On this page
- Key findings
- How I did it
- 01On ChatGPT, the cited pages are directories and review sites
- 02ChatGPT's local shortlist is a set of map cards
- 03Google AI Overviews cite the business itself, and Reddit
- 04The question gets an AI Overview. The keyword search does not.
- 05Google AI Overviews' picks rank on Google's first page far more often than ChatGPT's
- 06The cards are steady. The written picks change every run.
- What I would test
- Limits
- Data
- Cite this study
- About the author
Ask ChatGPT which plumber to call in Tampa and it hands you a shortlist, usually as a set of map cards, sometimes with sources. Ask Google the same thing and, for a question-style search, you often get an AI Overview with a different shortlist and different sources. I wanted to know where those shortlists come from. So I asked ChatGPT and Google the same 352 local-business questions, three runs an hour apart on one evening, and traced every page they cited and every business they recommended.
Key findings
- 82% of the pages ChatGPT cites when it recommends a local business are directories and review sites. The businesses' own websites are 3% of ChatGPT's cited links.
- ChatGPT's local shortlists are mostly map cards. 91% of the businesses ChatGPT recommended came as a business card with an address, and when ChatGPT recommended a business, that business's own site was among the answer's sources 1% of the time.
- Google AI Overviews cite the recommended business itself. When a Google AI Overview recommended a business, that business's own site was among the sources 58% of the time.
- The sources differ by platform. Expertise.com appears in 42% of ChatGPT's answers that recommend a local business and Angi in 25%. On Google AI Overviews, Reddit appears in 44% and Yelp in 42%.
- A Google AI Overview showed up for 93% of the question-style searches and 21% of the keyword searches. For "best plumbers in Tampa, FL", Google mostly showed its usual local results instead.
- Google AI Overviews' picks rank on Google's first page far more often than ChatGPT's. 45% of the businesses Google AI Overviews recommended ranked on Google's first page for the same search. For ChatGPT, 17%.
- The cards are steady. The written picks are not. Counting map cards, the businesses ChatGPT recommended overlapped 0.72 from one run to the next (Jaccard index), and 54% of prompt-business pairs appeared in all three runs. Leaving the cards out, 16% of ChatGPT's written recommendations appeared in all three runs.
How I did it
I asked ChatGPT and Google AI Overviews 352 questions, three runs an hour apart on 18 September 2026, and recorded every page they cited and every business they recommended.
- The panel 176 city-and-service pairs: 22 US cities across the Northeast, Midwest, South and West, and eight services: roofing contractors, plumbers, HVAC contractors, pest control companies, electricians, dentists, auto repair shops and landscapers.
- Two wordings per pair A keyword search, "best plumbers in Tampa, FL", and a question, "Which plumbers in Tampa, Florida would you recommend?".
- two platforms ChatGPT with web search, with the city in the question, and Google's results page with the AI Overview, requested from each city's own location so the results are the city's.
- three runs an hour apart, so every figure comes with a measure of how much it changes between runs.
A ChatGPT answer counts as recommending a business when it presents that business as an option to hire, as a map card, a list item, a table row or a named further option. Every recommendation was matched to the business's website where the answer linked one or the same business was linked elsewhere in the data. 60% of ChatGPT's recommendations resolved to a website this way, and 90% of Google's, so figures about a business's own site cover the recommendations that resolved.
Finding 01On ChatGPT, the cited pages are directories and review sites
| Publisher of the cited page | Share | Links |
|---|---|---|
| Directory or review site | 82% | 2,879 |
| Publisher, media or other business | 13% | 459 |
| The recommended business's own site | 2% | 53 |
| Public or educational | 1% | 52 |
| Another business's site | 1% | 40 |
| Community or social | 0% | 7 |
| Could not be classified | 0% | 7 |
| Google page or card | 0% | 1 |
On ChatGPT, 82% of citation links next to its local business recommendations went to directories and review sites, and 3% to businesses' own websites. Publishers, media, communities and public sites together were 15%. The 95% uncertainty range for the directory figure is 81% to 84%.
