ChatGPT local recommendations: stable details, shifting lists

What we asked and collected

For five services (dentist, plumber, personal injury lawyer, accountant, and physiotherapist, asked as “physical therapist” in the US) in six cities in each of the four countries, we asked ChatGPT “Who is the best {service} in {city}?” through DataForSEO’s LLM Scraper, which queries the ChatGPT consumer web app from the matching country.

• 26 September 2026: one run of each question (120 answers) and the Google Maps top 20 for “best {service} in {city}”.
• 27 September 2026: three more runs of the same question (360 answers), one run of each of four rewordings (480 answers) and the Google Maps top 20 again. The rewordings were “Which {service} in {city} would you recommend?”, “I need a good {service} in {city}. Who should I go to?”, “List the top {services} in {city}.” and “Who is the best {service} in {city}? Please cite your sources.”

All 960 requests returned an answer. ChatGPT showed no business list in 62 of them and cited no source in 31; both stay in every count below.

What we measured

Every measure has a direction and a denominator, and they differ, so the same “overlap” can give different numbers.

Present means the listing is the same business at the same location as a Maps top-20 entry. A listing and a Maps entry that share a website or a name are compared on street address, postcode and phone: a matching street or phone confirms the same branch; a different street or postcode makes it the same company at another location, which is not counted as present. A street or postcode conflict outranks a shared phone number, and a candidate found by phone alone needs the same street. Listings with nothing to compare (4 in all) are counted as present. Of the 7,020 listings that the first version’s matching rule links to Google Maps across both days, 6,699 are the same branch, 317 are another location of the same company (for example another STAR Physical Therapy clinic in Nashville), and 4 have nothing to compare.

Set overlap is the Jaccard index: the number of businesses on both lists divided by the number on either list. A business is identified by its Google Maps place when it is present, otherwise by its phone number, otherwise by its name and street.

Findings

The details match Google Maps

The match extends to details that exist only on Google listings. 95.5% of website links carrying Google tracking tags (such as ?utm_source=gmb) were identical to the Maps link, tags included, and all 115 keyword-stuffed names (such as “Business Name - Emergency Plumber - City”) were reproduced verbatim on 26 September.

Review counts are the exception, and the gap depends on the country. They matched for 99.3% of present listings in the UK and 98.3% in Australia, but only 17.8% in the US and 14.7% in Canada. When the count differed, ChatGPT showed fewer reviews than Google Maps in 95.7% of cases.

The first version of this study reported 95.6% identical names and 93.2% identical links. The difference comes from branches: that version counted a listing as matched when it shared a website with any Maps entry, including another branch of the same chain, whose details naturally differ.

Overlap with Google Maps, in each direction

On 27 September the Maps top 3 was listed in 69.4% of cases (answers with a list), within the interval above. Intervals on this page come from resampling the 120 questions, so they allow for answers to the same question being alike.

The same question, a different list

Per question, ChatGPT’s lists a day apart overlapped 0.278 less than Google Maps’ top 20 a day apart (95% interval 0.235 to 0.323). The result does not change when companies are counted by website domain instead of branches: the same-day overlap is still 0.65. The overlaps above use pairs where both answers showed a list; 40 same-day pairs and 66 pairs across days had one answer with no list, and counting those as zero overlap lowers the same-day figure to 0.57.

Whether a list appears at all also varies. 86 of the 120 questions showed a list in every answer and 34 in only some; every question showed a list at least once.

Order is only weakly related to Google’s ranking

Within an answer, ChatGPT’s order and Google’s Maps rank were weakly and positively related: the mean rank correlation was 0.28 (95% interval 0.19 to 0.37) across the 87 answers on 26 September with at least 3 present businesses. The first business ChatGPT named was in Google’s top 3 in 35.0% of lists (36 of 103). This is an association in the order of two lists; it does not show that Google’s ranking causes ChatGPT’s.

Can a listed business be traced to a page ChatGPT cites?

We fetched every page ChatGPT cited (1,041 of 1,424 could be read) and checked whether it names each business listed in the same answer, by its full name or its website address. Naming is the only thing checked: a page that names a business is not thereby evidence that the business is good, or that ChatGPT used the page to choose it.

So between 29.3% and 67.0% of listings are traceable to a cited page; among listings with at least one readable cited page, the share is 31.3% (25.9% to 36.7%). On the same basis it was higher for businesses present in the Maps top 20 (40.7%) than for those that were not (20.5%). 78.6% of answers with a list had at least one traceable business, and 75.1% of readable cited pages named at least one listed business. On 27 September the figures were similar (30.8% of listings named). Only 1.9% of citations on 26 September pointed to a listed business’s own website.

