A question we get constantly: "so how do I rank on AI?" — as if there were one AI. Our sweeps say otherwise. Each engine family assembled its answers from a recognisably different diet of sources, and the businesses that surfaced differed accordingly.
- Gemini (with Google grounding) reads the local editors. Urban List, Denizen, Heart of the City — its recommendations track local editorial coverage more than any other engine's. It also consistently surfaced the most businesses per query in our runs.
- OpenAI's web search leans on aggregators and hard facts. TripAdvisor, Yelp, directories — fewer picks, more verification-flavoured. At the other end of its model range we hit a quirk worth knowing: in our plumber run, the cheapest tier sometimes refused to name businesses at all, answering in generic advice. The model tier a user is on changes whether you exist.
- Claude reads the community. Reddit threads, forum consensus, community institutions. (One methodological honesty note: Claude's cited sources in our runs are agent-self-reported rather than transcript-verified, so we treat its source mix as indicative, not measured.)
- Siri's local layer is Apple Maps — powerful in food and retail, blind elsewhere. It surfaced 25–30 venues per restaurant vertical, more raw coverage than any LLM. Then we asked about plumbers: 0 of the 67 plumbers any other engine surfaced. Board-game shops: 1 of 19. If your vertical isn't a strong Maps category, the Siri shelf effectively doesn't exist. (We measured this layer as one Maps search per vertical rather than per query, so nuanced intents collapse there by design.)
Same city, different shelves
The overlap between engines was small enough that "AI visibility" as a single number is close to meaningless. A restaurant beloved by local editors can dominate Gemini and barely register on OpenAI; a shop with a devoted subreddit presence does the reverse on Claude. Two calibration notes: every reach figure here is single-run — in our repeat testing the ordering between engines held while exact counts wobbled, so trust the contrasts, not the points. And these are API instruments; a logged-in consumer app layers your personal history and A/B buckets on top of everything we measured. Query wording mattered less than you'd think — in our barber run, five different phrasings of the same need converged on nearly the same list — but which engine and which intent mattered a lot (asking for a board game gift was the only intent that pulled answers away from the specialist shops).
Check yourself on more than one engine. A single ChatGPT screenshot tells you about one shelf out of at least four.
Match effort to engine diet: local editorial gets you into Gemini's world, review platforms and directories into OpenAI's, genuine community standing into Claude's — and your Apple Maps listing is the entire game for Siri, so make sure it's complete and current.