A study published this week puts a number on something many of us suspected: when an AI model goes out to search for information to answer a purchase question, it searches for brands it already knows 3.2 times more often than brands it doesn't. In other words, before a single search happens, before your content even gets a chance to compete, the model has already decided who it's going to look at first.
The study comes from geoSurge, a company that measures brand visibility in AI, and while it carries the bias you'd expect from someone selling exactly that, the methodology is solid and the data is strong enough to take seriously. Let me walk you through what they measured, why this happens and, the part I find most interesting, what a brand that isn't Coca-Cola can do to get into the conversation anyway.
What exactly did the study measure?
The researchers took 66 typical U.S. consumer purchase prompts, along the lines of "which running shoes do you recommend for a marathon", and ran each one 60 times between late May and early June. That gave them 3,960 responses to analyze.
The interesting part isn't the responses themselves, but what happens behind them. When you ask this kind of question to a model with integrated search, it doesn't fire off one search: it fires off several in parallel, reformulating your question from different angles. These are the so-called fan-out searches, which I already covered when I analyzed why ChatGPT switches to English even when you talk to it in Spanish. The study captured 13,281 of those internal searches and cross-referenced them with 1,416 brand-level observations.
That means they measured something that normally stays hidden: the searches the model writes on its own, in its own words, before building the answer you actually see.
What did they find?
Several numbers, all pointing in the same direction.
Brands the model already knew well concentrated 55.7% of brand searches, while brands outside the familiarity top 10 shared barely 17.4%. And when the model decided to search for a specific brand, 63% of the time that search was about one of its five most familiar brands.
The pattern repeats across every industry they measured, with ranges going from 41% to 82% preference for known brands, against 9% to 23% for unknown ones. In no category does the result flip.
There's one more figure, and to me it's the most useful one in the whole study: only 31% of broad searches (the category ones, with no specific brand) included any company name. Hold on to that number, because that's where the door is. I'll come back to it below.
Why do models do this?
Because they work the way we do. If you ask a friend to recommend a sushi restaurant, they won't run an exhaustive survey of every restaurant in your city: they'll name the three they know, maybe google one to confirm it's still open, and answer with that. Their "search" inherits the bias of their memory.
AI models operate in a similar way. They learned about brands during training, by reading essentially the entire public internet, and that prior familiarity defines which names come naturally to them when they compose their searches. Live search comes afterward, as a layer that complements that memory and often confirms it. If the model already knows you, it searches you to verify facts and cites you. And if it doesn't know you, it's unlikely to discover you on its own initiative, because its own searches are written around what it already knows.
This connects with something I've been repeating for a while: Google and the AIs no longer evaluate just your page, they build a profile of your company with everything they can find. This study adds a layer to that idea, because that profile defines not only whether they recommend you, but also whether the model even bothers to go look at you.
Why does this matter now and not a year from now?
Because the share of decisions flowing through AI answers stopped being marginal. Measurements published this same week place Google's AI Overviews between 43% and 48% of all searches, depending on whether you look at Similarweb's or Semrush's data. Nearly half of Google queries already show a generated answer, and that's without counting the people who go straight to ChatGPT, Gemini or Perplexity.
If models favor the brands they know, and models intermediate a growing fraction of purchase decisions, then your brand's familiarity became a distribution variable you can measure, with the same logic you'd use to measure coverage in physical channels.
What can you do if your brand isn't one of the known ones?
Here's the full recipe, in order of effort.
First, target category searches. This is the practical use of the number I asked you to hold on to: if 31% of broad searches carry a brand name, the remaining 69% don't. These are searches like "best cream for sensitive skin" or "invoicing software for small businesses", and there what matters is the relevance and quality of your content more than the fame of your brand. The study itself says so: strong content still lets newer brands show up. Your priority is ranking for the searches models launch when they don't have a brand in mind yet.
Second, work on mentions on third-party sites. Models learn which brands exist by reading press, comparison articles, directories, forums and reviews. A mention in an industry publication or a vendor ranking does double duty: it exposes you to human readers today and enrolls you in the memory of the model's next training run. In my experience with clients, third-party comparison pieces ("the best X in Chile") are among the sources most often cited in AI answers.
Third, get your entity in order. A consistent name everywhere, well-implemented structured data, complete company profiles in the places where models verify information. I already wrote a full guide on why structured data is critical for brand visibility in AI, so I won't repeat myself: the point is that when the model finally does search for you, what it finds has to confirm who you are, what you do and for whom.
Fourth, format your content so it's citable. Direct answers up top, data with sources, clear structure. That's the foundation of what I covered in the Generative Engine Optimization guide.
And fifth, measure. Ask about your category in the main models systematically and record who shows up. If you're only reporting organic sessions, you're missing this part of the picture: in the metrics to measure organic in 2026 I explained how to build that dashboard.
Wrapping up
Tomorrow's exercise takes half an hour and costs nothing: write down the five most important purchase questions in your category, ask them in ChatGPT and Gemini, and note which brands appear and which sources get cited. If your brand doesn't show up, look at the cited sources: that list of sites where you should be mentioned is, quite literally, your visibility work plan for the coming months. The study confirms the field is tilted in favor of the well-known players, but it also shows where the entrance is: most of the searches models run still have no owner.