Day to day

ChatGPT showed a list more often on 27 September than on 26 September: 94.2% against 85.8% of answers, a difference of 8.3 percentage points (95% interval 1.9 to 15.6). Lists from different days overlapped 0.057 less than lists from the same day (95% interval 0.029 to 0.091). The two days differ in time of day as well as date (04:30 and 16:20 UTC), and the first day has one run, so this does not separate a day effect from ordinary variation.

Four other wordings

Compared with the original wording on the same day, paired by question, we tested four outcomes for each rewording: whether a list appeared, whether a source was cited, how much the business list overlapped with the original runs (against the overlap between original runs), and the share of identical links. After correcting for the 16 tests, no difference was significant. Before correction, three intervals excluded zero: the lists for “I need a good {service} in {city}” and “List the top {services} in {city}” overlapped the original runs slightly less (by 0.041 and 0.029), and asking for sources gave 4.8 percentage points fewer identical links (95% interval 1.9 to 8.0). We treat these as leads. The Google Maps comparison is not a wording outcome, because the Maps search did not change with the wording.

How well the matching works

We checked the rules against blind coding by two AI models (Claude Opus and Claude Sonnet) on a targeted sample of 140 items: 25 matches through a website, 20 multi-location businesses, 25 cases the rule first called another location (24 after the postcode fix in the design record), 10 borderline matches, 20 unmatched listings and 40 cited-page decisions. The two coders agreed on all 100 match items and on 39 of 40 page items. Where the coders disagreed with the rule, we settled the case from the raw data and recorded the reason.

The page rule’s four errors were all chains: the page named another branch of the same brand, which the rule counts because the company name or website domain matches. For chains, “named on a cited page” is therefore a company-level measure. The one missed match was a plumber whose Maps address differs from the address ChatGPT shows but whose phone and branch web page are the same. An earlier round of model coding on stratified random samples of 200 match decisions (first-version matching rule) and 120 page decisions gave agreement of kappa 0.96 and 0.93. This validation is model-coded: no person coded the sample.

Observed, inferred and unknown

What we observe. ChatGPT’s displayed details for a business agree with the same business’s Google Maps entry field by field, including Google-only tracking tags and keyword-stuffed names, and stay the same from answer to answer. Which businesses ChatGPT lists changes much more from run to run than Google’s top 20 does from day to day. The pages ChatGPT cites name only part of the businesses it lists.

What we infer. The listing details most likely come from Google business data or a source that mirrors it, since tags and padded names that exist only on Google profiles reappear verbatim. The selection of businesses behaves like a draw that varies from answer to answer, loosely related to Google’s ranking. The cited sources are not a full account of where the list comes from.

What remains unknown. Which provider supplies the data; what ChatGPT retrieves or reads before answering (the consumer app shows neither); why it picks one business over another; whether a business named on a cited page was chosen because of that page; and whether the day-to-day difference is a change in ChatGPT or ordinary variation.

Where this sits in GEO research

Studies of generative engine optimization follow a chain: content → retrieval → selection → citation → prominence → fidelity → stability. Controlled studies test the first links by changing content. This study observes the middle of the chain in real local answers: how selection relates to conventional local search, how faithfully business details are represented, how traceable the list is to citations, and how stable each of these is. It does not test content changes, and it cannot see retrieval.

How this compares with other studies

The source of ChatGPT’s local data is disputed. A widely repeated claim says 60% to 70% of ChatGPT’s local results come from Foursquare; Steady Demand traced that figure to a small May 2025 study and, in 4,607 ChatGPT runs in August 2026, found Yelp as the grounding source for business cards in 95.83% of runs and Foursquare in 0.00%. Steady Demand also found ChatGPT’s named businesses resembled Google’s set more than Foursquare’s. SOCi’s 2026 Local Visibility Index found ChatGPT recommended 1.2% of the locations of multi-location brands, against a 35.9% appearance rate in Google’s local 3-pack.

Our study measures something different: not which source ChatGPT names, but whether the details it shows agree with Google’s. They do, field by field. Both results can be true at once if the listing card is built from Google-derived data while the text draws on review sites, or if a data provider mirrors Google’s listings; our data cannot say which. None of our answers cited Yelp, Foursquare or TripAdvisor; the most-cited sources were review and ranking sites such as birdeye.com and expertise.com.

Sources: Steady Demand (opens in a new tab); SOCi (opens in a new tab).

What this means

What follows is interpretation rather than measurement.

• Check that the Google Business Profile is accurate. ChatGPT’s details agree with the profile nearly field for field; if the two share a source, an error on the profile would likely appear in ChatGPT too.
• One answer is one draw. A business can appear in one answer and not the next; checking visibility needs several runs, not one screenshot.
• Being on the cited pages is not the whole story. Between 29.3% and 67.0% of listed businesses were named on a page the answer cited, and for a third no readable cited page named them, so citations alone do not explain who gets listed.

