Compare Gemini vs ChatGPT local search results and the two engines answer from different parts of the web. Across 1,487 identical queries they named the same top business 4.2 percent of the time, and their cited domains overlapped just 8 percent. The source mix behind each engine also keeps moving.
That gap has practical consequences for any business trying to work out whether it is visible in AI search, and for anyone selling a single answer to that question.
Gemini and ChatGPT cite different parts of the web for the same local search
The two engines draw from different parts of the web entirely. Running the same 1,487 local queries through both on July 30, 2026, Steady Demand found Gemini citing the business's own site 59.9 percent of the time against ChatGPT's 15.9 percent, while ChatGPT pulled 41.7 percent of its citations from social and community forums and 34.6 percent from general business directories, against Gemini's 14.4 and 10.4 percent (Steady Demand).
Three weeks later the mix had moved. A re-run of the identical design on August 20 found ChatGPT's citations to reddit.com had gone from 41.7 percent to zero, with the share redistributed to businesses' own sites, up from 15.9 to 42.8 percent, and general business directories, up from 34.6 to 40.8 percent. Gemini's Reddit share was unaffected (Steady Demand).
Two things survive that revision. Both engines now lean hardest on a business's own site, so the website carries weight in either one. And the engines still disagree with each other on which sites to cite, which is why a strategy tuned to one of them tells you very little about the other. The off-site half of the work is what we lay out in off-page SEO for AI search.
Grounding drift means the same local search returns different citations
Ask an AI engine the identical question twice and the sources change. Steady Demand measured byte-identical repeat queries and found cited-domain overlap of 0.463 on back-to-back calls, 0.410 about three and a half hours later, and 0.265 across the next day's own repeats (Steady Demand). Same words, same day, different answer.
The cause sits upstream of retrieval. The search strings the model generated internally to go and find those sources overlapped only 0.011 to 0.056 across rounds, far less than the citations that came out the other end (Steady Demand). Asking the same question twice does not run the same search twice.
Google's local pack stayed stable while Gemini's recommendations moved
The obvious objection is that local search is just noisy. A control test answers it. Across 500 metro and vertical combinations run on the same cadence, Google's classic local pack returned the same top listing 90.2 percent of the time while Gemini named the same top business 7.9 percent of the time (Steady Demand).
Conventional local ranking is not the unstable part. Whatever produces the variation belongs to generative answer synthesis, which means your Google Maps position and your AI answer visibility are now two separate things to manage. The traditional half still runs on the fundamentals in your Business Profile categories.
One AI visibility check is closer to a coin flip than a verdict
The single most useful takeaway for an owner-operator concerns how you test. A business that asks ChatGPT once, sees no mention, and concludes it is invisible may simply have caught one draw of a stochastic process, and the same is true in reverse for a business that checks once and relaxes (Steady Demand).
So build the check into a routine. Repeat the identical question at least twice, try one genuinely different phrasing, and log the results on a cadence. Treat any single screenshot as one data point. That measurement discipline is the same one behind share of LLM as a metric.
A single local AI search strategy will not cover both engines
The engines barely agree on anything. Across the identical query set, the cited domains matched only about 8 percent of the time between Gemini and ChatGPT, and the two named the same top business in 4.2 percent of cases (Steady Demand). Optimizing for one tells you very little about your position in the other.
Practically, that means a two-track plan. Track one is your own site: accurate, crawlable, structured, and specific about services and locations. Track two is everywhere else people describe you: directories, review platforms, and community discussion, whose relative weight in ChatGPT moved sharply between the July and August measurements. Both tracks feed the broader answer engine optimization work.
Your own website is the most drift-resistant local citation you control
One finding is good news. The business's own website was the largest single citation category in the dataset, and it is the only source in the mix that a business fully controls without renting it from a third-party platform (Steady Demand). Everything else depends on someone else's indexing behavior on top of the model's.
Which makes the site the first investment, not the last. Clear service pages, real location pages, accurate hours and contact details, and content that answers the question a searcher actually asked. Reviews still carry weight in how you are described once you are found, as we covered in Google reviews for multi-location operators.
Measuring local AI visibility without fooling yourself
Build a small, fixed query set and stick to it. Pick the six to ten questions a real customer would ask, write them down word for word, and run them monthly against both engines, recording which businesses and which domains get cited. The study's own scope gives you the template: 1,487 queries across 50 metro areas and 10 service categories, repeated across rounds (Search Engine Land).
Then judge the trend and ignore the individual result. A month where you appear in four of ten runs and the next where you appear in six is progress, and neither number means anything on its own. If you want help setting up that tracking for your locations, we can build the query set with you.
