centric storytelling.

According to the social reports, Fox is perceived through a lens of “intense, character-driven drama” and high-stakes emotional depth, appealing to those who value mature storytelling. Simultaneously, Kimberly-Clark’s social presence is described as “authentic and engaging,” characterized by a “vibrant” and “charismatic” brand presence that uses humor and social commentary to create a “deeper connection beyond entertainment.” In this digital layer, Fox’s dramatic clips and Kimberly-Clark’s community-focused narratives occupy the same neighborhood. They are no longer judged as “The News” vs. “The Diaper Company”; instead, they are unified as Established Lifestyle Entities that use high-quality content to build emotional resonance with their audiences.

Invisible Infrastructure Brands

The Atlas also reveals a unique category of brands that operate as “Invisible Infrastructure.” These are entities that sit at the center of the perception map not because they are famous, but because they are foundational. One striking example is the proximity of Pfizer and Kyndryl.

Kyndryl is one of the world’s largest IT infrastructure services companies, spun off from IBM in 2021. It manages the mission-critical technology backbone for hospitals, banks, airlines, and governments around the world. Most consumers have never heard of its name. Its revenue is enormous; its consumer presence is almost zero. And yet, in our perception map, it sits directly next to Pfizer.

Figure 9: Screenshot of the Brand Perception Atlas showing Kyndryl and Pfizer, with Kyndryl selected as the focus brand

On the BAV survey, both brands score highly on imagery such as intelligent, reliable, progressive, and original, placing them far from lifestyle or aspirational brands. The depth of the BAV dataset further reveals that this “Expertise” profile is universally recognized across demographics. For example, both Male and Female audiences consistently identify these brands as Visionary and High-Performance leaders, regardless of their personal familiarity with the company’s day-to-day operations. Despite radically different visibility levels, they share the same underlying promise: critical systems do not fail because of us. Kyndryl has no social media presence. Post-Covid, Pfizer is perceived as controversial on Wikipedia and Gemini.

Kyndryl, a brand name most consumers don’t recognise, sits directly next to one of the world’s most recognised pharmaceutical giants in perception space. Not because they sell similar products, but because they occupy identical emotional territory: the invisible infrastructure of modern life that you only notice when it fails. The data from our sensors provides strong evidence for this “Invisible Infrastructure” connection between Pfizer and Kyndryl. Despite their radically different levels of public fame, they share a distinct perceptual fingerprint. When we look at the BAV data, the reason for this proximity becomes clear. Both brands are anchored by attributes associated with competence and essentiality rather than lifestyle appeal.

The Gemini sensor perceives Kyndryl as a “Powerful and Reliable” entity that provides the backbone for critical systems. The report emphasizes its “indispensable” nature to global commerce. Pfizer, while highly recognized, is described as a “polarization” where the brand is viewed as a “scientific savior” clashing with a “profit-driven corporate giant.” The Wiki and Gemini and reports specifically mention it is a “Big Pharma” entity whose “trust is a highly contested issue.”

Even though one is a household name and the other is a ghost to the average consumer, the Atlas recognizes them as the same thing: the high-stakes machinery that keeps the modern world running. One is noticed because of its global news cycle; the other is only noticed when the “mission-critical backbone” fails.

Conclusion

Understanding brand perception has always been one of the central challenges of marketing and brand strategy. In the digital era, the difficulty lies not in the absence of information but in its overwhelming abundance. Organizations now face a paradox: more data than ever before, yet less clarity about what it means.

The Brand Perception Atlas represents an effort to bring structure to this complexity. By integrating diverse perception signals, ranging from social media discourse to public narratives and LLM summaries, and anchoring them against the long-term ground truth of BAV brand equity data, the atlas makes it possible to observe how brand meaning evolves across different layers of public perception.

This unified view reveals something important: the signals surrounding a brand are not random noise. They form patterns. Brands cluster around shared emotional territories, compete in perceptual spaces that often cut across industry boundaries, and reveal whether social conversation reinforces or diverges from long-standing brand equity.

Rather than replacing traditional research methods, this approach extends them. It allows brand leaders to see not only what people say about brands today, but also how those conversations relate to the deeper beliefs that define brand meaning over time.

In a world where social media amplifies attention but does not always reshape perception, the real strategic advantage lies in understanding the relationship between visibility and meaning. The Brand Perception Atlas offers a way to navigate that relationship. Not as a replacement for human judgment, but as a new map, one that helps brand leaders see where their brand truly sits in the landscape of consumer perception, and how that landscape is changing.

Because in the end, the challenge is not simply measuring brand performance, it is understanding the territory of meaning that a brand occupies in people’s minds and learning how to navigate it.

Learn more about WPP Brand Asset Valuator® (BAV) and how you can access this data for your own analysis.