How AI recommends local businesses
We asked ChatGPT, Claude, and Gemini to name the best local business 1,350 times. The models agreed 6.5% of the time.
We asked three web-connected AI engines — OpenAI (the model behind ChatGPT search), Anthropic’s Claude, and Google’s Gemini — the same simple question a customer would ask: “What’s the best med spa / personal-injury law firm / digital-marketing agency in [city]?” We asked it 1,350 times, across 15 US cities of three sizes, five times each, and recorded every business named and every source cited.
The result is the first apples-to-apples map of how AI answer engines pick local businesses — who gets named, where the engines disagree, which sources they trust, and how often they refuse to answer at all.
Method (the short version)
- Engines (all web-grounded): OpenAI gpt-4o-mini-search, Anthropic Claude Haiku 4.5 + web search, Google Gemini 2.5 Flash + Search grounding.
- Categories: med spas, personal-injury law firms, digital-marketing agencies.
- Cities (15): 5 large (New York, Los Angeles, Chicago, Houston, Miami), 5 mid (Tampa, Pittsburgh, Cincinnati, Richmond, Boise), 5 small (Asheville, Savannah, Boulder, Bend, Springfield MO).
- Two phrasings per category, 5 runs each, default sampling — we wanted the natural run-to-run variability a real user sees.
- 1,350 total calls (1,225 web-grounded). We sent no system prompt — coaxing the models to “be specific” would have masked how often they refuse.
Finding 1 — the engines barely agree
Across 45 category-by-city match-ups, the average overlap between the three engines’ business lists was 6.5%. They are, in effect, three different search engines.
But it’s not random noise. In 91% of cities there was at least one business all three engines named — they reliably agree on one top pick and then diverge on everything else.
Gemini named 27 different med spas; ChatGPT named 11; Claude named 9. All three agreed on exactly three: Lux Med Spa Brickell, Monaco Med Spa, and Avabello Aesthetics. Everything else was unique to one engine.
“Rank on AI” is the wrong mental model — there is no single AI ranking. Optimizing so ChatGPT recommends you does almost nothing for your Gemini visibility. You’re being judged by three different juries that agree only on the obvious front-runner.
Finding 2 — nobody owns AI search yet
We expected winner-take-all. We found the opposite.
| Vertical | Top business’s share | Top-3 share | Distinct businesses / city |
|---|---|---|---|
| Med spas | 12% | 32% | ~32 |
| Personal-injury law | 10% | 25% | ~46 |
| Digital-marketing agencies | 8% | 19% | ~55 |
Even the single most-recommended business in a city captures only 8–12% of mentions, drawn from a pool of 30–55 different names. Digital-marketing agencies are the most fragmented of all — the “best agency in town,” according to AI, is a coin-flip among dozens.
No incumbent has locked up AI recommendations in your market. Unlike Google’s first page, where the top results take most clicks, AI answer-share is spread thin and contestable. Visibility here is winnable right now — the businesses that structure their content for AI today are claiming an open field.
Finding 3 — AI reads your website, not your Yelp page
When we bucketed every cited source, the conventional SEO wisdom — “get on the directories” — mostly broke down:
| Engine | Cites the business’s own site | Major directories (Yelp/Avvo/Clutch) |
|---|---|---|
| ChatGPT (OpenAI) | 77% | ~0% |
| Gemini | 45% | 14% |
| Claude | 26% | 18% |
ChatGPT overwhelmingly cites the business’s own website and almost never a directory. But the picture flips by vertical and engine:
| Vertical | Directory reliance |
|---|---|
| Med spas | Near zero everywhere — own-site dominates (ChatGPT 81%, Gemini 64%) |
| Personal-injury law | Real: Claude 34%, Gemini 20% (Avvo, Super Lawyers) — ChatGPT still 73% own-site |
| Digital-marketing | Real: Claude 24%, Gemini 21% (Clutch, DesignRush) — ChatGPT still 77% own-site |
This is the actionable “where to invest” map. For a med spa, optimize your own website — directories barely register. For a law firm or agency, your own site is what ChatGPT reads, but a strong Avvo or Clutch presence still feeds Gemini and Claude. And across the board: if your own website isn’t legible to an AI crawler, ChatGPT has almost nothing to cite you from.
