You Fixed The Copy And The AI Still Doesn’t Say Your Name
Somebody rewrites the product page, tidies the headings, drops an FAQ block at the bottom. Then they open ChatGPT, ask the obvious question about their own category, and their company still isn’t anywhere in the answer.
I’ve been on the receiving end of “so what do we fix next” and had nothing to offer. There was no page left I could point at.
Which, in my experience, means I’m standing in the wrong place and looking hard.
Because before it writes you anything, the model goes off and runs searches of its own. Somebody counted 13,281 of those searches and checked whose names went into them.
Which Brands Does The Model Search For Before It Answers?
Of the brands the model already held in memory, 55.7% got searched at all. Of the brands it didn’t hold, 17.4% — a ratio of about 3.2 to 1 (geoSurge, published early June 2026).
The measurement ran from 29 May to 9 June 2026. Sixty-six questions across nine industries, each one asked 60 times, and not a single brand name written into any of the questions — so every name that surfaced got there on its own.
That produced about 3,960 answers, 13,281 internal searches and 1,416 brand-level observations. “Held in memory” here just means the model can produce the brand without being handed the name, and geoSurge scored that with its own method (their scale, so nobody else’s numbers plug into it). The search side was watched on Gemini 3.5 Flash.
An internal search is what people have started calling fan-out: the queries a model fires at itself before it drafts a word, which you never see. Ask it for the best team chat tools and it doesn’t type that string into a search box; it splits the question up — “team chat tools comparison”, “Company A pricing”, “Company B case study” — and goes looking.
Whether your name lands in one of those slots or not is the whole gap between 55.7% and 17.4%. And it sits a long way before the step everyone actually measures, which is whether the finished answer says your name at all. So that’s three steps where most of us were counting one: get remembered, get searched, get quoted.
The Searches Pile Onto A Handful Of Names
Of the searches that named a specific brand, 63% went to one of the five best-remembered ones. Everybody else divided up what was left.
Those two figures aren’t the same species, and they’re worth keeping apart. The 55.7% and the 17.4% are shares of observed brands that got searched at all; the 63% is how the brand-named searches were distributed. Different denominators, so they don’t stack — although they lean the same way, which is that the concentration happens upstream of anything you can read off a finished answer.
But 17.4% isn’t zero. Even without a memory entry, about one in six of those brands still had its name used as search material.
That’s the number with room in it.
Move The Work One Step Outside Your Own Pages
Whatever the model soaked up about your category, most of it accumulated somewhere other than your own domain — comparison articles, trade press interviews, review listings, industry reports. Rewriting your site doesn’t touch that pile.
That pile is a different surface, worked by different people, funded from a different line.
Your own pages haven’t stopped mattering. They’re later in the order, that’s all. The model uses your name to run a search, the search has to land somewhere, and that somewhere is your site — so the page copy earns its keep once you’ve been looked up.
I had our own AI visibility money sitting in the advertising column and nowhere else, and I’d never once asked whether the step it was buying was the step we were losing at.
How Far Does 3.2 To 1 Actually Travel?
Not as far as a target on a slide. Six things sit between this figure and any goal you’d set from it, and I’d keep every one of them (which is more caution than I’d normally carry into a meeting).
- The publisher states plainly that this is an association in exploratory data, not causation
- Well-known brands are both easier to remember and easier to search for, so part of the association may well be fame doing the work (the publisher says this too)
- The industry-by-industry figures are described as suggestive rather than settled — again, the publisher’s own wording
- The search side was watched on one model, Gemini 3.5 Flash. Nobody has measured how the others fan out
- geoSurge sells AI visibility measurement, so this is a conclusion that happens to be good for its business
- The questions were written for US buyers. Nothing here tells you how any of it behaves in another market
The first one changes how you’re allowed to say this out loud. “Get remembered and you’ll get searched” is a causal sentence, and this study doesn’t reach that far. Maybe a cleaner design gets there later; it hasn’t yet.
Although the six don’t reach the part that’s true by definition: a name that never enters one of those searches can’t come out of the answer, whether the ratio is 3.2 or 2.5 or something else on a model nobody has tested. What they do reach is the step before that one — whether getting remembered is a lever anybody can pull, or just what fame looks like from the model’s side.
Work Out Which Of The Three Steps You’re Losing At
Get remembered, get searched, get quoted — each one gets checked a different way, and none of the checks costs much.
Counting how often a brand shows up in finished answers measures the third. Putting the category questions to a model and reading the searches it fires off gets the second. For the first, I ask a model about the company name and nothing else and read what comes back as a rough signal. Since the study watched a single model fan out, all three counts want running across several — there’s no promise the others reach for names in the same order.
If it’s the first step where you’re falling over, more work on your own pages won’t get there. The money goes wherever third parties write your name down.
Then the planning conversation comes round again and “where does the AI budget go” finally has an answer with a step number in it. We’re losing at step one, so we’re funding communications. I’d have liked that sentence about a year earlier than I got it.
Sources
- geoSurge, “AI Searches What It Remembers”, published June 2026 (measurement run 29 May–9 June 2026: 66 questions across 9 industries, each asked 60 times with no brand name in the question, producing about 3,960 answers, 13,281 internal fan-out searches and 1,416 brand-level observations. Brands held in model memory were searched in 55.7% of observations against 17.4% for brands that were not, a ratio of about 3.2 to 1; 63% of brand-named searches went to one of the five best-remembered brands. Memory was scored with geoSurge’s own method and the search side was observed on Gemini 3.5 Flash alone. The publisher states that these are associations in exploratory data rather than causation, that brand fame plausibly drives part of the association, and that the industry-level figures are suggestive rather than settled. geoSurge sells AI visibility measurement tooling. The questions were written for US buyers.)
- Search Engine Land, “AI models favor familiar brands in search: Study”, 2026 (secondary coverage of the geoSurge measurement)
