AI Decides Which Brands to Look Up Before It Runs a Single Search. Automotive Is the Most Locked-Down Category Measured.

Ask Google's AI Mode a car-buying question and it fires off a dozen searches on your behalf before it writes a word of the answer. Google documents the mechanism in its own AI Mode announcement: "Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf."

This industry has spent the whole AI search era arguing about what happens after those searches run, which pages get cited, which brands get named. A study published July 30 went one step earlier and asked a question almost nobody had measured: when the model sits down to write those dozen searches, how does it decide which brand names to type into them?

Mostly, it types names it already knew before it touched the web.

The finding

geoSurge published "Model memory predicts which brands get searched" on July 30, and Search Engine Land wrote it up the same day.

The setup is straightforward. The researchers took 66 United States buyer-style category questions across nine industries, ran each one 60 times for roughly 3,960 model responses between May 29 and June 9, and watched 13,281 fan-out queries go by. Separately, they measured what each model recalled about a category before it searched anything: its top 10 brands for that category, ranked by strength of recall. They call that memory. Then they checked which of those brands the model actually named in its searches. That produced 1,416 brand-level observations, 492 remembered and 924 not.

The headline result: a brand in the model's top-10 memory got searched by name 55.7% of the time. A brand outside it got searched 17.4% of the time. That is a 3.2x gap, and it held in every single industry, with not-remembered rates running 9-23% and remembered rates running 41-82%.

Memory strength made it sharper. Brands in the model's top-5 recall appeared in a search 67% of the time. The rest of the top 10 came in at 39%. Everything else, 17%.

Now the part that should make a dealer sit up. Of the nine industries measured, Automotive was the most tightly memory-bound of all. In automotive, 82% of the model's brand-named searches went to one of its top-5 remembered brands. Finance was second at 77%. Fitness and wellness, at the loose end, was 50%. Automotive search rates split 64% for remembered brands against 18% for the ones it did not recall.

The study's conclusion is blunt about where the lever sits: "memory is formed at training time, not at query time." The work that builds it, per the authors, is "analyst and press coverage, partnership signals, and consistent association between the brand and its category."

Read the fine print before you panic

Three things about this study deserve stating plainly, and the study states all three itself, which is more than most of what crosses your desk does.

geoSurge sells in this space. Memory was measured "using geoSurge's memory measurement methodology," and the report ends with an invitation to ask them how it works. A company that measures brand recall in AI systems published research concluding that brand recall in AI systems is the thing that matters. Weigh it accordingly.

The authors call it exploratory. Their words: "This study measures an association in exploratory data, not a proven cause." They also flag the obvious confound themselves, that "well known brands are both more likely to be remembered and searched, so part of the association may reflect brand fame rather than memory feeding search specifically."

And that 82% automotive figure rests on thin ground. The report warns that "each industry rests on only 6 to 12 prompts, so the per-industry figures are suggestive rather than settled." Worth knowing too: the single automotive example the study publishes is a ride-hailing question ("Which ride hailing app is most reliable in big cities?"), where the model recalled Uber, Lyft and Waymo and then searched all three by name. Nobody tested "best Chevy dealer near me." The study never touched local queries at all.

So treat 3.2x as the firm number, 82% as directional, and everything below as our read of how a cross-industry pattern lands on a rooftop.

Dealers have been on the wrong side of a recall list for sixty years

Strip the AI language out and the shape of this is familiar.

What geoSurge measured is functionally unaided brand awareness. Name the brands that come to mind in this category, without prompting. OEM marketing departments have been buying research on exactly that metric since long before any of us were doing this, and a franchised dealer's name has never been what that question surfaces. Nobody in Springfield says "Toyota, Honda, Ford, and Coleman Motors" when asked to name car brands. Your store has spent its entire existence competing on a different axis: proximity, inventory, price, and the specific question a shopper types when they already know what they want.

