Why AI recommends a local business: what 12 AI assistants said about 130,996 picks

Summary

Key findings

1. Reviews are the reason AI gives most. Across all 12 assistants, 31% of the businesses recommended came with a review or rating as the reason. ChatGPT (50%) and Claude (47%) used reviews most; Microsoft Copilot least (5%). Only the "affordable" question made price the main reason.
2. AI almost only quotes high ratings. Of the 17,940 star ratings the assistants quoted, 90% were 4.5 or higher and 1.6% were under 4.0. When AI gives a rating, it is nearly always a high one.
3. Web-search assistants quote numbers. They gave a review count for 13% of their picks, against 2% for assistants that answer from memory and 1% for the voice assistants. A typical quoted count was about 200 reviews.
4. Memory assistants quote numbers they cannot look up. Mistral gave a star rating for 31% of its picks and a phone number for 41%, without searching the web. Its ratings agreed with the web-search assistants' to within 0.1 stars only 36% of the time; the web-search assistants agreed with each other 70–82% of the time.
5. "Affordable" changes the answer. For "affordable" questions, 71% of picks came with a price reason, against 16% for "best". Half of those answers (51%) gave a dollar figure; Meta AI did so in 99%.
6. Some assistants will not name a lawyer. Grok named no lawyer in 87% of its legal answers, DeepSeek in 63% and Alexa in 44%. ChatGPT, Claude, Gemini and Mistral almost always named one.
7. AI hands customers a checklist. 29% of answers told people to check a licence and 28% to get several quotes. In roofing, 67% said to get several quotes.
8. AI points people to a short list of sites. Google came up in 38% of answers, Yelp in 29% and the Better Business Bureau in 20%. For lawyers it was Avvo and the state bar (47% each); for real estate, Zillow (66%).

What it means. When AI recommends a local business, it usually says why, and the reasons are things a business can show: a high rating with many reviews, a licence, years in business, a price or a free estimate, a specialty. These are the reasons the assistants wrote, not proof of how they choose. But a business whose rating, reviews, licence and prices are easy to find gives every assistant something to say about it.

Scope. The October 2026 edition of AI Syndicate Local: 1,000 markets (50 metros in 9 states, times 20 types of business), three fixed questions to each of 12 assistants, asked between 1 and 7 October 2026. This paper reads all 31,516 answers that came back and the 130,996 business recommendations in them. Method and limits are at the end.

What did we read, and how?

Each month AI Syndicate Local asks 12 AI assistants three questions about a type of business in a metro area: the best one, the one to recommend, and a good affordable one nearby. In October 2026, 31,516 answers came back from 1,000 markets. This paper reads what the assistants said about each business they named. We call each one a pick, take the words that describe it, from its name to the next business named, and sort the reasons into 13 kinds:

The twelve fall into three groups by how they answer:

• Web search: ChatGPT, Claude, Gemini and Perplexity answer with web search on. Google AI Overviews is the AI Overview Google shows for the question. Microsoft Copilot is the Copilot answer Bing shows; when Bing shows none, we reconstruct one from the same Bing results.
• Model memory: Grok, Meta AI, DeepSeek and Mistral answer from what the model learned, because no consumer live-search access is available for them.
• Voice: Siri and AAlexa cannot be asked programmatically, so we reconstruct their answers in the short, spoken style each uses. That style shapes what they say: a typical Siri answer is about 40 words and names one business.

A reason in this paper is one the assistant wrote, not proof of why it picked the business. The limits are set out at the end.

Does AI say why it recommends a business?

Most of the time, yes. 72% of picks came with at least one reason. The rest were a name, sometimes with an address or phone number, and nothing about why.

Price and reviews lead, but for different reasons. Price comes almost entirely from the "affordable" question: in answers to "best" and "recommend", only 16–19% of picks mention price, while 29–32% mention reviews. Reviews are the reason AI gives most when nobody has asked about cost.

Assistants that search the web give reasons more often (75% of picks) than assistants that answer from memory (69%) or the voice assistants (64%). The web-search assistants lean harder on evidence: reviews (34% against 26% for memory), licences (14% against 9%) and years in business (15% against 5%). Memory assistants talk more about price (35% against 30%) and size (11% against 6%). The voice assistants give one short reason, most often reviews (35%) or service (20%).

Each assistant has its own habits. ChatGPT names licences and credentials for 37% of its picks, about three times the average. Gemini describes what a business specializes in (41%) and how it treats customers (32%). Perplexity writes short answers, about 90 words, and 43% of its picks come with no reason.

The first pick gets the most. In answers that name five or more businesses, the first is described in a median 44 words and the fifth in 32. 16% of first picks come with no reason, against 26% of fifth picks.

How does AI use reviews and ratings?

