What 200,000 Local AI Searches Reveal About Which Businesses Get Recommended
Google Maps has long been the reference point for local visibility.
This study asks a narrower, more practical question: if a business ranks well on the map, does it also get named when someone asks an AI assistant for a local recommendation? And when the same question is asked again, does the answer hold? The analysis covers 200,085 localised, non-branded prompts run across ChatGPT, Google AI Mode and Google AI Overviews, drawn from 1,342 business locations over a 60-day window, plus 1.9 million AI citations. Each prompt was repeated across time and across points on a map grid, so the study can separate three things that are easy to blur together: how much the answer set changes between runs, how often an individual business comes back, and how many businesses get named at all. The findings below are captured study evidence, not a live benchmark; treat the figures as a dated snapshot of the platforms as measured.
How the study was run
The team ran multiple reports (an average of three) for 1,342 locations across 60 days. Each report contained a set of prompts checked across nine map grid points, and every check was localised to that business's location. Branded prompts were excluded so they would not skew the results.
In total the analysis covered 200,875 non-branded prompts and 1.9 million AI citations. The same prompt was re-run several times over the period and from different points on the map, so the businesses returned could be compared run to run and place to place.
Three measures do most of the work in this study:
- Variance — how much the set of recommended businesses changes between runs of the same prompt.
- Persistence — how often a business that appeared once appears again in later runs.
- Count — how many businesses a single response names.
A manual version of the method, for one platform, looks like this. Take the prompt "need an emergency dentist open late in brooklyn tonight" and log who is named on each run. The illustrative run-by-run log is shown in the table below.
Repeat runs change the list, but individual businesses keep coming back
Across repeat searches of the same prompt, only 20–33% of each result set overlapped on average: ChatGPT 23%, AI Mode 21%, AI Overview 33%. That sounds like near-total churn, and in one sense it is — the exact combination of names shifts every time.
Persistence tells a steadier story. A business that appears once is seen again in roughly half of subsequent attempts: 50% persistence for ChatGPT and AI Mode, 58% for AI Overview. At the same time, 71–80% of businesses appear in half the attempts or fewer, so inconsistent inclusion remains the norm rather than the exception.
A useful way to read this: if you run a prompt four times, a single business is likely to appear in two of those runs. Across the four, the answers will contain a different set of businesses, and often a different number of them.
The pizza example makes the arithmetic concrete. The prompt "best pizza place in manhattan for a tourist that only has time to try one pizza" was run four times on ChatGPT; the run-by-run log is in the second table below. For that individual example, similarity between runs averaged 43% and persistence averaged 50%. One response added three businesses that appeared only once, which pulls both averages down — occasional longer lists make the responses look less alike than the recurring names suggest.
AI Overview showed the highest consistency of the three. It also produced fewer usable searches, because it does not appear for as many responses, whereas the other two surfaces always return something. That higher consistency may simply reflect its grounding in traditional search — and, as the next section shows, it also returns the shortest lists.
How many businesses each platform names
ChatGPT returns the longest lists, averaging 4.1 businesses per response (range 1–7, median 4). Google AI Mode averages 3.5 (range 0–6, median 4). Google AI Overview is the tightest, averaging 2.5 (range 0–4, median 3).
Empty responses are not rare: 10% for ChatGPT and 11% for both Google surfaces. Because each response can name a different number of businesses as well as a different set, list length is one of the quieter reasons answers look inconsistent. It is context for the rest of the findings rather than a reason to rebuild a local strategy from scratch.
A strong Maps position does not carry straight over to AI
Google Maps mentioned a tracked business in 66% of searches. The AI surfaces did so far less often: AI Overview 38%, ChatGPT 33%, AI Mode 32%. Maps was also markedly more stable, with 99% average agreement across attempts versus 91–96% for the AI platforms.
The platforms disagree with each other too. AI Mode and AI Overview shared only 29% of named businesses on average, despite both leaning heavily on Google Business Profile — which accounted for 28.5% of all citations across the three platforms. ChatGPT overlapped just 20% with AI Overview and 19% with AI Mode. Its different source mix, including Yelp and Bing, helps explain the gap; all three platforms use Yelp, and the Google surfaces also lean on Facebook significantly.
Part of the reason is mechanical. A Local Pack is chosen by a distinct local algorithm, while an AI answer is assembled from research across its index. For each prompt the AI runs what is called a fan-out query:
- It breaks the original prompt into several related questions or keywords.
