What AI searches for when someone asks it to recommend a business

How much each engine searched

A “search” here is one query string the engine issued while answering, as the provider returned it with the response. We did not infer these; they are stored strings.

Perplexity’s zero is a limit of what its API returns, not a finding that it does not search. It cites sources in its answers, so it plainly retrieves. We cannot see the queries, so Perplexity is absent from everything below.

Source. 01_Intel/research/icp_test.json (128 answers) and 05_Engineering/work/runs/cinc_panel.json (60 answers), Vector Index’s own research runs · no client data · 583 of the 595 search strings are distinct.

157 of ChatGPT’s 478 searches named one site to check

26.4% of the searches used the site: operator, which restricts a search to a single domain. They targeted 97 distinct domains across 38 answers. Every one came from ChatGPT; Gemini used it in none of its 117.

Two patterns run through them, and both are verbatim below. In the first, the engine has a company name already and goes to a third party to check it:

• site:bbb.org Cincinnati HVAC BBB Corcoran Harnist Heating Air
• site:bbb.org Cincinnati HVAC BBB JonLe Heating Cooling
• site:bbb.org Cincinnati HVAC BBB Arlinghaus Heating Air
• site:napfa.org Cincinnati OH fee-only advisor truepoint wealth counsel
• site:ohiobar.org certified specialist family relations law Cincinnati attorney

In the second, it goes to the company’s own domain looking for particular words:

• site:plasticsurgerygroup.net Cincinnati plastic surgery group official board certified
• site:theplasticsurgerycenterofnorthernkentucky.com Cincinnati plastic surgeon official board certified
• site:sportysacademy.com Cincinnati flight training official Sporty’s Academy Clermont County
• site:trihealth.com Cincinnati cosmetic plastic surgery official board certified

The word “official” appears in 162 of the 595 searches, 27.2%, and 28 of those also use site:. The most-targeted domains were bbb.org (13 searches), plasticsurgery.org (9) and abplasticsurgery.org (7).

What this does not show: that a page satisfying one of these searches changes the answer. We recorded the search, not its effect. The honest reading is narrower and still useful. On this sample the engine was running named checks against specific sites, so the question “does our own site state plainly what we are and what we are certified in” is answerable rather than rhetorical.

Source. 157 of 595 search strings contain the substring site: · domains parsed from the operator · 38 of the 188 answers · all ChatGPT.

What words the searches contained

Each row counts search strings containing the substring, pooled across engines. These are string matches, which is all they are: a search containing “bbb” is not proof the engine read a BBB page.

“Near me” appears in none of the 595. The engines rewrote proximity into a place name instead, which is the opposite of how the phrase is usually discussed.

The two engines behave differently enough that one test tells you little

47 of Gemini’s 117 searches contain “best”, 40.2%, against 43 of ChatGPT’s 478, 9.0%. Gemini mostly restated the question as a category-and-city query. Its searches, verbatim:

• fiduciary fee only financial planner Cincinnati Ohio
• best fee only financial advisor Cincinnati OH
• top wealth management financial advisors Cincinnati
• best cosmetic surgery centers Cincinnati Ohio

ChatGPT, on the same kind of question, went from the category to named companies and then checked them one at a time. Reading only one engine would leave a business optimising for the wrong step: a category-and-city page for Gemini, a checkable record on third-party sites for ChatGPT.

Source. 117 Gemini and 478 ChatGPT search strings · the “best” counts are substring matches on those strings.

How we counted

• What a search is. One query string the engine issued while producing the answer, returned by the provider alongside the response and stored with it. No search was reconstructed, guessed or parsed out of prose.
• Engines. ChatGPT (gpt-5.4), Gemini (gemini-3.6-flash) and Perplexity (sonar), each through its API with web search enabled. API answers are not consumer-app answers: system prompt, retrieval and routing all differ.
• Location. None was set, and no account was signed in. Where a city appears, it is in the wording of the question only, which is not evidence of where a user is.
• Substring counts. Matched case-insensitively against the search string. A search counts once per row however many times the substring occurs.
• Distinctness. 595 searches, 583 of them distinct strings. Repeats across answers are counted each time, because each one was issued.
• Questions. 48 distinct wordings. The 6 in cinc_panel.json are a subset of the 48 in icp_test.json, checked rather than assumed, so 48 is the union and not a sum. 45 of the 48 produced at least one search.
• Client data. Excluded by construction: only the two research files above were read, neither of which is under the client folder, and every search string was additionally screened against every client name and domain on file. 0 were screened out, and the screen was tested by running it with two substrings that do occur, which removed 478 of the 595.

Limits

• Perplexity is missing from the findings. Its API exposed no searches in 32 answers. That is a gap in what we can see, not a finding about the engine.
• What the provider chooses to show us. These are the searches each API reported. An engine may run searches it does not report, and 46 of the 188 answers exposed none at all. Nothing here is the complete retrieval behaviour of any engine.
• One city, mostly. Most questions name Cincinnati. 10 of the 48 name no city, such as “best executive outplacement firms”, and those produced 63 of the 595 searches. Nothing here says what happens in another metro.
• Dated. 24 and 25 August and 16 September 2026, on the model versions named. Retrieval behaviour is a product decision and can change without notice.
• Searches, not effects. We recorded what the engines looked for. We did not test whether satisfying any of those searches changes an answer, and nothing on this page is a causal result.
• The questions were ours. We wrote them for our own research. They are not drawn from search logs and are not a sample of what buyers ask.
• Not measured: how many businesses each answer named. The stored research file carries a per-answer list of firm names, and it is not fit for that purpose, so we have not published a count. For one home-services answer it holds 171 entries including a concrete contractor and an electrician, neither of which appears in an answer about replacing a furnace. It is a harvest from the retrieval layer, not from the answer. Counting named businesses properly needs an extractor whose output has been read against every answer, and this pass did not do that.

Every search we recovered

All 595, one per row, with the question that produced it, the engine, the model, the date and the substring flags used above. The counts on this page can be reproduced from this file alone.

Download the data (CSV, 595 rows)

Some rows name real local businesses, because the engine named them in its own search. Their presence says only that: the engine searched for them on the date shown. It is not a statement by us about any of them.

Which sites AI engines cite · How we measure · Why ChatGPT does not mention a business · Our scope and prices