Does ChatGPT Recommend Local Businesses? 250 AI Picks Tested

Does ChatGPT Recommend Local Businesses That Actually Exist?

Almost always, in this test. Here is the short version across all four assistants, from 250 recommendations collected on October 7, 2026:

1. 97.6% were real businesses serving the metro area. 94.0% matched a Google Maps listing in the area and another 3.6% were real operators confirmed another way, while the last 6 were 2 hard failures, 3 businesses outside the area and 1 unresolved.
2. 22.8% were in Google's top 3 map results for the matching search. An answer can match at most three businesses and most named about five, so the most possible was about 56%.
3. The four AIs never all listed the same business first, not once in 12 questions. Only 6 of 72 head-to-head comparisons of first picks matched.
4. Only 2 recommendations failed outright, both from Claude: one company with no live listing and a web address serving gambling spam, and one that no longer operates in the city.
5. Phone-number mismatches were more common than wrong businesses. In 7 recommendations, the number the assistant gave did not match the main number on the company's live Google listing.

Where this started

Before the test, I had a Claude agent ask ChatGPT for the best emergency septic service in Des Moines. One of the options it listed was desmoinesseptic.com, which is my own rank and rent site (a real local septic company takes its calls).

That site doesn't name the company behind it, though. By this test's own rules, an outside checker couldn't verify the operator, so it would count as a lead-gen failure.

In other words, ChatGPT put a site with no named operator right next to real septic companies in an emergency answer. My site showing up in that answer is part of why the test ran in Omaha, Nebraska and Boise, Idaho, where I own nothing. My sites and client sites are excluded from every number on this page.

The same answer showed Bob's Septic & Portable Restroom Service with a 1.0-star rating from 2 reviews, plus two different phone numbers. On Google Maps, Bob's has 4.3 stars from 84 reviews, and only one of those numbers matches its listing.

When the same setup ran the question again after the test, ChatGPT made Bob's its top pick and left my site out of the answer. The 1.0-star card and the two phone numbers were still there.

What Happened to All 250 Recommendations

Each dot below is one recommendation, grouped by what the checks found. A business can appear more than once, since several were named by more than one assistant.

Most AI Picks Are Real, and Most Aren't Google's Top 3

ChatGPT came closest to Google, and even so, only 31.0% of its picks were in Google's top 3. The other three assistants matched less often.

Claude's lower match rate comes mostly from naming smaller operators that are real but had no Maps listing the checks could find. Counting those, 95.1% of its picks were real businesses serving the metro area.

Overlap by trade

Top-3 overlap varied more across trades than across assistants in this sample.

On the original Omaha roofing search, none of the four named any of the three roofers Google showed first. When Claude agents pulled that exact search twice more the same afternoon, three minutes apart, the second pull had swapped one of the three for McCoy Roofing, which was ChatGPT's first pick.

Google's map pack moves too. Treat any overlap figure, including mine, as a snapshot.

Two more numbers came up after the test rather than being planned. Looking from Google's side, the four AIs together mentioned 29 of the 36 businesses in Google's top 3 (three per search, 12 searches) at least once.

Those top-3 businesses were also more likely to be listed first. They made up 35.4% of first picks, compared with 22.8% of all recommendations.

For a related but different question, SOCi studied multi-location brands and found ChatGPT recommends just 1.2% of brand locations, compared with an average 35.9% appearance rate in Google's local 3-pack.

Explore Every Answer

Pick a city and a trade to see Google's top 3 next to every business each assistant recommended, in the order it listed them.

The Four AIs Rarely List the Same Local Business First

Across these 12 questions, the four assistants usually listed different businesses first.

Each assistant's first pick was compared with every other assistant's first pick. That makes 72 head-to-head comparisons, and only 6 matched (8.3%).

The clearest agreement was Omaha septic, where ChatGPT, Gemini and Perplexity all put Crown Pumping first. Claude said there was no single best and listed Midland Pumping Service first.

This lines up with the Steady Demand AI Citation Ledger, which found ChatGPT and Gemini name the same top business only 4.9% of the time.

The answers also drift. The test reran 12 of the 48 answers, and an average of 79.1% of the businesses from the first run showed up again.

Where the AI Recommendations Went Wrong

A bad recommendation can waste a homeowner's call and send the job away from a real business. These were the clearest misses, each re-checked by Claude agents against live listings on October 7, 2026.