Reviews are the evidence AI reaches for most. 31% of picks mention reviews or ratings. 14% quote a specific star rating and 7% a number of reviews.

When an assistant quotes a rating, it is almost always high. Of the 17,940 star ratings quoted across the 12 assistants, 90% were 4.5 or higher. 1.6% were below 4.0.

This does not prove that AI filters out businesses below 4.5, because many picks come with no rating at all. But in the answers we read, ratings under 4.0 almost never appeared, and a rating that does appear works as a selling point.

Web-search assistants back the rating with a count. They gave a number of reviews for 13% of their picks; memory assistants did so for 2% and the voice assistants for 1%. The typical count quoted by a web-search assistant was about 200 reviews.

Gemini and Microsoft Copilot are the exceptions among web-search assistants. Gemini talks about reviews in general terms (27% of picks mention them) but quotes a star rating for under 2%. Copilot almost never mentions reviews (5%) and quoted no rating at all.

Can you trust the numbers AI quotes?

Not equally. A web-search assistant can read a business's current rating. An assistant answering from memory cannot. It can only repeat what it learned, which may be out of date or wrong.

Mistral, which answers from memory, quoted a star rating for 31% of its picks, more than any of the 12, and a phone number for 41%. We checked each assistant's ratings against the web-search assistants. Where Mistral and a web-search assistant rated the same business in the same market, they agreed to within 0.1 stars for 36% of 276 businesses, and were 0.3 stars or more apart for 25%.

The web-search assistants agreed with each other to within 0.1 stars for 70–82% of businesses. Alexa matched them as often. Meta AI and Mistral, which answer from memory, matched far less. The gap is our measure of how far a remembered rating drifts from a current one. Some samples are small, so read the exact figures as a guide.

Phone numbers raise the same question. ChatGPT gave a phone number for 37% of its picks and Mistral for 41%; Perplexity (16%) and Google AI Overviews (13%) did so sometimes, and the other eight rarely. We did not check the numbers, but a number from memory is only as current as the model's training.

What changes when someone asks for "affordable"?

The "affordable" question changes what AI talks about more than anything else we measured. For "affordable" questions, 71% of picks came with a price reason, against 16% for "best" and 19% for "recommend".

Half of "affordable" answers (51%) give real numbers, such as an hourly rate, a price range or a typical job cost. Among the web-search and memory assistants it is about six in ten. Meta AI did so in 99% of its "affordable" answers, Gemini in 80% and Mistral in 76%. Grok did so in 11%; it often gave advice on finding a fair price instead of naming businesses. Siri and Alexa, which answer in a few sentences, almost never did.

Free estimates and free consultations become a reason too: 8.5% of "affordable" picks mention one, against 1.7% for "best". What drops is credentials. Licences and certifications come up for 13% of "best" picks and 8% of "affordable" ones, and awards fall from 6% to 2%.

For a business, this means one customer question can be answered with two different shortlists, and the "affordable" one is built from price signals. A business with no published prices, ranges or free estimate gives AI little to say when a customer asks for a good deal.

Does AI give different reasons in different trades?

Yes. What AI says about a plumber is not what it says about a lawyer or a hotel.

Share of picks in each trade group mentioning each reason, 12 assistants, 1,000 markets, October 2026. Home and auto services has 11 trades, legal 4, health and beauty 2 (med spas and cosmetic dentists), hospitality 2 (hotels and restaurants) and real estate 1.

• Home and auto services are judged on reviews. 40% of picks mention them and 19% quote a star rating. They are also the only group where availability (24/7, emergency, same-day) and warranties come up often.
• Lawyers are described by what they handle. Specialty leads, and lawyers are the only group where results such as verdicts and settlements appear often. "Free consultation" and "no fee unless we win" carry the price reason.
• Med spas and cosmetic dentists are often given with a street address (19% of picks), more than twice as often as any other group.
• Hotels and restaurants are picked for where they are and what they cost. Location is mentioned for 30% of picks, and reviews for only 12%.

When does AI refuse to name a business?

Sometimes an answer names no business at all and gives advice instead. Across the 12 assistants, 13% of answers did. How often depends far more on the assistant than on the question.

Web-search assistants named no business in 4% of answers, voice assistants in 13% and assistants answering from memory in 24%. The split is not simply memory against web: Mistral, which answers from memory, almost always named someone, while Grok declined in half its answers and DeepSeek in nearly four in ten.

Lawyers are where assistants hold back most. Grok named no lawyer in 87% of its legal answers, DeepSeek in 63% and Alexa in 44%. ChatGPT, Claude, Gemini and Mistral named none in 3% or fewer.

Hotels and restaurants are the exception: every group names them freely. For a lawyer, a med spa, a real estate agent or a contractor, being well known is not enough to be named by every assistant; some will decline to name anyone. The web-search assistants, which almost always name businesses, are where these trades are most reliably found.

