Public AI models can answer almost anything. The question is whether you can trust the answer.
Every executive evaluating an AI-powered intelligence platform right now is doing the same mental math. The models are remarkable. The demos are persuasive. And somewhere in the back of the mind sits the experience of asking ChatGPT something you actually know the answer to and watching it confidently get it wrong.
That tension is the real story of AI in business intelligence today. Public models trained on the open web are extraordinary general-purpose tools, but used alone for competitive intelligence they carry real risk: delayed visibility into what's actually happening in market, a bias toward the loudest and most heavily promoted ads, inconsistent outputs across prompts and model versions, no normalized definitions of brands or products or channels, and conclusions you cannot verify.
You get quick answers. You don't get reliable intelligence.
This is the gap MediaRadar has spent the last year engineering against. Earlier this year, we built and began rolling out new AI classification technology to close it — starting with AVOD and social, and now expanding across the broader digital ecosystem. The lesson we keep returning to, and the one we want to share here, is this: AI is only as powerful as the intelligence underneath it. What follows is one proof point of what that looks like in practice.
Most of the AI conversation in ad intelligence asks the wrong question: which data vendor’s AI is better? It's the wrong question because the model isn't the differentiator. The frontier models are increasingly available to everyone, including every vendor you might evaluate.
The right question is: what is the AI being fed?
For two decades, MediaRadar's edge has come from methodology. Anyone can collect ad occurrences. What turns occurrences into intelligence is the layer underneath: a proprietary taxonomy, a Brand Identity System, and the rigor to map every creative against them consistently. That methodology is why our customers can answer questions other datasets can't – which specific product is being promoted, by which advertiser, in which competitive context.
The opportunity in AI, for us, was never to replace that methodology. It was to scale it into corners of the market where the volume of creatives (think: social!) has always moved faster than human review could follow.
Learning 1: The model isn't the differentiator. What you feed it is.
Let’s walk through how most ad intelligence platforms classify a social creative today. A new ad appears. The system grabs whatever structured metadata the platform exposes – landing page URL, advertiser name, sometimes a category tag – and uses that to guess what the ad is about.
This works reasonably well for established brands with consistent metadata hygiene. It works poorly for everything else: long-tail advertisers, performance creatives with stripped-down landing pages, video ads where the metadata says almost nothing about what's actually on screen. The result, across the industry, has been a familiar pattern: a clean top of the leaderboard and a growing "miscellaneous" bucket below it that nobody can quite explain.
The honest answer is that this isn't a tagging problem. It's an input problem.
Learning 2: Hygiene is critical. If you only look at the wrapper,
you only know what the wrapper tells you. The look inside is critical.
Our Classification Engine reads the creative itself – video, audio, on-screen text, visual context – and maps it directly against our taxonomy and product catalog. It runs in two passes. The first extracts structured information from the ad: brand, advertiser, product, visual description, contextual signals. The second takes that information and evaluates it against the most relevant candidates from our taxonomy, returning a final classification with a confidence score.
In between sits the part that makes it work at scale: a search layer that, given the signals from the first pass, finds the handful of taxonomy entries most likely to match – out of a catalog spanning the entire advertising universe – in milliseconds.
As the premier provider of advertising intelligence, we're on a mission to map the entire advertising ecosystem — an undertaking that demands we increase scale without sacrificing quality. We engineered this pipeline to process up to one million ads per day – roughly twelve every second. This will enable our expansion plans: classification quality on social today, the broader digital ecosystem next, additional channels after that. The capacity is in place because our roadmap demands it.
When you build a system like this, the temptation is to reach for the biggest, best-known models on the market. The narrative writes itself: cutting-edge AI, premium components, headline-friendly vendors.
Of course, we tested the obvious commercial option. Then we did something less glamorous: we tested a smaller, retrieval-specific alternative trained for exactly the kind of matching our pipeline needed. The smaller model performed measurably better for our problem – and ran faster and more efficiently while doing it.
The reason this matters is simple: defensible isn't the same as effective. Because the intelligence we feed the model is the differentiator, it made sense to ask a different question: which model fits this specific problem, applied to our specific data, against our specific taxonomy? That strategic choice has always paid dividends for us in speed and accuracy downstream — it's the same discipline that made our methodology industry-leading in the first place. We're just applying it to the new layer of the stack.
Learning 3: The right model beats the biggest model. Fit to the problem, not the brand name.
A system engineered to classify a million ads per day cannot be reviewed creative-by-creative by a human. But classification quality cannot be left entirely to a model, either – that's exactly how you end up with the hallucinations and inconsistencies that make public AI unreliable for serious decisions.
The answer is not a binary choice between automation and human expertise. It's an architecture that uses each for what it's best at. Our model does the first pass at scale. Human analysts focus on aggregate monitoring and detailed sampling to drive quality control, identify edge cases, and deliver feedback that improves the model over time. Every correction becomes a training signal. Every ambiguous case becomes a data point. The system gets sharper with use rather than drifting with it.
This is what we mean when we say AI is a multiplier, not a replacement, of high-quality methodology. The methodology stays human; the application of it scales.
Learning 4: AI is a multiplier of methodology, not a replacement for it.
There's no shortage of vendors telling AI stories right now. The question worth asking them is what they’re applying AI to: is the AI amplifying real intellectual property or papering over its absence?
We've spent two decades building the taxonomy, the Brand Identity System, the product catalog, and the methodology that turns ad data into intelligence. The Classification Engine is the next layer of that work, not a replacement for it. And it's one of several places we're putting AI to work – alongside creative enrichment, spend modeling, and the agent-led workflows on our near-term roadmap – each of them built on the same principle: better AI starts with better intelligence.
Learning 5: Better AI starts with better intelligence — that foundation, not the AI on top of it, is the differentiator.
Our improved engine is rolling out incrementally. We began with AVOD (~600 ads per day, with 100% human review), then expanded to all of social (~40,000 ads/day).
These rollouts have driven measurable gains: greater efficiency with the same high bar for quality in AVOD, and improvements to quality and scale in social, including:
Given this success, we're excited to expand this engine to every creative we process daily, starting with Digital – and someday expanding to Podcasts, Print and more. Any new media we add to coverage will also automatically go through the Classification Engine as part of the enrichment process. We’re also using this technology to extract even more content from the ads we process, giving additional depth and power to our Creative Intelligence offering.
What clients get, in practical terms: a more complete picture of who's advertising, what they're promoting, and where the long tail of competitive activity actually lives.
Innovation, for us, isn't a launch. It's a posture. The engine we built this year will not be the engine we run two years from now, because the inputs, the formats, and the channels will keep changing. What stays constant is the discipline: methodology first, validated technology second, humans where they matter most.
That's the partnership we offer. Build your AI on intelligence you can trust.
MediaRadar's AI Classification Engine is rolling out across paid social in 2026, with expansion to additional digital channels and platforms planned. To see what improved visibility looks like for your specific competitive set, reach out to your MediaRadar account team.
Ready to go deeper on why AI is only as good as the intelligence behind it? Join us live for our State of the Industry: AI in Advertising webinar on August 20th at 1PM EST.
ABOUT THE AUTHOR | Lauren Amira
Lauren Amira is the Vice President of Product Marketing at MediaRadar, where she leads with over 12 years of experience translating complex offerings into clear, compelling go-to-market strategies. A seasoned product marketing leader, Lauren has successfully guided both startups and established brands to craft resonant narratives that align with strategic objectives and drive market impact.