- It searches for each of those separately.
- It combines the answers, weighs them for relevance, and returns a blended response.
A family-law prompt such as "best lawyer in downtown LA for family law" might fan out into searches like "best family law attorney Downtown Los Angeles divorce custody reviews" and "Los Angeles family law attorneys certified family law specialist downtown Los Angeles". Sometimes there will be more of these sub-queries, sometimes fewer. The practical consequence is that a grid full of strong map rankings is no guarantee that an AI platform will pick your business.
Moving across town changes the answer, and ChatGPT ranges furthest
Running the same prompt from different points in the same town produced overlap of 36.2% on ChatGPT, 46.9% on AI Mode and 47.4% on AI Overview. Only 3.5% of ChatGPT results were present across every point, against 8.8% for AI Mode and 25.0% for AI Overview.
Proximity still matters. Between 52.8% and 59.3% of picks fell within 5km of the search point, and 71.7–78.2% within 10km. The difference is in how far each platform's reach extends: the radius covering 90% of ChatGPT's results stretched to 287km, against 64km for AI Mode and 38km for AI Overview. Both Google AIs keep tight radii, with roughly 5% of chosen businesses within 1km. ChatGPT starts close but weakens as distance grows: about 53% of its picks fall within 5km, against 53% for AI Mode and 59.3% for AI Overview.
For a local business the takeaway is that "near me" has not gone away. Keep your address easy to find on your site, keep name, address and phone number consistent across off-site listings and mentions, and make location pages clear and specific to each place you operate from.
Your own website is still the most common kind of source
Of roughly 116,000 domains used as sources, about 108,000 were business websites — 93% of unique domains. Those business sites accounted for 42% of the 1,967,762 total citations. Third-party profiles on Yelp, Tripadvisor, Facebook and Google Business Profile were plentiful, but a business's own site remained central to giving AI platforms accurate information.
There is a plausible mechanism behind this. ChatGPT in particular has been observed using "site:" searches within its fan-out queries, which suggests the model often has a candidate in mind and then goes to that business's own site, or its brand pages on review sites, to confirm details.
What to make clear on your site:
- Expertise — why you are credible in your area.
- Reviews — why customers rate you.
- Structure — obvious location and service pages, and answers to common customer questions.
A practical way to find the questions worth answering: open Search Console, go to Performance, set the date range to 12 months, add a filter, choose Query and switch to Custom (regex). Two patterns are worth trying. To surface conversational queries, use ^(?:\S+\s+){9,}\S+$. To find "site:" style searches for your business, use (?i)(site:.*official|official.*site:). Add the questions that show up as tracked prompts, then answer them on your site. Bing Webmaster Tools and Google Search Console can both show the queries you receive impressions for.
Finally, make it easy for customers to contact you through the site. As agentic tools mature, bots may increasingly fill in forms on a customer's behalf, so keep forms simple and robust.
Directories and listings still shape the answer
Google Business Profile was by far the most important single source, especially on AI Mode and AI Overview. Beyond it, directories and listings such as Yelp, Facebook, Mapquest, Better Business Bureau and Angi remain influential. AI platforms also draw on industry-specific directories within particular verticals, so being listed in the right ones for your sector matters.
The table below lists the ten most-cited domains across all three platforms, with citation counts, share of total citations, the number of businesses citing each domain and that domain's share of total businesses. Read the two percentage columns together: a domain can account for a small share of all citations while still being used by a large share of businesses.
Storefronts are named slightly more often than service-area businesses
On average, storefront locations had a 33.6% mention rate against 29.4% for service-area businesses. The pattern held on all three platforms, as the table below shows. The gap is modest, but a fixed public address does appear to make a recommendation somewhat more likely.
What to do with these findings
The study's own recommendation is not to write off inconsistency as pure volatility. Some businesses recur far more often than others, and the ones that recur tend to be the ones doing the work. A business appearing in one of every two runs is a different situation from one appearing in one of every ten.
A short checklist drawn from the findings:
- Track across all three platforms, over time. A single check tells you little; the study's whole point is that answers move.
- Log who else is named and cited. Note competitors and the sources behind their mentions, then look at whether you can earn a place in the same places — more reviews, relevant local or industry directories, or a closer look at your Google Business Profile category.
- Keep a clear website. It does not need to be elaborate, but it should state who you are, where you are, what you are expert in and why you are trustworthy.