The sites doing the work are not the ones a local business usually thinks about. expertise.com appeared in 42% of ChatGPT's answers that recommended a local business, angi.com in 25%, birdeye.com in 24%, consumeraffairs.com in 21% and bestprosintown.com in 15%. Yelp appeared in 1% of them and Reddit in 1%.
The pattern holds for every service. Directories and review sites were 99% of ChatGPT's cited links for landscapers, 88% for roofers, 84% for plumbers, 83% for HVAC contractors, 85% for electricians, 81% for auto repair shops, 73% for pest control companies and 67% for dentists.
Finding 02ChatGPT's local shortlist is a set of map cards
91% of the businesses ChatGPT recommended came as a business card with a street address, the kind of card a map shows, rather than as a written pick. The cards carry no link, so the business's website is usually not among the answer's sources at all. When ChatGPT recommended a business, that business's own site was among the sources 1% of the time, and 3% of the time for the written picks.
Where ChatGPT does put a link next to a written pick, it is usually the business's own page: 62% of the written picks that resolved to a site had their own page linked beside the name, 16% a directory page, and 19% no link at all. Those inline links are a different thing from the answer's source list, which is what the citation figures above count.
ChatGPT recommended 3,409 distinct businesses across 1,052 answers, and 1,043 of those businesses account for half of everything it recommended.
Finding 03Google AI Overviews cite the business itself, and Reddit
| Publisher of the cited page | Share | Links |
|---|---|---|
| Google page or card | 62% | 4,271 |
| The recommended business's own site | 15% | 1,052 |
| Directory or review site | 8% | 577 |
| Community or social | 6% | 382 |
| Another business's site | 5% | 368 |
| Publisher, media or other business | 3% | 206 |
| Could not be classified | 0% | 10 |
| Public or educational | 0% | 6 |
Google AI Overviews look different. 62% of citation links next to Google's recommendations were Google's own pages and business cards. Of the rest, the recommended business's own site was the largest single source: when an AI Overview recommended a business, that business's site was among the sources 58% of the time, and 68% of the time for written picks.
Among the pages that are not Google's own, reddit.com appeared in 44% of Google's AI Overviews that recommended a local business and yelp.com in 42%, then angi.com in 19% and consumeraffairs.com in 13%. Expertise.com, ChatGPT's most cited source, appeared in 1% of Google's AI Overviews.
By service, businesses' own websites were 58% of the AI Overview links for landscapers, 43% for electricians and 34% for pest control companies, and 10% for dentists.
Finding 04The question gets an AI Overview. The keyword search does not.
Google showed an AI Overview for 93% of the question-style searches and for 21% of the keyword searches, 57% of all Google searches in the study. For "best plumbers in Tampa, FL", Google's page mostly carried its usual local results without an AI Overview. By service, an AI Overview appeared in 69% of the searches for pest control companies and 48% for auto repair shops.
Finding 05Google AI Overviews' picks rank on Google's first page far more often than ChatGPT's
45% of the businesses Google AI Overviews recommended ranked on Google's first page for the same search in the same city. For ChatGPT, 17%. The same holds for the cited pages: 65% of the domains Google AI Overviews cited were on the first page for that search, against 15% of the domains ChatGPT cited.
Finding 06The cards are steady. The written picks change every run.
I ran every question three times, an hour apart.
Counting the map cards, ChatGPT's local shortlists are fairly steady: the median overlap of recommended businesses between one run and the next was 0.72 (Jaccard index: shared businesses divided by all businesses named in either run), and 54% of prompt-business pairs appeared in all three runs.
The written picks are another matter. Leaving the cards out, only 16% of ChatGPT's prompt-business pairs appeared in all three runs, and 67% appeared once. On Google AI Overviews, 7% of the written picks held for all three runs. For comparison, in my study of SEO agency recommendations, run over the same three-run window, 17% of ChatGPT's prompt-agency pairs and 30% of Google's held for all three runs. Local business picks change more.
Cited domains moved too: the median overlap between runs was 0.33 on ChatGPT and 0.44 on Google AI Overviews.
If you are reporting a local business's AI visibility from one screenshot of a written answer, most of what you see can change on the next run.
What I would test
A snapshot shows what AI cites and recommends, not what changes it, so each point below is something to test rather than a rule.
