A new citation study puts a number on something local SEOs have suspected for months: what an AI engine cites today, it often will not cite tomorrow. Steady Demand, a local-SEO analytics vendor whose co-founder Ben Fisher led the research, tracked 14,472 citations drawn from 1,487 local queries, spread over ten service categories and the fifty biggest metro markets in the United States. The team ran the same queries through Google’s Gemini and OpenAI’s ChatGPT to compare citation patterns across platforms. The headline number, nearly 60% of Gemini’s citations pointing to business websites, matters less than what happened when the same query ran twice.
Close to 60% of what Gemini cited was a business’s own site, the study found, a bigger slice than directories, review platforms and forums held between them. Reddit took second place on 13.7%, outranking every local-service directory put together, Angi and Thumbtack and HomeAdvisor among them. For operators who assumed local citations still ran mostly through structured directory listings, that ordering inverts the priority: a business’s own site drew more Gemini citations than the platforms built to aggregate business data.
Fisher named the deeper problem “Grounding Drift.” Ask Gemini the identical question twice and roughly 40% of the cited sources carry over. The business it recommends first survives a rerun only about 7% of the time. Google’s local pack, used as a control, held its number one listing across repeats at close to 90%.
Running the same 1,487 queries through ChatGPT produced a different pattern. Overlap between the two engines was slight: identical domains surfaced in 8% of cases, and the same business came out on top in only 4.2%. Where Gemini leaned on business websites, ChatGPT relied far more on Reddit and business directories, the study found.
Fisher’s own framing counsels caution. One run of an AI search, in his description, amounts to a snapshot and should not be trusted to describe visibility over any real span of time. That distinction matters because it draws a line this study does not cross: unstable citation behavior is not the same claim as random citation behavior, and nothing in the findings shows Gemini choosing sources without a pattern, only that the pattern shifts run to run.
The study is a vendor-run analysis, not an independently replicated one. Steady Demand has not said whether outside researchers have reproduced the findings, and the reporting discloses query count and metro coverage but no further methodology.
The instability finding follows a report Search Engine Land published Aug. 13: local-pack AI Overviews had cited one lawyer’s self-promotional listicle as the source for multiple competing firms. That case was a single documented instance, not a measured pattern, and it involved a different generative engine and a far smaller sample than Steady Demand’s 1,487-query run. Together, the two point at the same operational problem: citation sourcing in generative local search is not yet consistent enough for a single spot-check, in either direction, to stand as proof of visibility.
Local SEOs tracking AI citations should stop treating any single Gemini or ChatGPT check as a status update. A monitoring practice needs repeated sampling of the same query, across both platforms, before a missing or present citation is worth acting on.
Search Engine Land reported these findings, in an article by Danny Goodwin, on Aug. 19, 2026.
