ChatGPT gets its information from two places: what its models learned in training, from public web pages, partner data and people, and a live web search it runs through outside search providers when a question needs current facts. For a local business the live search is the part you can see and act on, and its sources are not the obvious ones: across 9,434 ChatGPT answers to home-services questions in 100 US and Canadian metros between August 31 and September 20, 2026, ChatGPT cited Expertise.com in 32.1% of answers and Angi or HomeAdvisor in 29.4%, but Yelp in 0.6%.
That matters if you run 40 locations, because the pages deciding whether ChatGPT names a branch are mostly pages you did not write. Of the ChatGPT answers that cited anything, 84% cited at least one directory, review platform or "best of" list.
Last verified: September 23, 2026.
The numbers come from the Cheers Market Baseline, which asks four AI engines the same questions every week; the method is at the end of this page. To check one of your own locations against live AI answers, run the free Cheers AI Visibility Grader.
What OpenAI says about where ChatGPT's information comes from
OpenAI documents both halves, in two help center articles updated in August 2026.
The training half is described in "How ChatGPT and our foundation models are developed." OpenAI says its models learn from three primary sources: information publicly available on the internet, information it accesses through partnerships with third parties, and information that users, human trainers and researchers provide or generate. It adds that it increasingly uses synthetic data, and that site owners can keep their pages out of training by disallowing GPTBot in robots.txt.
The live half is described in "Searching the web with ChatGPT." ChatGPT may search the web automatically when a question would benefit from current information. When it does, it rewrites your question into one or more targeted queries and sends them to search providers, and the page points readers to Microsoft's and Shopify's privacy policies for how those providers handle the queries. It also estimates a general location from your IP address, so OpenAI's own example turns "good restaurants near me" into "top restaurants San Francisco." For local questions, OpenAI says ChatGPT may share that location with trusted third-party providers that supply nearby business listings, and on iOS and Android the answer can include a map.
What OpenAI documents
Two sources, one answer
- 01Training
Trained knowledge
Public web pages, data from partners, and data from users, trainers and researchers, learned before your question
- 02Live
Decide to search
ChatGPT may search automatically when current information would help
- 03Live
Rewrite the query
Into targeted queries, with a general location taken from your IP address
- 04Live
Ask search partners
OpenAI points to Microsoft's and Shopify's privacy policies for this step
- 05Answer
Cite the pages
Cited links in the answer and a Sources list you can open
- 06Answer
Add the map
On iOS and Android, a map with business listings from third-party providers
OpenAI Help Center: Searching the web with ChatGPT; How ChatGPT and our foundation models are developed. Both updated August 2026, read September 23, 2026.
- How ChatGPT answers a local question, from OpenAI's help center
- 1. Trained knowledge: Public web pages, data from partners, and data from users, trainers and researchers, learned before your question
- 2. Decide to search: ChatGPT may search automatically when current information would help
- 3. Rewrite the query: Into targeted queries, with a general location taken from your IP address
- 4. Ask search partners: OpenAI points to Microsoft's and Shopify's privacy policies for this step
- 5. Cite the pages: Cited links in the answer and a Sources list you can open
- 6. Add the map: On iOS and Android, a map with business listings from third-party providers
OpenAI does not name the business-listing provider behind the map. When it launched ChatGPT search on October 31, 2024, it said only that it had partnered with news and data providers for categories including weather, news and maps. Outside testing fills in part of the picture: in May 2025, Search Engine Land's Damian Rollison found that ChatGPT pulled web results from Bing's index for local questions but ordered them with its own logic, and that it did not see the profile details businesses manage in Bing Places. On ranking, OpenAI's help article gives site owners one sentence:
ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.
The one thing OpenAI tells site owners to do is technical. Allow OAI-SearchBot in robots.txt and make sure your host or CDN is not blocking OpenAI's published search IP addresses. OpenAI's crawler documentation is blunt about the alternative: sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. GPTBot is the separate training crawler, so blocking it does not remove you from search.
So for a question like "who are the best plumbers for water heater repair or replacement?" asked from a phone in Denver, the answer you can inspect is built from two visible layers: the pages ChatGPT cites and the business cards on its map. Those are the two layers we measured.
What sources does ChatGPT use for local businesses?
Directories and ranking lists, then contractor websites, then almost nothing else. In the three weeks we measured, ChatGPT's 9,434 answers carried 43,785 citation links. Named platforms led: Expertise.com in 32.1% of answers, Angi or HomeAdvisor in 29.4%, HomeStars in 16.0%, BBB in 14.5%, Thumbtack in 5.0% and Reddit in 4.9%. A long tail of other directories and "best of" list sites, a group that includes Three Best Rated, Best Pick Reports, HomeGuide, Yellow Pages and dozens of smaller directories, showed up in 59.9% of answers. Sites that rank local companies, the "best of" list format, appeared in 64.2% of the ChatGPT answers that cited anything.
