The collapse of brand perception measurement and the rise of unstructured data
Brand perception isn’t a score, it’s a shifting construct. Learn how to track real-time narratives, detect early signals, and outpace competitors.
On my mind this week
For years, brands have operated under the belief that brand perception is something they can track, quantify, and optimize. Awareness, sentiment, and engagement were turned into structured metrics, feeding a sense of control over how brands are perceived.
These frameworks worked when perception was shaped by broadcast media, traditional PR, and brand-controlled messaging, a time when public opinion moved predictably.
That world no longer exists. If anything, it’s crazier than ever.
Brand perception is not a static metric. It’s a construct. It’s fluid, contextual, and constantly reshaped by decentralized, real-time conversations.
What a company says about itself matters far less than what is being said in group chats, private communities, Reddit or AI-generated overviews, outside of the brand’s reach.
Yet, most companies are still relying on structured sentiment trackers, NPS surveys (OK sure, NPS has its place for tracking customer satisfaction over time, but it’s inadequate for capturing dynamic brand perception shifts), and brand lift studies as if brand perception follows a linear, measurable path.
These models are failing because they were never designed for how brand perception actually forms today.
Why I Am Talking About This
Maybe most people don’t know this about me, but I come from a product background. I was at it for a decade before I transitioned to digital analytics. I have a VERY passionate entrepreneur inside me, so, for the past 2 years, I’ve been building AI products in an agency environment.
The products I’ve built and am constantly refining are meant to solve for measuring brand perception and customer voice.
Working with some of the largest brands in the world has thought me a lot about the impasse we are in today and how legacy frameworks for measuring brand perception are failing under the pressure of our current reality.
The industry is obsessed with tracking brand sentiment, while the real competitive edge lies in understanding narratives, detecting shifts before they happen, and mapping the impact of brand perception in ways that actually inform business decisions.
And this is not a theoretical conversation. Companies that fail to adapt are already:
Misreading signals, focusing on high-level sentiment scores that hide deeper shifts.
Reacting too late, treating brand perception as a static KPI instead of a living, evolving force.
Optimizing for misleading metrics, building strategies based on engagement numbers that do not correlate with actual influence.
Beyond the Mean
Most brand perception measurement today is built around structured data. Clean, categorized, and easy to quantify.
Structured data works well for CRM records, sales or controlled survey responses. It is easy to report on, but it only captures what can be explicitly measured, not the complexity of how people build perceptions.
Unstructured data however, represents the vast majority of the brand perception ecosystem.
It consists of:
Conversational data (social media, chat logs, reviews, Reddit discussions).
Behavioral signals (session recordings, rage clicks, customer service transcripts).
Market discourse (WhatsApp groups, industry Slack communities, AI-generated content, Reddit threads, etc).
In fact, IBM estimates that 80-90% of enterprise data is unstructured, despite how most analytics tools still operate within structured models.
By relying solely on structured data, brands are tracking only a small, sanitized slice of brand perception, missing the reality of how narratives evolve in these decentralized ecosystems.
Why Unstructured Data is Hard(er) to Analyze
Unstructured data presents big challenges for traditional analytics models.
AI advancements have made it easier to process it, but generalized LLMs (GPT-4, Claude, Gemini) lack the precision needed for brand-specific insights unless fine-tuned.
Here are some reasons I came across that account for this.
1. Identifying Patterns Over Isolated Mentions
A sudden spike in mentions does not equal a perception shift. Neither does a wave of negative sentiment. Most companies respond to momentary volume shifts rather than mapping whether those conversations are actually reshaping perception over time.
To do this correctly, brands must:
Distinguish noise from real influence. Not every viral moment changes perception. A Twitter (yes, still calling it that) storm might fade in hours, while a quiet shift in Reddit sentiment can redefine a brand’s reputation long-term.
Track how narratives move across platforms. A complaint that starts on niche forums can build legitimacy on LinkedIn before being amplified by mainstream media.
Look at repetition, not just volume. A negative news cycle that disappears in a week is different from persistent themes that resurface in multiple contexts.
What matters isn’t whether people are talking about your brand, it’s how the conversation is evolving and whether it has the power to stick.
Additionally, this unstructured data is spread across multiple sources, and LLMs are not fine-tuned for platform-specific variations. The models you use must be fully trained on the data sources you attach to them and the type of unstructured data you want them to analyze.
2. Contextuality vs. Statistical Averaging
Most sentiment models reduce language into binary or numerical classifications, but brand perception is inherently context-driven.
A phrase like "I love how terrible this company is" is often misclassified as positive.
Criticism from engaged users (e.g., Starbucks fans debating different drinks) gets labeled as negative when it actually signals high retention and excitement.
