We compared the Google Maps top 20 for 120 local searches in the United States, United Kingdom, Canada and Australia with the businesses ChatGPT listed for the same question: one answer per search on 26 September 2026, and seven more per search on 27 September 2026 (three repeats of the same question and four rewordings). In all, 898 answers showed a business list, giving 16,294 pairs of a Maps business and an answer.
Version 1.1 of this study adds the second day, the repeats and the rewordings, a regression model that adjusts for all signals at once, and measures of how stable ChatGPT’s picks are. One finding from version 1.0 did not hold up: keyword-stuffed business names were not reliably favored once more answers were examined.
The short version
- Review volume was the one signal that held everywhere: businesses with more reviews than the median for their search were more likely to be listed at the same Maps rank in all 8 sets of answers (by 15.8 to 22.0 points). In a model that adjusts for rank, every other signal, country, service, wording and day, the difference was 19.5 points (95% interval 15.0 to 23.0).
- Maps rank matters most. On 26 September ChatGPT listed 67.7% of the businesses ranked 1 to 3 in Maps and 25.4% of those ranked 11 to 20; after adjusting for the other signals, a rank of 11 to 20 instead of 1 to 3 lowered the chance by 34.1 points.
- A rating of 4.8 or more added 9.9 points in the model; a rating above the local median added 7.1 points, which version 1.0 (one answer per search) could not detect.
- Keyword-stuffed business names did not replicate: the 10.6-point advantage seen on 26 September was significant in only 2 of the 8 answer sets, and 2.4 points (not significant) in the model. We no longer report it as a finding.
- ChatGPT’s picks are fairly stable. Two repeat runs of the same question shared 0.74 of their listed Maps businesses on average (Jaccard); a run and a rewording shared 0.71, and the two days shared 0.68. Of the Maps businesses listed at least once in three repeat runs, 60.0% were listed all three times.
- Stable picks have more reviews. Among businesses listed at least once, 69.2% of those with more reviews than the local median were listed in all three repeats, against 44.1% of those with fewer.
The questions we tested
We treat each claim as a question the data can answer, rather than a conclusion:
- Is a business higher in Google Maps more likely to be listed by ChatGPT?
- At the same Maps rank, are businesses with more reviews, higher ratings, more photos or keyword-stuffed names more likely to be listed?
- Does each association hold across repeat runs of the same question, four rewordings and a second day?
- When the question is repeated, how much of the listed set stays the same, and which businesses stay on it?
The data shows which businesses ChatGPT lists, not why: we observe the answer, not the retrieval behind it.
What we measured
For each of the 120 local questions in our earlier study (“Who is the best {service} in {city}?”, five services, six cities in each of four countries), we took the Google Maps top 20 for “best {service} in {city}”, collected on the same day as the ChatGPT answers. Each Maps business was labeled listed or not listed in each ChatGPT answer that showed a business list.
Listed means ChatGPT named the same business at the same location, using the branch-aware matching of our local recommendations study: a listing and a Maps entry that share a website, name or phone number are compared on street address, postcode and phone, and another branch of the same company does not count. Version 1.0 counted every Maps entry that shared a listed business’s website as listed, including other branches of the same chain; with the stricter rule the 26 September listing rate is 40.8% instead of 42.9%.
Reviews, photos and ratings mean different things in different cities, so each business is compared with the others in the same search: above or below that search’s median. We estimate each signal in two ways:
- Within rank bands (the version 1.0 method): the difference in listing rate with and without the signal inside Maps rank bands 1 to 3, 4 to 10 and 11 to 20, averaged, separately for each of the 8 answer sets.
- A pooled model: a logistic regression on all 16,294 business-answer pairs with every signal, the rank band, country, service, reworded or not, and day, with uncertainty that treats each search as one cluster. We report each signal as an average marginal effect: the change in the predicted chance of being listed when the signal is switched on for every business.
All 95% intervals come from resampling whole searches, so answers and businesses from the same search are never treated as independent.
Findings
Maps rank
On 26 September, with one answer per search:
| Google Maps rank | Businesses | Listed by ChatGPT |
|---|---|---|
| 1 to 3 | 282 | 67.7% |
| 4 to 10 | 637 | 50.9% |
| 11 to 20 | 914 | 25.4% |
| All | 1,833 | 40.8% |
The pattern repeated in every answer set on 27 September: between 68.4% and 71.1% of the Maps top 3 were listed, and between 24.5% and 27.5% of ranks 11 to 20.
Profile signals, all adjusted at once
| Signal | Businesses with it | Change (points) | 95% interval |
|---|---|---|---|
| More reviews than local median | 49.5% | +19.5 | 15.0 to 23.0 |
| Rating 4.8 or higher | 84.0% | +9.9 | 4.7 to 14.7 |
| Rating above local median | 27.1% | +7.1 | 2.5 to 11.3 |
| More photos than local median | 48.7% | +3.2 | 0.5 to 6.3 |
| Keyword-stuffed name | 13.0% | +2.4 | −2.2 to 6.8 |
| Maps rank 11 to 20 (vs 1 to 3) | 49.9% | −34.1 | −40.4 to −28.5 |
“Change” is the model’s estimate of how much the signal changes the chance of being listed, holding the other signals, rank, country, service, wording and day fixed. “Businesses with it” is the share on 26 September.