Contractor websites were the other big source. At least 44.2% of ChatGPT answers cited a website belonging to a business in our dataset, and those sites took 26.8% of all ChatGPT citation links, about the same share as the long tail of directories. The trade changes the split. In plumbing, 53.7% of ChatGPT answers cited a plumber's own site; in roofing, only 17.8% of 600 answers cited a roofer's site while 40.0% cited Expertise.com.
Google and Yelp barely registered. ChatGPT cited a Google page in 0.1% of answers and Yelp in 0.6%, 52 answers out of 9,434. Yelp shows up somewhere else: in the five days from September 18 to 22, when we captured ChatGPT's map in full, 34% of the rated business cards on it linked to a Yelp business page, and 98% of the answers showing a business under 4.0 stars involved one of those Yelp cards. We cover that rating gap in why ChatGPT shows a lower star rating than Google.
When ChatGPT did cite a contractor's site, it cited the homepage 37.4% of the time, a service page 24.1%, a page naming a service and a city 17.4%, and a location page 10.7%. Blog posts got 1.2%. We turned that into a page spec in what location pages should include for AI search.
Next step
Is AI recommending your business?
Find out how visible you are across ChatGPT, Gemini, Perplexity, and AI Overviews.
Four engines, four source lists
The same 32 questions, asked in the same 100 metros in the same weeks, produced four very different source lists. Each cell below is the share of that engine's answers citing at least one page from the source.
Four engines, one question set
Who cites what
Cheers Market Baseline, August 31 to September 20, 2026: 32 home-services questions asked weekly in 100 US and Canadian metros, 34,742 usable answers. Aggregate only.
| Source | ChatGPT (9,434 answers) | AI Mode (9,435 answers) | Perplexity (6,509 answers) | Gemini* (4,141 with citations) |
|---|---|---|---|---|
| Business websites | 44.2% | 92.5% | 89.0% | 88.3% |
| Google's own pages | 0.1% | 95.5% | 1.1% | 0.2% |
| Other directories and 'best of' lists | 59.9% | 9.2% | 81.7% | 37.8% |
| Expertise.com | 32.1% | 0.2% | 30.9% | 6.0% |
| Angi or HomeAdvisor | 29.4% | 9.3% | 30.8% | 13.4% |
| BBB | 14.5% | 1.7% | 31.0% | 8.4% |
| Yelp | 0.6% | 19.7% | 13.9% | 0.0% |
| 4.9% | 9.1% | 10.8% | 1.2% | |
| Online booking pages | 0.0% | 31.2% | 0.2% | 0.0% |
ChatGPT is the directory engine, with the long-tail lists, Expertise.com and Angi or HomeAdvisor covered above, and it almost never cited Google (0.1%), Yelp (0.6%) or a booking page (0.0%).
Google AI Mode is the opposite. It cited Google's own pages in 95.5% of answers and a business website in 92.5%, linked an online booking page in 31.2% and Yelp in 19.7%, while Reddit came in at 9.1%, Angi or HomeAdvisor 9.3%, long-tail directories 9.2%, BBB 1.7% and Expertise.com 0.2%. Only 38.3% of its answers with citations cited any directory, review platform or list, against 84.2% for ChatGPT.
Perplexity cites everything at once: business websites 89.0%, long-tail directories 81.7%, BBB 31.0%, Expertise.com 30.9%, Angi or HomeAdvisor 30.8%, Yelp 13.9%, Reddit 10.8%, Google 1.1% and booking pages 0.2%. Across its answers with citations, 93.5% cited a directory, review platform or list.
Gemini is the hardest to read, because we captured citations for only 4,141 of its answers. Within those, it cited business websites in 88.3%, long-tail directories 37.8%, Angi or HomeAdvisor 13.4%, BBB 8.4%, Expertise.com 6.0%, Reddit 1.2% and Google 0.2%, and we saw no Yelp or booking-page citations. Across its answers with citations, 55.8% cited a directory, review platform or list. For the engine-by-engine version of this page, read the source differences between ChatGPT, Gemini, Perplexity and Google AI Mode.
The practical result is that a brand can be strong on one engine and missing on another. On occasions when all four engines answered the same question for the same city in the same week, 77% of the businesses named were named by only one of them. Our AI local ranking factors summary lays out those differences factor by factor, from ratings to booking links.
The US and Canada are different source markets
If you operate on both sides of the border, ChatGPT is reading two different directory sets.
