Ask Gemini the same local service question a day apart and roughly three quarters of the sources it quotes have changed. Google's ordinary local pack returned the same top business 90.2% of the time in the same test.

The author ran a 41-day controlled experiment on a Brazilian business directory of roughly 2,400 pages. Twenty-one sections were split into two groups, 11 reached by hard-coded HTML links and 10 reached only by links injected with JavaScript, giving 1,062 pages in the test hierarchy (748 HTML-linked, 293 JavaScript-linked). GPTBot reached 748 HTML pages and zero JavaScript-linked pages. ClaudeBot, the same: 748 and zero. Amazonbot reached 735 and zero. Bingbot reached 282 HTML pages and 3 JavaScript ones. Google was the outlier in both directions: Googlebot reached only 5% of the HTML pages and 2% of the JavaScript ones during the window, while GoogleOther reached 66% and 48%.

Caveats worth stating: this is one site, run by the author, on a directory rather than a service business, with no third-party replication. It is a well-designed single test, not a study.

Method first, because it decides how much of this we can use. Data was collected 27 and 28 July 2026. The design was 50 metros by 10 service verticals by 3 phrasings, 1,487 queries completing out of an intended 1,500 (13 lost to API failures), yielding 14,472 citations across the 50 largest US metros and 10 service categories. Gemini was queried through the API as `gemini-flash-latest` using the Google Search grounding tool, capturing grounding chunks and the model's own web search queries. ChatGPT was run through DataForSEO's AI Optimization endpoint rather than the consumer product.

The source mix diverges sharply. For Gemini: the business's own site about 60%, social and community 14.4% (Reddit alone 13.7%), general business directories 10.4%, local service directories 10.3%. For ChatGPT: social and community 41.7%, general business directories 34.6%, the business's own site 15.9%.

Agreement between engines was close to absent. Across the 1,487 identical queries, average cited domain overlap was 8.3% and the two named the same top business 4.2% of the time.

The most useful part is a separate, smaller repeat test. Six fixed queries were run across four rounds, 72 calls total, in New York over about 20 hours. Cited domain overlap was 46.3% back to back, 41.0% at plus 3.5 hours the same day, and 26.5% the next day. Overlap in the search strings the model wrote for itself was 0.056, which points at the grounding step rather than at ranking. As a control, Google's conventional local pack returned the same top listing 90.2% of the time under the same repeat conditions.

Caveats that must travel with these numbers. This is a Gemini API grounding test, not the consumer Gemini app and definitely not AI Overviews, so it cannot be quoted as an AI Overviews finding. Gemini averages 8 to 10 citations per query against ChatGPT's roughly 1.9, which mechanically depresses any domain-overlap measure between them, so the 8.3% figure is softer than the 4.2% top-business figure. More than 400 long-tail domains were classified as "business's own site" without individual verification, so the 60% headline is the softest number in the set. The repeat test is 6 queries in one city and the authors themselves call it "a small, deliberately narrow experiment". The publisher is a local SEO agency that sells Google Business Profile management, so a finding that business websites dominate is a finding that suits them. The data is also three weeks old and predates the 14 August ChatGPT source shift logged yesterday.