- Keep NAP and location pages consistent. Proximity still shapes results, so make your location easy to confirm.
- Remember that most local SEO tactics still apply. Claim your Google Business Profile and gather reviews across the relevant third-party sites.
A simple record to keep for each prompt run: platform, map point, run number, businesses named, and sources cited. That is enough to reproduce the study's three measures — variance, persistence and count — for your own business.
| Run one | Run two | Run three |
|---|---|---|
| Willoughby Dental | Willoughby Dental | Willoughby Dental |
| Emergency Dentist Brooklyn | Emergency Dentist Brooklyn | |
| NYC health + hospitals | ||
| Williamsburgh Dental Works | ||
| NYU Langone Health | NYU Langone Health |
Manual illustration of the logging method for one AI platform, not a study result.
| Run one | Run two | Run three | Run four |
|---|---|---|---|
| John’s of bleecker street | John’s of bleecker street | ||
| Joe’s pizza broadway | Joe’s pizza broadway | Joe’s pizza broadway | Joe’s pizza broadway |
| Prince street pizza | Prince street pizza | Prince street pizza | |
| l’industrie pizzeria – west village | |||
| Lombardi’s | |||
| NY Pizza Suprema |
Manual illustration; for this example similarity averaged 43% and persistence 50%.
| Platform | Storefront mention rate | Service area business mention rate |
|---|---|---|
| ChatGPT | 33.7% | 29.4% |
| Google AI Mode | 31.9% | 27.4% |
| Google AI Overviews | 38.4% | 36.9% |
Captured study figures; a dated snapshot.
Frequently asked questions
If AI recommends different businesses each time, is tracking pointless?
No — the study found that a business appearing once is seen again in roughly half of later attempts (50% persistence for ChatGPT and AI Mode, 58% for AI Overview). The list changes, but not everything in it changes. What matters is the rate: recurring in one of two runs is a very different position from appearing in one of ten.
Does ranking well in Google Maps mean I will be recommended by AI?
Not directly. Google Maps mentioned a tracked business in 66% of searches, against 32–38% for the AI surfaces. Maps was also far more stable across attempts (99% agreement versus 91–96%). The two systems work differently: a Local Pack is chosen by a local algorithm, while an AI answer is assembled from fan-out queries across its index.
Do ChatGPT and Google's AIs recommend the same businesses?
Largely not. AI Mode and AI Overview shared only 29% of named businesses on average, and ChatGPT overlapped just 19–20% with either Google surface. The different source mixes — Google Business Profile for the Google surfaces, Yelp and Bing for ChatGPT — help explain the gap.
Do I still need my own website for AI visibility?
The study's data says yes. 93% of unique source domains were business websites, and they made up 42% of all citations. Third-party profiles matter too, but your own site is where AI platforms confirm who you are, where you are, what you do and why you are credible.
How far away can a recommended business be?
Most picks are close: 52.8–59.3% fell within 5km and 71.7–78.2% within 10km. The reach differs by platform — the radius covering 90% of ChatGPT's results stretched to 287km, against 64km for AI Mode and 38km for AI Overview. Proximity still shapes local AI results.
| Domain | Citations | % of total citations | Businesses citing | % of total businesses |
|---|---|---|---|---|
| google.com (Google Business Profile) | 563,272 | 28.63 | 1276 | 94.17 |
| yelp.com | 187,599 | 9.53 | 1147 | 84.65 |
| facebook.com | 43,876 | 2.23 | 1261 | 93.06 |
| tripadvisor.com | 35,246 | 1.79 | 441 | 32.55 |
| opentable.com | 23,237 | 1.18 | 274 | 20.22 |
| mapquest.com | 8,282 | 0.42 | 741 | 54.69 |
| bbb.org | 6,519 | 0.33 | 397 | 29.30 |
| angieslist.com | 4,627 | 0.24 | 257 | 18.97 |
| reviews.birdeye.com | 3,612 | 0.18 | 343 | 25.31 |
| wanderlog.com | 2,694 | 0.14 | 268 | 19.78 |
Captured study figures; a dated snapshot, not a live ranking.
| Storefront mention rate | Service area business mention rate | |
|---|---|---|
| ChatGPT | 33.7% | 29.4% |
| Google AI Mode | 31.9% | 27.4% |
| Google AI Overviews | 38.4% | 36.9% |