Which sources ChatGPT, Perplexity and Gemini cite when people ask who to hire (2026)

What I take from it

What should a business do with this?

Start with the sites the engines already read for your category. If you run a local service or a travel business, that means the directories and review sites: in this sample they were 37.1% and 50.3% of the citations. If you sell software, it means the “best X” and “X alternatives” articles and your own docs.

Then fix your own site so it names you plainly. Company sites got half the citations, but a cited company was named only 26.7% of the time. ChatGPT named the cited company most often (46.6%), Gemini next (30.0%) and Perplexity least (16.3%), partly because Perplexity cites so many more pages per answer.

Last, never judge one answer. With names changing on every run, a single screenshot of ChatGPT naming you, or not naming you, says very little. Ask each question at least three times and count a win only when you show up in most runs.

Method

How I measured it

Engines and models

• ChatGPT: the OpenAI Responses API, model gpt-4o, with the web search tool required on every call. 1,206 answers.
• Perplexity: the Perplexity API, model sonar, which searches live on every call. 1,179 answers.
• Gemini: Google Vertex AI, model gemini-2.5-flash, with Google Search grounding on. 733 answers. Gemini was not switched on for every question set that week, so it has fewer answers.

These are the APIs, not the consumer apps. Nobody was signed in, so no answer was shaped by a personal history. The question was the whole input, with no system prompt.

Questions, runs and dates

280 distinct questions, written the way a buyer types them: “best HVAC company in [city]”, “who should I hire to finish my app”, “best LLM gateway”. They come from 25 question sets we track for clients, prospects and category scans, grouped into six categories. Each question went to each engine three times in a row. Answers were collected from 2026-09-24 to 2026-09-30 across 41 probe runs.

Cleaning

• The store held 8,042 rows. 3,895 were answers from one of the three engines, from a live probe, with the model recorded and the cited links kept. Older imports that did not keep cited links were left out.
• 126 answers about a local election were dropped: they are not a buying question.
• 651 rows were same-day cached copies of an answer already counted (two question sets asking the same question on the same day), so they were dropped. A further 0 rows repeated the exact text of an answer already counted and were dropped too.
• That left 3,118 answers. 26 of them cited nothing.

How a citation is counted

A citation is a URL the API itself returns as a source: OpenAI's URL citation annotations, Perplexity's citations list, and Gemini's grounding sources (resolved from Google's redirect links to the real page). The same URL twice in one answer counts once. Links that only appear inside the answer text are not counted.

How a source type is assigned

By the URL, with a written rulebook: known lists of directories, review sites, marketplaces, news outlets, forums, video and social sites; “best”, “top N”, “alternatives” and “vs” in the page path for listicles; docs subdomains and paths for docs. I read the most-cited domains the rules left untyped and added the directories, job boards, local news sites and research papers among them by hand. Anything left over counts as a company or vendor site, so that bucket also holds some independent blogs.

Named vs cited

For each cited company site, I took its domain name without the ending (“acmeair” for acmeair.com) and checked whether it appears in the answer text, ignoring spaces, case and punctuation, after removing the links. This used the full answer text for 2,950 answers; answers where only a snippet was stored were left out of this one measure.

Run to run

Cited domains: every question and engine with three fresh runs in the same probe (970 sets) was checked for an identical set of cited domains. Named businesses: on four local-service leaderboards (360 answers, probed on 2026-09-26 and 2026-09-28), a list of the business names in the answers was built first, then each run was checked for which of them it named.

Privacy

The questions were written for businesses I work with and for prospects. This page reports categories and domains only. Our clients' and prospects' own sites (17 domains, 594 citations) are counted in every total but left out of the domain table. Travel sites cited only for the travel questions are left out of the table too.

Limitations

What this study cannot tell you

• The questions are not a random sample. They were chosen because our clients, prospects and category scans care about them. Local services (1,032 answers) and travel (757 answers) are a large share, and that shapes the totals and the top-domain list.
• APIs are not the apps. The ChatGPT, Perplexity and Gemini apps can use other models, other search settings and your history. Treat these numbers as what the engines read, not a screenshot of what you will see.
• One week, one set of models. Engines change their search stacks often. A re-run next quarter may differ, which is why the script is saved and dated.
• Source types come from rules, not a person reading every page. Some pages will be in the wrong bucket, mostly between company sites, blogs and listicles.
• The named check is strict. It looks for the domain name in the answer, so a business whose name differs from its domain is missed by this check.
• The name check across runs covers local services only, the four categories where we had a verified list of the businesses named.
• Gemini has fewer answers (733) than ChatGPT (1,206) and Perplexity (1,179), because it was not switched on for every question set that week.

Check your own business

The AI Answer Check runs the same probe on your buyers' questions: 12 to 15 questions, three runs each on ChatGPT, Perplexity and Gemini, with who gets named instead and which sites they read. It costs $299. For how to act on it, see how to get your business mentioned in ChatGPT and how to get cited by Perplexity.