Test whether presence on the directories AI reads changes anything. On ChatGPT the cited pages are Expertise.com, Angi, Birdeye, ConsumerAffairs and BBB profiles, and on Google AI Overviews they are Reddit threads and Yelp. Those are the pages to check a business's presence on, and repeated runs would show whether a new or better profile goes with more recommendations.
Test whether the Google Business Profile behind the map card matters more than the website. 91% of ChatGPT's recommendations and 69% of Google AI Overviews' came as map cards, and ChatGPT's cards held steady between runs while its written picks did not. Whether the profile's reviews, categories and photos move a business into those cards is a question for repeated runs after a change, not for this snapshot.
Test whether ranking on Google's first page goes with being in the AI Overview. 45% of the businesses Google AI Overviews recommended were on the first page for the same search. Whether climbing onto the first page brings a business into the AI Overview is a before-and-after question.
Test with repeated runs, not one screenshot. Three runs an hour apart gave different written picks almost every time. Any AI visibility report for a local business should say how many runs it is based on and whether it counts map cards.
What this study does not show
The prompts were written by me, not pulled from real user logs. The 352 prompts are a deliberate panel of 22 cities and eight services, not a random sample of what people ask.
ChatGPT was collected as the consumer product with web search, with the city written into the question, and Google as a results page requested from each city. A study that queries ChatGPT through its developer API with web search sees a different product: no map cards, and different sources. Comparisons with such studies should say which product was measured.
Business names were extracted from answers with rules, not by hand. To check the extraction, Claude labelled 100 answers by hand, blind to the extractor's output: 994 business names. Against that set the published extraction has 94.5% precision and 88.8% recall: 97.1% and 89.9% on ChatGPT, 80.7% and 82.4% on Google AI Overviews. A second pass by an AI model then labelled every extracted name, and 431 names it called headings, address lines, credentials or directories were removed. That model was checked blind on 100 of its labels by a separate model, Claude Sonnet 5, which found it right on 53 of 53 business calls and 18 of 18 non-business calls (weighted accuracy 99.7%). 17 names it called publishers, other kinds of business or software were kept as businesses, because the check did not bear those labels out: 9 of 20 right. Matching each business to its website was automated too: 60% of ChatGPT's recommendations and 90% of Google's resolved to a website, and the own-site figures cover only those. A business I could not match could not be linked to citations, and a wrong match links the wrong site, so those figures can be off in either direction.
Publisher labels come from a dictionary of well-known directories, review sites, communities and platforms, from the matched business websites, and from page titles for the rest. The labels were checked three times on blind samples of 100 pages from the unreviewed tail, labelled by Claude: 75, 80 and 86 of 100 right, with the rules and the known-domain list corrected after each round (63 domain decisions). An AI model then labelled the 1,510 pages still unreviewed; a blind check of 100 of its labels by Claude Sonnet 5 found it right on 97 (weighted accuracy 97.5%), and its label replaced the automated one on 159 domains under a recorded rule. Reviewed or checked labels cover 87.0% of the cited links.
Nothing here is causal. I measured what two AI surfaces cite and name for a fixed panel of questions on one evening. Whether a new directory profile, a better Google Business Profile or a first-page ranking changes what they say is a question for an experiment, not a snapshot.
Data
2,112 searches: 352 questions, run three times each on both platforms, between 21:30 and 23:50 UTC on 18 September 2026. 1,056 ChatGPT answers came back, and 1,052 of them recommended at least one business. Google produced 603 AI Overview answers, and 553 of them recommended at least one business. ChatGPT cited 3,498 links across 222 domains; Google AI Overviews cited 6,872 links across 722 domains. The agency study this compares with is at https://ivanbuilds.com/ai-agency-citations.html.
Cite this study
Ivan Smiljkovic (2026). 82% of the Pages ChatGPT Cites When Recommending a Local Business Are Directories and Review Sites (Study of 2,112 AI Searches). Ivan Builds. https://ivanbuilds.com/ai-local-business-citations.html
Every section has its own link: hover a heading and use the # to copy it.
About the author
I study how Google and AI assistants choose what to cite, using public data I collect myself, and I publish the method and limits with every study.