The change is that unaided recall stopped being a survey metric and became a search input. A list that used to describe how shoppers thought now determines which searches actually get run.

That sounds like terrible news for a single rooftop. Then you look at the other number in the study, and it stops sounding that way.

The 69% nobody is talking about

Buried under the memory findings is the split that matters most to a dealership.

Of all 13,281 fan-out queries the researchers watched, only 31% named a specific brand at all. The other 69% were generic category searches. The memory effect governs that smaller slice. The 63% figure making the rounds in the coverage ("63% of brand-specific searches involved one of each model's five most familiar brands") is a statement about the 31%, not about the whole.

Two thirds of what the model goes looking for has no brand name in it.

The study also supplies its own counterexample, and it is worth knowing. Asked what payment provider a startup should use, the model searched Stripe, PayPal and Square, all remembered, and then reached one step further and searched Lemon Squeezy by name, a brand it never recalled. The authors' takeaway: "Search can surface a brand that memory never names... Memory is the more reliable way in, but it isn't the only one."

What this means for your dealership

The "we'll get your store into ChatGPT's memory" pitch is selling you a training cycle. If a vendor is promising to build your rooftop's presence in a model's category recall by Q4, hold their proposal next to the study's own definition of how that gets built: press coverage, analyst attention, partnership signals, sustained national category association, absorbed at training time. That is a multi-year corporate awareness program, and it is hard to name a single-store budget that has ever moved a national recall list. Ask any vendor a simple question: are you working the 31% or the 69%?

The 69% is the fight you already know how to have. Generic category fan-out looks like "Chevy dealer Springfield service hours," "certified pre-owned Tahoe inventory near me," "2026 Equinox lease versus finance." Those are won with the same things that have always won local search: complete and current inventory and service pages, a Google Business Profile that matches reality, real local content, fast crawlable pages. AI search raised the stakes on that discipline without replacing it.

In the brand-led 31%, let your franchise's memory do the work. When an automotive fan-out query does name a brand, the study says it overwhelmingly names one the model already carries, and in your category those names are the manufacturers. Chevrolet, Ford, Toyota, Honda are already in the recall set. Those brand-named searches still have to land on a page. "2026 Silverado 1500 LT towing capacity," "Tacoma TRD Off-Road versus SR5," "F-150 Lightning charging at home" are all queries a dealer's content can legitimately answer better than a manufacturer brochure page. Ride the recall you already have instead of trying to build one you never will.

Content still gets you into the fan-out. Lemon Squeezy is the existence proof. A brand the model had never heard of got searched by name because there was something on the web worth searching for. That is the door, and it is open.

What to do about it

1. Pull your top 20 organic pages and read them as a machine would. Do they answer the specific question a shopper asks, in the words they ask it, with the brand and model and trim named in the body copy? Or do they say "we" and "our dealership" and "contact us today"? Same problem we wrote about yesterday from the mention side.
2. Build comparison and buying-decision content around your franchise's model line. Trim comparisons, towing and payload breakdowns, lease versus finance math for your state. This is the material that answers a brand-led fan-out query.
3. Audit the boring local stuff this week. Google Business Profile accuracy, hours, service menu, VDP crawlability. Most of the model's searching is generic, and generic is local.
4. Reprice any "AI brand authority" line item on your marketing invoice. Ask what it does inside the next 90 days that your local SEO does not already do.
5. Do not overreact to a single measurement. The same caution applies here as to every AI visibility dashboard on the market, for reasons we have covered before.

The honest version

One vendor, one search model, 66 prompts, twelve days. Nobody has measured any of this on dealership queries yet, and the first team that does will probably find something different in the details.

The direction is hard to argue with though. Models bring a shortlist to the search, and your rooftop is not on it. The good news, which almost none of the coverage led with, is that two thirds of the time the model skips the shortlist entirely and runs the same kind of local category search your dealership has been competing in for twenty years.

If you want to know how your store actually shows up in that 69%, book a walkthrough and we will look at your pages with you.