ChatGPT, US and Canada
Two countries, two directory sets
Cheers Market Baseline, August 31 to September 20, 2026: 5,659 US and 3,775 Canadian ChatGPT answers. Aggregate only.
| Engine | US | Canada |
|---|---|---|
| Expertise.com | 53.5% | 0.0% |
| Angi or HomeAdvisor | 49.1% | 0.0% |
| HomeStars | 0.0% | 39.9% |
| BBB | 9.8% | 21.5% |
| Business websites | 36.7% | 55.4% |
Angi and HomeAdvisor appeared in 49.1% of ChatGPT's 5,659 US answers and in 1 of its 3,775 Canadian answers. Since September 16, 2026, Angi has also been testing a paid, labeled chat that opens from ChatGPT ads in the US, which we explain in what ChatGPT Sponsored Agents are. HomeStars fills that role in Canada, at 39.9% of Canadian answers and none in the US. Expertise.com appeared in 53.5% of US answers and none of the Canadian ones; we saw no Expertise.com page for a Canadian city. BBB was cited in 21.5% of Canadian answers against 9.8% in the US, and Reddit in 7.9% against 2.9%, a gap we cover in why Reddit is rarely cited by AI for local businesses.
Canadian ChatGPT answers also leaned harder on contractor websites, 55.4% against 36.7% in the US, and on long-tail directories, 68.0% against 54.4%.
What to do this week
Pick five metros and one money question per trade, the kind a customer asks when they are ready to hire. Then spend the week on the sources that answer it.
- Ask ChatGPT the question for each metro and save two things: the cited links and the business cards on the map. The free AI Visibility Grader gives you a one-location head start across ChatGPT, Gemini and Perplexity.
- Open every directory page it cited: the Expertise.com list for your trade and city in US metros, HomeStars in Canadian ones, Angi, BBB and whatever "best of" sites came up. Check whether each branch is listed and whether the name, phone, service area and hours are right, and fix the profiles you control.
- Open each location's Yelp page and read it the way a customer would, because in US metros it is often the rating ChatGPT puts on its map card.
- Ask your web team to confirm that robots.txt allows OAI-SearchBot and that the CDN is not blocking OpenAI's published IP ranges.
- Make sure each location has a service page with the city in it, so ChatGPT has something of yours to cite besides the homepage.
If the same wrong phone number or closed branch keeps coming back, trace it with the business-information correction workflow.
Methodology
Every week, Cheers asks ChatGPT, Gemini, Perplexity and Google AI Mode the same 32 home-services questions in 100 metro areas across the United States and Canada, and records every business each engine names, with its star rating, review count and map position where the engine shows one, and every source it cites. This analysis covers the three weeks from August 31 to September 20, 2026: 34,742 usable answers (9,434 from ChatGPT, 9,435 from Google AI Mode, 9,364 from Gemini and 6,509 from Perplexity) across 60 US and 40 Canadian metros. We excluded empty or cut-off answers, most of them from a Perplexity collection outage. Each question was asked once per city per week, so week-to-week changes include normal answer-to-answer variation. Business names were extracted automatically; a hand check of 40 answers found the extraction caught about three in four of the businesses a reader would count, so counts are lower bounds. Star ratings and review counts are the figures each engine displayed; for ChatGPT and Gemini we captured them from September 18 and September 14 onward. Findings describe associations in this sample, not how any engine ranks businesses.
Source groups come from hand-built domain lists. "Business website" means a domain that matches the website of a business named anywhere in the dataset, 12,667 domains, so it is a lower bound. Citation links were de-duplicated within each answer. Gemini citations were captured for only 4,141 of its answers, so Gemini shares use those answers as the denominator and a missing citation is unknown, not zero. The ChatGPT map-card figures come from September 18 to 22, the only days we captured those cards in full. Full definitions are on our research methodology page.
Sources
Every link below was opened and checked on September 23, 2026.
- How ChatGPT and our foundation models are developed. OpenAI Help Center, updated August 2026. The three primary sources of training data, synthetic data, and the GPTBot robots.txt control.
- Searching the web with ChatGPT. OpenAI Help Center, updated August 2026. Automatic search, query rewriting, search providers, approximate location, third-party business listings, maps, ranking and OAI-SearchBot eligibility.
- Overview of OpenAI crawlers. OpenAI developer documentation. The difference between OAI-SearchBot, GPTBot and ChatGPT-User, and what opting out of OAI-SearchBot does.
- How does ChatGPT conduct local searches?. Search Engine Land, May 2, 2025, by Damian Rollison. Side-by-side tests showing ChatGPT drawing on Bing web results without Bing Places profile data. A dated outside observation, not OpenAI documentation.
- Introducing ChatGPT search. OpenAI, October 31, 2024. Partnerships with news and data providers for weather, stocks, sports, news and maps.
- Cheers Market Baseline, August 31 to September 20, 2026: 32 home-services questions, 100 US and Canadian metros, 34,742 usable answers, aggregate only. Research methodology.
Dylan Allen-Arnegård is the CEO & Co-Founder of Cheers, the local search platform for multi-location service businesses.