The 7 phone-number mismatches can mean an old number, a tracking line or a plain mistake. Either way, call the number on the listing.

The ASAP Septic card is a correction of the test itself. The Claude agents' first pass flagged 7 recommendations as closed, not found or lead-gen, and a live re-check cleared 4 of them, with 1 more unresolved either way.

For a separate check of a different engine, Whito tested Google AI Mode in 20 UK towns and found 5.2% of the businesses it called trustworthy could not be verified as active companies trading in that town. It checked company registry records, so its unverifiable category is not directly comparable to the 2 failures here.

What This Means If You Run a Local Business

The errors documented here included stale or conflicting details: an old closed listing, a dead domain, a phone number that differs from the listing. The test can't prove what caused each error, but those gaps are the ones you control.

• Keep your Google Business Profile current, especially your hours and phone number.
• Use one main phone number everywhere: your listing, your site, Facebook and every directory.
• Keep your domain renewed. A dead domain can still get handed to customers.
• Close or merge old listings when you move, rebrand or sell, so a stale record cannot outlive the real one.

None of that is new. My take from this pilot: for local service businesses, getting named correctly by AI starts with the same fundamentals that already pay off on Google.

Getting named more often is a different question. I covered which sources ChatGPT actually reads in my test of how to rank in ChatGPT.

If you are on the other side of this, looking for a plumber at midnight, open the business's Google listing before you call the number an AI gave you.

How I Ran This Field Test

• Questions. "Who is the best [service] in [City, ST]?" for 6 urgent home services (septic pumping, emergency plumber, HVAC repair, roofer, junk removal, garage door repair) in Omaha and Boise. That is 12 questions and 48 answers.
• Assistants. ChatGPT through its consumer app with web search on, collected with DataForSEO's AI Optimization API. Claude, Gemini and Perplexity through their APIs, with web search confirmed on every answer but no user location. The results apply to these four setups and may differ in the consumer apps.
• Recommendations. Every business an answer presented as a recommendation, which came to 250 entries covering about 160 different businesses. Businesses named only as warnings did not count, and Claude's first-listed business counted as its first pick.
• Google comparison. The local pack top 3 for the matching Google search (for example, "roofing company Omaha NE"), pulled the same day in each city.
• Checks. Claude agents matched each recommendation to Google Maps listings and websites, checked the 57 top-3 matches against the saved search results, live-checked every closed, lead-gen or not-found flag, and confirmed the 7 phone mismatches against live listings.
• Reliability. A separate Claude model re-checked 50 randomly drawn recommendations against live listings or the saved search results and agreed with the labels on 48 (96%), above the 90% bar set in advance. Both disagreements were the same plumber, a real business matched to the wrong Maps listing on its name alone. That fix is in the numbers above, but other name-only Maps matches were not re-checked one by one, so the 94.0% Maps figure could run slightly high.
• One inconsistency. Claude sometimes included names from a third-party list. Eight of those were counted in three answers but left out of a roofing answer, and removing them moves every headline figure by about a point or less.
• Disclosure. Claude ran the checks and is also one of the four assistants tested, and I sell a Claude Code product. The same rules applied to every assistant, and Claude finished near the bottom on top-3 overlap and had both hard failures.

Before collecting any data, the plan set criteria for expanding to a full 10-city study. Top-3 overlap had to land at 50% or above, or 20% or below, or hard failures had to reach 10% (closed, not found, or a lead-gen site with no verifiable operating company behind it).

Overlap came in at 22.8% and hard failures far below 10%, and the pilot also ran over the plan's cost and workload limits. So the test stopped at two cities, and I'm reporting it as a field test.

The agreement figures, the 29-of-36 coverage figure and the first-pick figure were not part of that plan, so treat them as observations. The bigger limits: two cities, 12 questions, one day.

Make Your Business Easy for AI to Get Right

Does ChatGPT recommend local businesses that are real? In this test, the four assistants named real, local businesses 97.6% of the time.

They often disagreed with Google and with each other, and the details they repeated were not always current. The part you control is whether those details are right when an AI does name you.

I ran this test from Claude Code. If you want the marketing skills I run there every week, they are in my Claude Code Skills Stack. It is my own product, so weigh my Claude Code bias accordingly.

Doser, R. (2026). Does ChatGPT recommend local businesses? A field test of 250 AI recommendations in Omaha and Boise. ryandoser.com. https://ryandoser.com/research/does-chatgpt-recommend-local-businesses/