A claimed profile, a website and a category that matches the service are left out of the model because 98% or more of the businesses have them, which leaves too few without them to compare.
The median listed business on 26 September had 235 reviews against 133.5 for businesses ChatGPT left out; both groups had a median rating of 4.9 and a similar number of photos (25 and 22).
Does each signal hold across runs, rewordings and days?
The within-band difference, estimated separately in each set of answers:
| Answer set | Answers | Reviews | Rating | Stuffed name |
|---|---|---|---|---|
| 26 Sep, original | 103 | +21.3 | +2.5 | +10.6 |
| 27 Sep, repeat 1 | 112 | +20.4 | +8.5 | +6.0 |
| 27 Sep, repeat 2 | 115 | +20.4 | +4.5 | −0.6 |
| 27 Sep, repeat 3 | 112 | +22.0 | +7.3 | +6.0 |
| 27 Sep, wording A | 113 | +18.5 | +5.6 | +1.7 |
| 27 Sep, wording B | 116 | +15.8 | +7.4 | +4.6 |
| 27 Sep, wording C | 115 | +21.3 | +5.7 | +2.3 |
| 27 Sep, wording D | 112 | +18.6 | +8.2 | +3.6 |
| Significant in | – | 8 of 8 | 6 of 8 | 2 of 8 |
Differences are in points: more reviews than the local median, a rating above the local median, and a keyword-stuffed name. The wordings were A “Which {service} in {city} would you recommend?”, B “I need a good {service} in {city}. Who should I go to?”, C “List the top {services} in {city}.” and D the original question with “Please cite your sources.”
Review volume is the robust finding. A rating above the local median is a smaller, fairly consistent advantage that a single answer per search was too noisy to detect. The keyword-stuffing advantage came mostly from the first day’s answers and should be treated as chance.
How stable are ChatGPT’s picks?
We compared the sets of Maps top-20 businesses listed in two answers to the same search with the Jaccard index (shared businesses divided by all businesses listed in either; 1 means identical).
| Comparison | Pairs of answers | Mean overlap | 95% interval |
|---|---|---|---|
| Repeats, 27 Sep | 283 | 0.74 | 0.72 to 0.77 |
| Original vs rewording, 27 Sep | 1,141 | 0.71 | 0.70 to 0.73 |
| 26 Sep vs 27 Sep | 251 | 0.68 | 0.66 to 0.71 |
In the 100 searches where all three repeats showed a list, 939 of 1,829 Maps businesses were listed at least once. Of those, 60.0% were listed in all three repeats, 19.3% in two and 20.8% in one. A top-3 business listed at least once was listed all three times 84.1% of the time; a business ranked 11 to 20, 41.1% of the time.
The Maps top 20 itself barely moved between the two days (mean overlap 0.87). Of the 1,438 businesses in both days’ top 20, 88.0% had the same status on both days (listed on 26 September and in at least two of three repeats on 27 September, or neither).
How widely ChatGPT spreads its picks
Across the seven answers per search on 27 September, a typical search had 8.7 Maps businesses listed per answer but 12 different ones across all seven. Counting how often each was listed, the picks behaved like 9.9 equally used businesses (median effective number; concentration index 0.101), and the three most-listed businesses took 34.4% of all listing slots. ChatGPT’s picks rotate among a core of businesses rather than settling on a single fixed list.
Google Maps is not the whole picture: 45.1% of the businesses ChatGPT listed on 27 September were not in the Maps top 20 for that search at all (46.8% on 26 September). This study is about the Maps top 20 only.
How this compares with other studies
| Source | Sample and date | Finding |
|---|---|---|
| This study | 120 searches, 898 ChatGPT answers over two days, 4 countries, September 2026 | More reviews than the local median: +19.5 points in all answers pooled, significant in 8 of 8 answer sets |
| Ibrahim and Zaki (arXiv) | Four service domains in the 100 largest US metros, September 2026 | “a 3-5x review-count premium” but “a rating premium of at most a tenth of a star” |
| SOCi 2026 Local Visibility Index | About 350,000 locations of multi-location brands, January 2026 | ChatGPT recommends 1.2% of locations; “ChatGPT-recommended locations average 4.3 stars” |
| Whitespark 2026 Local Search Ranking Factors | Survey of 47 local search experts, November 2025 | Experts rank presence on curated “best of” lists first for AI visibility, and high Google ratings eighth |
Our results and Ibrahim and Zaki’s point the same way: review volume goes with being recommended by AI. With eight times as many answers, we do see a modest rating advantage within a search, which is consistent with their “at most a tenth of a star” once most businesses are rated 4.8 or more. Whitespark’s ranking is expert opinion rather than measurement, and SOCi’s figures describe multi-location brands, a different population from our independent businesses.