LLMs are trained on diverse, broad datasets rather than the domain-specific language of brand perception. This is why specialized language models, trained on the nuances of industry-specific discourse, are necessary. Especially for industries like automotive, finance, tourism.
3. Brand perception isn’t just shaped through words.
Meme culture, video narratives, YouTube content now play a significant role in how people interpret brands, yet most perception analysis tools still focus entirely on text.
Focusing on text only unstructured data means we are ignoring:
How a brand becomes a cultural reference. If your brand keeps appearing in meme formats, it’s signaling something, even if nobody is explicitly talking about it.
Non-verbal sentiment shifts. A sarcastic video review can have more impact on perception than a hundred written complaints.
The role of AI-generated content in brand narratives. AI-generated articles, chat summaries, and even fake product reviews are sadly now influencing how brands are perceived, yet few companies track these as part of their brand perception strategy.
And truthfully, text-based only analysis tools alone cannot ever capture the full complexity of how perception evolves today.
The GIANT Data Privacy Problem
The rush to integrate AI into brand perception analysis has led to a reckless disregard for data privacy. And this is an understatement.
Companies are feeding proprietary customer conversations into OpenAI, Claude, and Google models without knowing how that data is stored, retained, or repurposed. Worse, executives are making brand decisions based on AI-generated reports with no verifiable source attribution.
To be compliant, your AI models should be:
Privacy-first and fully controlled. Do not outsource proprietary brand perception data to external LLMs. OR ANY DATA FOR THAT MATTER.
Auditable and explainable. AI-generated insights should be traceable back to raw data, not black-box summaries.
Strategic, not automated. AI should assist human interpretation, not replace it.
Little note here, it’s not ALL BAD.
If you use Gemini as part of GCP AI Studio you are not under the same scrutiny as if you’d use the public Gemini LLM. It’s actually very safe and powerful.
OpenAI has improved privacy options. (e.g., API users can disable training data retention).
Regulatory bodies (e.g., the EU, FTC) are increasingly scrutinizing AI-driven data processing.
5. Brand Perception moves in real time, yet most brands are still tracking it like a quarterly KPI.
Brands that measure perception through quarterly surveys and sentiment reports are working with yesterday’s data.
Real-time perception tracking requires:
Continuous signal detection. Catching shifts when they start, not when they show up in structured reports.
Adaptive AI models. Static machine learning models will fail at tracking evolving narratives.
Focus on movement, not moments. A single viral controversy is less important than a slow, consistent shift in how a brand is talked about over months.
So how can you go about measuring Brand Perception, the right way?
Well, it’s not all doom and gloom.
Brand perception is not a score, a sentiment percentage, or a one-time measurement. As we discovered, it’s a living construct that shifts over time, across platforms, and through different formats of communication. Measuring it correctly requires a mix of network analysis, narrative tracking, and multimodal intelligence.
Here’s what works better. (for now at least; AI changes so fast, this write-up might become obsolete in a month)
Spotting Brand Perception Shifts Before They Become a Crisis
Most brands react too late to perception shifts because they rely on volume-based tracking (spikes in mentions, social listening alerts) instead of how narratives spread.
Use network analysis to track who is shaping the conversation, not just how often a brand is mentioned.
Set up anomaly detection models that flag early deviations in sentiment and discourse patterns.
Track cross-platform narrative movement. What starts as an isolated forum post often hits mainstream media weeks later.
Understanding Where You Win and Lose vs. Competitors
Perception is more than your brand, it’s pretty relative. That’s why measuring perception in isolation leads to blind spots in competitive positioning.
Use entity resolution to compare how your brand and competitors are discussed across platforms.
Track recurring themes associated with competitors that haven’t yet surfaced in your brand discussions. These could be early warning signs.
Look at trust vs. engagement metrics. A competitor might have high sentiment but declining purchase intent, revealing an opportunity.
Tracking Recurring Narratives, Not Just One-Off Mentions
A single negative story doesn’t define perception, but when the same themes repeat in different contexts, they do.
Use topic modelling over time to identify themes that keep resurfacing.
Analyze when certain perceptions peak and fade. Are they temporary spikes or long-term trends?
Separate new narrative formations from old ones resurfacing. If something keeps coming back, it’s shaping reputation.
Separating Skepticism from Actual Brand Damage
Not all negativity is a problem. Loyal customers argue with brands they care about, while disengaged customers can leave silently.
Use emotion and intent classification to separate criticism that signals high engagement from criticism that signals detachment.
Track shift in tone over time. Skepticism turning into disengagement is the real red flag.
Monitor whether high-sentiment customers are still making purchase decisions because sentiment alone isn’t predictive.
Going beyond text. Brand Perception is multimodal.
As discussed, most brand tracking relies entirely on text, but perception is increasingly shaped by memes, short-form video, AI-generated content, and voice.