Finding 4 — the model matters: Claude is the cautious one
ChatGPT and Gemini named specific businesses 100% of the time. Claude refused or hedged (“it depends on your needs…”) far more often — and the rate is revealing:
| Vertical | Claude hedge / refusal rate |
|---|---|
| Med spas | 17% |
| Personal-injury law | 44% |
| Digital-marketing | 22% |
Claude declines to name a personal-injury lawyer nearly half the time — unsurprising for a high-liability category. And it gets more cautious in bigger markets:
| Market size | Claude hedge / refusal rate |
|---|---|
| Large metros | 41% |
| Mid-size | 27% |
| Small towns | 15% |
Which assistant your customer uses changes whether they get a recommendation at all. In a big city, a customer asking Claude for a lawyer often gets “here’s how to choose” instead of a name — while the same question to ChatGPT always returns specific firms.
Finding 5 — Gemini is the fickle one
We ran every query five times and counted how many different businesses each engine named across those five identical runs (fewer = more consistent):
| Engine | Consistency across 5 identical runs |
|---|---|
| ChatGPT | Rock-steady — ~7–8 distinct businesses, regardless of city size |
| Claude | Steady — ~7–11 |
| Gemini | Volatile, and it scales with market size |
In a large metro, Gemini named 24 different med spas and 35 different marketing agencies across five identical queries. In a small town, that dropped to ~20 on average. The market-size stability effect is real — but it lives almost entirely in Gemini:
| Vertical | Gemini churn: large metro → small town |
|---|---|
| Med spas | 24.1 → 11.7 |
| Personal-injury law | 23.1 → 18.3 |
| Digital-marketing | 35.4 → 28.3 |
On Gemini, a single great answer is no guarantee — ask twice and you may not appear. The bar isn’t “get named once,” it’s “get named consistently,” and that’s hardest in big markets on the most volatile engine.
Bonus: the businesses AI already favors
The businesses named by all three engines in a city are its current AI front-runners. A sample:
- Med spas — New York: Tribeca Med Spa, Skinney Med Spa · Miami: Lux Med Spa Brickell, Monaco Med Spa, Avabello Aesthetics · Chicago: Dvida Med Spa, Gold Coast Med Spa · Asheville: Mountain Radiance Medical Spa.
- Personal-injury law — Los Angeles: Panish Shea & Ravipudi · Chicago: Horwitz Horwitz & Associates, Levin & Perconti · Pittsburgh: Edgar Snyder & Associates · Boise: Hepworth Holzer.
- Digital-marketing — New York: SmartSites, Jives Media, The Charles · Los Angeles: Ignite Visibility · Chicago: Comrade Digital · Tampa: WebFX, ROI Amplified.
What this means if you’re the business (or the agency)
- There’s no single “AI ranking” to chase. Track all three engines separately — they agree only on the obvious leader.
- The field is open. No competitor owns more than ~12% of AI answer-share in these categories. This is the land-grab phase.
- Fix your own website first — it’s what ChatGPT reads. Then, for law and marketing, shore up the directories Gemini and Claude lean on.
- Consistency beats a one-off win, especially on Gemini and especially in big cities.
For agencies specifically: this is your service now. “We’ll get you ranked on Google” is becoming “we’ll get you named by ChatGPT, Claude, and Gemini” — three engines, three source diets, measured separately. And: how visible is your own agency when a prospect asks AI for the best agency in your city? In our data, the top agency captured just 8% of mentions. That’s the opportunity.
Caveats (stated plainly, because they matter)
- These are web-grounded API queries. They approximate, but are not identical to, the ChatGPT/Gemini apps or Google’s AI Overviews.
- The city is named in the prompt. API calls have no GPS, so “near me” is not exercised.
- Web-grounded only: Claude chose not to search on ~28% of its queries; those are excluded from the visibility and source numbers (but counted in the refusal rate). ChatGPT and Gemini grounded 100% of the time.
- We used the efficient search models. A flagship spot-check (gpt-4o-search, Claude Sonnet, Gemini 2.5 Pro) is the planned follow-up.
- Five runs per cell. Per-city results are reported as appearances; percentages are reported at the tier level.
- Source buckets are heuristic: “own site” means the cited domain matched a recommended business’s name; the residual “other” includes small local sites we couldn’t classify.
- A dated snapshot — June 2026. Models and the live web change.
This is Volume 1 of the PromptRank AI Visibility Index, and the methodology is reproducible — we’ll re-run it as the models and the web evolve.