AI knows your brand. It just won’t recommend you.

A new study from Victorious landed yesterday with a number the industry should be sitting with for a week. Across eight AI platforms and 175 brands, 96% of brands were described accurately when asked directly. And 89% of those same brands never appeared in AI answers to category research questions.

That's the gap. AI systems know who you are. They will not name you when a buyer is comparing options.

This is the finding that finally makes the "brand is the moat" argument concrete, and it also quietly demolishes about half of what's currently being sold as GEO strategy. Most of the tactics being pitched to businesses right now are aimed at fixing recognition. Recognition isn't the problem. The problem is retrieval into a comparison set, and that runs on a completely different set of signals.

Recognition and mention are two different games

The Victorious study measured two things separately, which is why the result is useful. Recognition was tested by asking each AI directly to describe the brand. Mentions were tested by asking category-research questions — the questions a buyer actually types when they're comparing solutions.

Being known is not the same as being retrievable into a comparison set.

The recognition scores were remarkable. Google AI Mode, Gemini, ChatGPT, AI Overviews and Copilot all cleared 83% accuracy across every vertical. Perplexity and Meta AI were patchier, but the ceiling is high everywhere.

Then the mention rates cratered. 89% of brands, invisible in category answers.

The reason this matters is that the two states feel identical to a marketer looking at their own brand queries. You ask ChatGPT "what does [your company] do" and it answers correctly. You conclude your AI visibility is fine. It isn't fine. You've tested the easy question. The hard question — "what are the best options for [your category]" — is the one your prospects are asking, and there's an 89% chance you're not in the answer.

Being known is not the same as being retrievable into a comparison set.

What actually correlated with mentions

The Victorious analysis found the strongest relationships came from signals outside the brand's own website. Referring domains and third-party web mentions both showed moderate correlation with AI mentions — 0.49 and 0.45 respectively.

Neither is strong enough to be a lever on its own. Together they describe something we already knew but keep pretending we don't: your prominence across the wider web is what makes you recommendable. Not your schema markup. Not your llms.txt file. Not whether you've rewritten your homepage in FAQ format.

The study also tested whether link quality mattered more than quantity. It didn't, particularly. High-authority links added a modest lift on their own. What moved the needle was overall footprint — how many places on the internet talk about you, not how prestigious any single one of them is.

If that sounds like the SEO advice of 2014, that's because it is. The mechanism has shifted from PageRank to whatever retrieval-augmented generation is doing under the hood, but the underlying signal is the same: earned prominence across the web.

Why category retrieval is the harder problem

When you ask an AI "describe [brand]", it has one job: find whatever it knows about that named entity and summarise it. This is a lookup task. The brand name is the query, and the model has stored something about it during training or retrieves it live. Almost any brand with a functioning website and a Wikipedia entry passes this bar.

When you ask "what are the best [category] options", the model is doing something structurally different. It has to construct a shortlist. That means retrieving multiple entities that match a category descriptor, then ordering them by some notion of relevance or authority. This is a ranking task performed against a categorical query, not an entity query.

The signals for these two tasks are different. To be recognised, you need to exist in the model's knowledge in a coherent form. To be mentioned in a category answer, you need to be strongly associated with the category across the sources the model draws on — reviews, listicles, industry roundups, comparison articles, forum discussions, news coverage.

Most brands have solved recognition. Almost none have solved category association. And the tools currently being sold as "GEO platforms" mostly measure recognition-adjacent things: whether your brand appears when queried, how your brand is described, whether your schema is being parsed. They don't tell you why you're not surfacing in the shortlists that actually drive consideration.

The SparkToro release is more relevant here than it looks

Rand Fishkin's SparkToro added brand affinity data to its reports this week. It shows you which brands an audience is most familiar with and buys from, drawn from keyword, LinkedIn, and clickstream data.

Read that alongside the Victorious finding and something clicks. The shortlist of brands in an AI category answer and the shortlist of brands an audience already knows are converging. AI systems are drawing on the same web signals that shape human brand awareness — because those signals are what the training data and retrieval indexes contain. Which means the brands that surface in AI answers will increasingly be the brands with existing audience prominence.

That's the loop. And it's not a friendly one if you're a smaller player. The AI answer layer inherits the incumbency of the wider web, then reinforces it by shaping which brands the next round of buyers even see as options.

What this actually means for the work

If you're serious about being mentioned in AI category answers, the honest playbook looks depressingly like the SEO playbook from before anyone cared about AI. It's the earned-media, category-association, brand-building work that has always been unglamorous and always been slow.

The concrete implications:

The prompt-tracking dashboards showing how often your brand appears in ChatGPT are measuring a lagging indicator of a much longer chain of work. Useful for monitoring. Useless as a strategy input, because they don't tell you what to change.

The GEO-specific tactics — LLM-optimised content, chunked FAQ pages, structured data blitzes — mostly influence recognition and citation-formatting. They don't influence category retrieval, which runs on external signals you cannot manipulate from your own site.

Digital PR, journalist relationships, being included in third-party listicles, category comparison pages, forum presence, review sites — these are now doing double duty. They influence human brand awareness and they feed the retrieval index AI systems are drawing on. Same activity, two payoffs.

Category association is a specific asset that needs specific work. Being covered isn't enough — being covered in the context of a category is what matters. A profile of your founder is nice. A mention of your company in a "best [category] for [use case]" article is worth more, because it teaches every retrieval system that queries the category what belongs in the answer.

The uncomfortable bit

None of this is fast. The businesses that will surface in AI category answers over the next two years are largely the businesses that were already doing the boring earned-prominence work five years ago. If you're behind, you're behind, and the fix takes quarters, not weeks.

There's a version of this analysis that's genuinely bleak for challenger brands, and I don't want to pretend otherwise. The recognition-mention gap is wide, the signals that close it are slow-moving, and the AI layer is now compounding whatever brand prominence you already had.

But the honest reframe is that this is what "brand is the moat" actually looks like as an operational reality. Not a slogan. A specific finding, with specific mechanics, telling you that the shortcut everyone's been selling doesn't exist. The work is the work. The people who did it before AI search are the ones who'll benefit from AI search. The people who thought GEO was a schema project are going to keep wondering why their brand keeps getting described accurately and recommended never.