Use image recognition to track how your brand appears in memes and visual culture.
Analyze TikTok and YouTube content sentiment beyond just text captions. Non-verbal sentiment matters.
Set up real-time multimodal tracking for AI-generated content that references your brand,
Putting data privacy first and foremost.
Own your AI pipeline. Use where possible Specialised Language Models for brand perception analysis instead of gen-AI.
If you do use Gen-AI, make sure AI-generated insights are explainable and traceable. If you can’t trace back why an insight was generated, it won’t be that useful. Unless… it’s an emerging trend. (covering that below)
Adopt privacy-first architectures that allow brand perception tracking without compromising sensitive data. (PII redaction, DLP, the whole shebang)
Bonus section: Hallucinations aren’t really errors. They can predict emerging perception shifts
The AI industry is obsessed with eliminating hallucinations.
The assumption is that if a model generates something that isn’t factually verifiable, it’s a failure. But if we’re measuring brand perception, where opinion, interpretation, and narrative evolution matter more than static facts, is that really the right way to think about them?
Understanding Factual vs. Faithfulness Hallucinations
A 2024 study by Xu et al. argues that hallucinations are inevitable, no matter how much a model is fine-tuned, an AI predicting the next most probable word will sometimes generate content that isn’t explicitly sourced from reality.
Another paper (Huang et al., 2024) categorizes hallucinations into two main types:
Factual Hallucinations.
Inaccuracies that contradict real-world facts. These matter when dealing with legal, medical, or financial information where precision is non-negotiable.Faithfulness Hallucinations.
The model doesn’t contradict facts but goes beyond what was explicitly given, making inferences that weren’t stated directly.
That second type is where things get really interesting for the purpose of this newsletter.
If an AI model generates something that isn’t explicitly stated but aligns with weak signals in unstructured data, is that an error, or an early indicator of an emerging narrative?
If a model detects frustration with a checkout experience, even if no one explicitly says "this checkout sucks," is that a hallucination or a weak signal of sentiment forming?
If a model associates a brand with sustainability despite no explicit statement saying so, is it wrong, or is it capturing an evolving consumer perception before it becomes mainstream?
How to Use AI Hallucinations for Brand Perception Intelligence
Monitor faithfulness hallucinations for recurring themes. If an AI consistently "hallucinates" a connection between a brand and a specific issue (e.g., pricing concerns, sustainability, exclusivity), it might be detecting an association forming in consumer perception.
NB: Faithfulness hallucinations can highlight weak signals but must be cross-validated with real discourse trends.
Compare hallucinations with real-world discourse. Are niche Reddit discussions, Discord chats, or early-stage social conversations starting to support what the model is generating?
Reframe the question from "Is this true?" to "Is this useful?". If an insight helps detect shifting sentiment, surface risks earlier, or uncover hidden connections, does it really need to be 100% factual to be valuable?
Brands that dismiss AI hallucinations entirely might miss early signals of perception shifts before they become visible in structured data. Instead of eliminating hallucinations, the goal should be learning how to manage them as potential predictive tools.
OK, I went crazy long-form again, so let’s wrap this up.
Everything I wrote about is already happening. FAST.
Brand perception is something you interpret, track, and act on.
Stop treating perception as a static KPI and start mapping influence in real time.
Understand that brand narratives evolve through decentralized, multimodal ecosystems, not structured reports.
Recognize that AI hallucinations aren’t always failures, they can surface early indicators of consumer sentiment shifts.
Move from passive measurement to active intelligence, where the goal isn’t tracking what people say, but understanding what they mean.
The brand perception gap is widening between brands that still rely on structured, legacy measurement frameworks and those that understand perception as a dynamic, evolving force. Of course, to some degree you need a mix of both.
I will address this topic in greater depth at 2 conferences (so far) this year with my framework called Insights, Intent and Impact.
This is a framework I’ve built around unstructured data, shaped by real lessons from brands like Starbucks and global fintech companies I worked with. Specifically, it’s about why traditional measurement fails at capturing brand perception, and what to do about it.
Conversion Boost - Copenhagen, March 18th
Experimentation Elite - Birmingham, June 5th (use discount code ExEliteJJ10 for 10% of an a ticket)
PS: A quick shoutout. Thank you for the incredible response to last week’s newsletter. The conversations it created prove why long-form, deep-dive content still matters and why I refuse to dilute things just to hit shitty engagement metrics.
PPS: The latest episode of Standard Deviation Podcast is now live. Me and Simo are kicking things off with AI agents, code, and whether Simo should build a SIMO-bot to replace himself. Then, we switch gears for a real convo with Nicola Strand about imposter syndrome and how she made it work in her favor.
Until next time,
x
Juliana