Gemini Named the Same Top Business 8% of the Time. Google's Map Pack Did It 90%.

Somebody finally ran the control test that everybody selling AI visibility has been avoiding.

Ben Fisher of Steady Demand collected 14,472 AI citations from 1,487 local search queries covering all 50 of the largest U.S. metros and 10 local-service categories, then did the obvious thing nobody bothers to do. He asked the same question again. And again. Across 500 metro and category combinations and 5,000 pairwise comparisons, Gemini named the same top business 7.9% of the time.

Then he ran the identical repeat test against Google's classic local map pack as a control, 4,466 pairwise comparisons on the same combinations. The map pack returned the same top listing 90.2% of the time.

Same city. Same question. Same minute. One system is a standings table. The other one reshuffles the deck on every ask. Search Engine Land wrote it up on August 19, and the number that traveled was the good news about business websites. The comparison underneath it is the part that should change how your store spends money.

What the study actually did

The main dataset was collected on July 27 and 28, 2026, across ten categories: plumber, roofer, HVAC contractor, electrician, locksmith, pest control, house cleaning, personal injury lawyer, dentist, and auto repair shop. It logged 4,216 unique businesses and 3,611 unique domains, with a 93% Google Places match rate on business verification.

No car dealerships in that list. Auto repair shop is as close as it gets, and that matters, because your fixed-ops department competes in exactly that query space against every independent shop in your market. Read the percentages as directional for your store rather than as your store's numbers.

Now the finding that made the headlines. Of everything Gemini cited, 59.9% pointed at the business's own website. That is more than every directory, forum, and review platform in the study combined. Reddit came second at 13.7%, which by itself beat all the local-service directories put together at 10.4%.

Run the same 1,487 queries through ChatGPT and the picture inverts. ChatGPT sent 15.9% of its citations to business websites, 41.7% to social and community forums, and 34.6% to general directories. That lines up with what we covered when ChatGPT was answering "dealer near me" from Yelp and Google Maps feeds rather than from dealership websites.

The two engines agreed on cited domains 8% of the time and named the same top business 4.2% of the time. Auto repair was the closest-aligned category in the entire study at 14.9% overlap. Personal injury law came in at 0.5%.

Why the drift number is the important one

Fisher's own description of the mechanism is the cleanest explanation anyone has written: "asking the same question twice doesn't run the same search twice. Each call is its own independent, evidently stochastic decision about what to type into Google on your behalf."

The decay curve he measured tells the rest of it. Repeat a query back-to-back and the cited domains overlapped 46.3%. Wait three and a half hours and it dropped to 41.0%. Ask the next morning and it was down to 26.5%. Reword the same question three different ways and agreement on sources sat around 40%.

So there are two separate lotteries running. Which sources get cited drifts by roughly half. Which business gets named as the top pick drifts by more than ninety percent.

We have seen this movie, and dealers lost the first screening

For fifteen years, local search gave the car business a scoreboard. You were third in the map pack for "Ford dealer near me," your competitor was first, and everybody in the tower could look at the same screenshot and argue about it. That stability is what made rank tracking a product category, and it is what made "where do we rank" a coherent question at a Monday manager meeting.

The map pack still works that way. The 90.2% control number in this study is the proof.

The assistant layer being built on top of it behaves differently, and the industry selling to dealers has not adjusted. Vendors are still shipping the same artifact: a screenshot, a score, a single-query snapshot showing whether AI mentioned your store. Fisher's blunt summary is that "a single query snapshot isn't a reliable read on your AI visibility." At 7.9% consistency, a screenshot showing your store named first carries about as much information as one showing you absent. Both are one pull of a lever.

We have made this argument about AI visibility scores before. This is the first time somebody has put a control group next to it.

What this means for your dealership

Your website went back to being the asset, at least on Google's side. For years the honest advice was that your Google Business Profile and third-party listings carried more weight than your own site for local discovery. In Gemini's citation mix, your website is 59.9% of the answer. Every page a shopper would need, service pricing, hours by department, what a specific package includes, staff and warranty details, needs to be real crawlable HTML on your domain instead of buried in a widget or a PDF. This is the same conclusion we reached running 1,100 car shopping searches through AI Overviews, now backed by a much larger local dataset.

Stop paying for AI visibility screenshots. If a vendor sends a monthly report built on single queries, the number in it is dominated by noise. Ask them how many repetitions per query they run and how they report variance. If the answer is one, you are buying a coin flip with a logo on it.

Gemini and ChatGPT are two different channels. An 8% domain overlap means work that gets you cited in one has a weak relationship to the other. Even auto repair, the best-aligned category tested, only hit 14.9%. Budget and measure them separately.

Reddit is doing more work than the directories. At 13.7% against 10.4% for all local-service directories combined, community discussion outweighed the listing sites. For a franchised store that means the model-specific and market-specific threads where people ask which dealer to use are load-bearing. You cannot astroturf that, and you should not try. You can make sure the service and sales experience people describe there is worth describing.

Reviews get you into the pool and do not pick the winner. This is the counterintuitive one. AI-recommended businesses averaged 4.75 stars against 4.84 for a plain Google baseline, and 97% were rated 4.0+ against 99.6% for the baseline. The AI applied a slightly looser quality bar than ordinary search did. Meanwhile 35% of recommended businesses had 500 or more reviews. Volume gets you considered. Nothing in this data says a tenth of a star buys you the top slot.

Local still means local. Only 2.0% of businesses, 87 of them, were recommended in more than one metro. The national platforms that appeared in all 50 metros were Angi, HomeAdvisor, BBB, ConsumerAffairs and Reddit. Your competition for these answers is the store across town.

The caveats, and they are real

The study queried Gemini through the API as gemini-flash-latest, not through the consumer Gemini app and not through AI Mode in Search. Fisher says so in the methodology and lists it as a limitation: API behavior may differ from consumer apps. The ChatGPT side came through DataForSEO's scraper endpoint rather than direct API calls. It is one snapshot, two days in late July.

That cuts both ways. The consumer surfaces carry personalization, location signals and conversation history that could stabilize results well above 7.9%. They could also be running the same stochastic retrieval with a friendlier wrapper. Nobody outside Google knows, and anybody telling your store they do know is guessing.

What survives the caveats is the shape of it. The retrieval step is probabilistic where the map pack's is close to fixed, and no amount of optimization changes that property of the system.

What to do about it

1. Get your specifics into crawlable HTML this quarter. Service menu and pricing, hours per department, financing terms, what each trim and package includes. Real text, real tables, no widget required to render it.
2. Change what you ask your vendor for. Require repeated sampling with variance reported, or drop the AI visibility line item. One query is too small a sample to call a measurement.
3. Measure Gemini and ChatGPT separately. Different source mixes, 8% agreement, different work.
4. Audit your Google Business Profile anyway. The map pack is the stable 90% surface and it is still where most local intent lands.
5. Fix the things people write about. Reddit outranked every directory in this dataset. That thread gets written by your service drive rather than your marketing budget.
6. Re-check quarterly, and never off one screenshot. A number that drifts 20 points overnight cannot support an annual review.

The useful reframe here is that the assistant layer stopped being a leaderboard you climb and started being a distribution you show up in. You cannot own position one when position one is redrawn every time somebody asks. You can be one of the handful of businesses in your market with clean data, real content on your own domain, deep review volume, and a reputation people describe in public, so that whichever way the deck shuffles, your store keeps landing in the answer.

That is less satisfying than a rank report. It also happens to be the only thing the data supports.

If you want a read on what your store's website actually gives an AI assistant to work with, book a walkthrough and we will go through it with you.