Methodology

• Questions: 120, five services by six cities by four countries (US: Austin, Denver, Charlotte, Phoenix, Columbus, Nashville; UK: Manchester, Bristol, Leeds, Glasgow, Birmingham, Edinburgh; Canada: Toronto, Calgary, Ottawa, Vancouver, Winnipeg, Halifax; Australia: Brisbane, Perth, Adelaide, Melbourne, Sydney, Canberra).
• ChatGPT: the consumer web app via DataForSEO LLM Scraper, country-level location; 26 September 2026 at about 04:30 UTC (1 run) and 27 September 2026 at about 16:20 UTC (3 runs and 4 rewordings).
• Google Maps: DataForSEO Google Maps SERP, “best {service} in {city}”, same country, top 20, collected on each date; each answer is compared with that day’s results.
• Same business, same location: candidates share the listing’s website domain, name or phone; a street or postcode conflict means another location; a matching street, phone or (UK, Canada) postcode confirms the branch. A candidate found by phone alone needs the same street.
• Cited-page check: each cited page fetched once (plain request, then a rendering service for failures); a listing counts as named when the page text contains its full normalized name or its website’s domain.
• Uncertainty: 95% intervals from 2,000 resamples of the 120 questions; differences are paired by question, with permutation p-values and Holm correction within each family of tests.
• Design record: the design file records the rewordings and runs as fixed before the second collection; the file was first committed after that collection and has since been edited, so this cannot be independently verified. The branch-aware matching rule and the revised outcomes were set after a first analysis of these data and refined while inspecting cases; the analysis plan records each change, including a fix found through a validation item (reading the Maps postcode from the full address, because Google’s structured field sometimes holds the street number), with its effect.

Limitations

• Two dates at different times of day, so day-level change cannot be separated from ordinary variation or from changes in ChatGPT.
• The consumer app shows neither what ChatGPT retrieved nor what it read, so being in the Maps top 20 shows visibility in conventional local search, not retrieval, and the study cannot say why a business was chosen.
• Agreement with Google Maps shows that two displays hold the same data; it does not identify the provider.
• The validation was coded by two AI models, not by people. A sample for human coding was drawn but not coded.
• For chains, the cited-page check works at company level: a page about another branch counts.
• Cited pages were fetched on 27 September and may differ from what ChatGPT read; 383 of 1,424 could not be read.
• Google Maps results were collected at country level for a city search, which may differ from what a searcher in that city sees.
• Only four rewordings were tested; they are not a sample of all the ways people ask.

Data and downloads

• All 960 answers, with status, wording, run and date: s9_answers.csv
• All 12,271 listings, each with its same-day Google Maps match, match class and field comparisons: s9_listings_all_days.csv
• Every Google Maps top-20 entry by date: s9_google_maps.csv
• Every listing and cited page pair with the check result, and the fetch record of each cited page: s9_citation_traceability.csv, s9_cited_pages.csv
• The targeted validation sample with the rule’s output, both coders’ labels and every adjudication: s9_validation_targeted.csv
• The design and deviation log: analysis_plan.txt
• Every statistic on this page: stats.json
• Machine-readable methodology: methodology.json
• The first version’s 1,405 listings: s9_chatgpt_local_listings.csv

The data is free to reuse with attribution (CC BY 4.0). To cite: Underneath (2026), ChatGPT local recommendations: stable details, shifting lists, Underneath Research, https://underneath.agency/research/chatgpt-local-recommendations-study

Frequently asked questions

Does ChatGPT copy Google Maps?

Our data shows agreement, not copying. For businesses present in both, ChatGPT’s name, link, street and phone match Google Maps in 98% to 99.8% of cases, including tracking tags that exist only on Google profiles. That points to Google business data or a source that mirrors it, but the study cannot identify the provider.

Does ChatGPT give the same list every time?

No. Two answers to the same question on the same day shared businesses with an overlap of 0.65 on a scale where 1 means identical lists, and 34 of 120 questions showed a list in some answers but not others. The details of each business, by contrast, were identical in 95.7% of repeat appearances on the same day.

Does ChatGPT list the businesses at the top of Google Maps?

Often, not always. When it showed a list, ChatGPT listed 67.7% of the Maps top 3 and 40.8% of the Maps top 20. Its order was only weakly related to Google’s ranking.

Do the sources ChatGPT cites name the businesses it recommends?

Often not. A cited page named the business for 29.3% of listings; because some pages could not be read, the true share is between 29.3% and 67.0%. Naming is not support: the check does not show that a page recommends the business.

How can a local business appear in ChatGPT’s recommendations?

Our study did not test what makes ChatGPT choose a business. It shows that ChatGPT’s details for a business agree with its Google Business Profile, so checking that the profile is accurate is a sensible step, and that a single answer is one of many possible lists.

Related guides