Sources: Ibrahim and Zaki (opens in a new tab); SOCi (opens in a new tab); Whitespark (opens in a new tab); our ChatGPT local recommendations study.
What this means
This is our interpretation; the data shows associations, not causes.
- Reviews are the signal a business can most directly influence. Review volume was the strongest and most consistent profile signal, and a business can grow it honestly by asking every satisfied customer for a review.
- Ranking in Maps still matters most. Being in the Maps top 3 made a listing far more likely than any single profile signal, and top-3 businesses were also the most consistently listed.
- Judge AI visibility over several answers, not one. About a quarter of the listed set changed between two identical questions, so a single check can mislead in either direction.
- Don’t stuff keywords into the business name. The advantage we reported in version 1.0 did not replicate, and Google’s guidelines (opens in a new tab) do not allow it: a profile that breaks them can be suspended.
Methodology
- Searches: the 120 local questions of our ChatGPT local study (5 services, 24 cities, 4 countries).
- Answers: ChatGPT answers collected through DataForSEO’s LLM Scraper: the original question once on 26 September 2026; on 27 September 2026 three repeats and four rewordings (“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 the original with “Please cite your sources.”). 898 of 960 answers showed a business list.
- Businesses: the Google Maps top 20 for “best {service} in {city}” (DataForSEO Google Maps, country-level location), pulled on each collection day: 1,833 businesses on 26 September and 2,173 on 27 September.
- Listed: the same business at the same branch appears in the answer’s business list (branch-aware entity resolution from our local recommendations study).
- Signals: review count, photo count and rating compared with the median of the same search; rating of 4.8 or more; keyword-stuffed name (a name containing “ - ”, “|” or brackets); claimed profile, website and matching primary category reported but not modeled.
- Within-band estimate: listing-rate difference with and without each signal inside Maps rank bands 1 to 3, 4 to 10 and 11 to 20, weighted by band size, per answer set; 95% intervals from 500 resamples of searches.
- Model: logistic regression on all business-answer pairs with rank band, the five signals, country, service, reworded question and day; standard errors clustered by search; average marginal effects with 95% intervals from 200 resamples of searches (seed 20260928).
- Stability: Jaccard overlap of the listed Maps businesses between answers to the same search; listing frequency across the three repeats; effective number of businesses (inverse of the sum of squared listing shares) across the seven 27 September answers.
- Code: cite/pipeline/s14_v11.py (version 1.1) and s14_analyze.py (version 1.0).
- Update schedule: quarterly.
Limitations
- Associations, not causes. Businesses with more reviews may differ in other ways, and we cannot see which pages ChatGPT retrieved before answering. Testing whether a signal changes the answer would need a controlled experiment, which this study is not.
- Two days only. The repeats show short-term variation, not how picks drift over months.
- Maps results were collected at country level for a city search, so the Maps top 20 may differ from what a searcher in that city sees.
- The Maps top 20 is not the whole field. 45.1% of the businesses ChatGPT listed on 27 September were outside it and are not analyzed here.
- Simple signal definitions. The keyword-stuffing rule is a text pattern and misses stuffed names without separators; signals are split at the local median rather than measured on a scale.
- Matching was validated by model coders, not people. The branch-aware matching was checked in our local recommendations study with model-coded, adjudicated samples; no person coded them.
Data and downloads
- Every business-answer pair (version 1.1) with its signals and listing status: s14_business_answer_pairs.csv
- The version 1.0 businesses: s14_maps_businesses.csv and JSON
- Every statistic on this page: stats.json
- Machine-readable methodology: methodology.json
The data is free to reuse with attribution (CC BY 4.0).
To cite: Underneath. (2026). Which Google Maps businesses does ChatGPT recommend? (Version 1.1). Underneath Research. https://underneath.agency/research/chatgpt-local-picks-google-profile-study
Frequently asked questions
How does ChatGPT choose which local businesses to recommend?
Our data shows what the businesses it picks have in common, not how it decides. ChatGPT listed about two thirds of the Google Maps top 3, and at the same Maps rank, businesses with more reviews than the local median were about 20 points more likely to be listed, in every set of answers we collected.
Do Google reviews help a business appear in ChatGPT?
They go together. Businesses with more reviews than others in the same search were more likely to be listed in all 8 sets of answers, and more likely to be listed every time the question was repeated.
Does a higher star rating matter for ChatGPT recommendations?
Less than review volume. A rating of 4.8 or more added about 10 points, and a rating above the local median about 7 points, after adjusting for review count and Maps rank.
Does ChatGPT recommend the same businesses every time?
Mostly, but not entirely. Two answers to the same question shared about three quarters of their listed Maps businesses, and 60.0% of the businesses listed at least once were listed in all three repeats.
Should a business add keywords to its Google Business Profile name?
No. The advantage we saw in our first day of data did not replicate across the other answer sets, and Google’s guidelines for business names do not allow added keywords.