What advice does AI give customers?

Many answers end with a checklist. Across the 12 assistants, 29% of answers tell the customer to check a licence or insurance, 28% to get several quotes and 23% to check reviews. The voice assistants, which answer in a few sentences, almost never add advice; among the other ten it is about one answer in three. The advice depends on the trade.

Outside home services, the licence check comes up for med spas (38%), lawyers (20–30%, usually as "check their standing with the state bar") and cosmetic dentists (26%), and never for hotels and restaurants.

Memory assistants hand over more of the homework. They told the customer to check reviews in 50% of answers, against 9% for web-search assistants and 4% for voice assistants, and warned that their information might be out of date in 21% of answers, against 5% and 4%. DeepSeek gave that warning in 45% of its answers.

For a business, the checklist is a preview of what a customer who used AI will ask. In roofing, they will usually have been told to compare several quotes. For an electrician, they will often have been told to check the licence before calling.

Which sites does AI tell people to check?

When AI tells people where to look, it names a short list of sites. These are sites named in the text of the answer, such as "check their Google reviews", not the sources the answer cited.

Each trade has its own list:

When an assistant names no business, it sends the customer somewhere else. Of the 4,169 answers that named no one, 66% pointed to Google and 47% to Yelp; in legal answers, the state bar and Avvo. For a business that an assistant will not name, its profiles on those sites become the answer.

What does this mean for a local business?

These are patterns in what 12 assistants wrote across 1,000 markets. They show what AI says about the businesses it recommends, not a guaranteed way to be recommended. With that limit, the findings suggest where a local business can start:

• Make your rating and review count easy to find. Reviews are the reason AI gives most, and the ratings it quotes are almost all 4.5 or higher. Web-search assistants quote the count as well as the stars.
• Write down what AI says about your competitors. Licensed and insured, years in business, warranty, free estimates, 24/7 service, what you specialize in: these are the words in AI's reasons. Put them in plain text on your website and profiles.
• Publish prices or ranges if you want the "affordable" question. For that question, 71% of picks came with a price reason and half the answers gave a dollar figure.
• Keep your profiles current on the sites AI names for your trade. Google, Yelp and the Better Business Bureau for home services; Avvo, the state bar and Super Lawyers for lawyers; Zillow and Realtor.com for real estate; RealSelf and Healthgrades for med spas and dentists.
• Keep your phone number, address and rating consistent everywhere. Some assistants quote them from memory, and those can be out of date.
• If you are a lawyer, check each assistant. Grok, DeepSeek and Alexa often decline to name a lawyer; ChatGPT, Claude, Gemini and Mistral almost always name several.
• Expect customers who arrive from AI to compare. In several trades most answers told them to get several quotes or check a licence. Have both ready.

How this data is used. This research feeds the AI Syndicate platform, so our clients have the leading edge. Clients see where they stand in their market's AI Local ranking, and how they have moved, inside their dashboard, and their markets are measured more often than the rest of the index. What we learn across the whole index, such as which reasons the assistants give, which sites they send customers to and how each assistant answers, is built into the twelve-assistant measurement the platform runs for each client's own business and questions.

How was this measured, and what can it not tell you?

Method. The October 2026 edition of AI Syndicate Local asked 12 assistants three fixed questions in each of 1,000 markets, with the city named in the question and no user location or account history. A model extracts the business names from each answer. Each pick's description runs from its name to the next business named, ending at a new heading. We then matched the description against fixed word lists for each kind of reason, and the whole answer against word lists for advice and sites. We checked a random sample of matches for each kind by hand and tightened lists that caught unrelated text.

How the twelve are reached. ChatGPT, Claude, Gemini and Perplexity answer with web search on. Google AI Overviews is the AI Overview Google shows. Microsoft Copilot is the Copilot answer Bing shows, reconstructed from the same Bing results when Bing shows none; in this edition every Copilot answer was a reconstruction. Siri and Alexa are reconstructions in the short spoken style each uses. Grok, Meta AI, DeepSeek and Mistral answer from model memory, because we cannot reach their live search; their consumer apps may answer differently.

What this cannot tell you.

• A reason in this paper is one the assistant wrote. Models do not reliably report why they chose something, so these are stated reasons, not causes.
• Siri, Alexa and Microsoft Copilot are reconstructions. Their answers follow each assistant's style as we reproduce it, and for Siri and Alexa that style is a few sentences, which is why they give fewer reasons and almost no advice. Read them as close approximations, not recordings.
• Word lists miss reasons phrased in unusual ways and occasionally catch text that is not a reason. "Service" and "location" are the broadest lists and the least precise.
• We did not check whether the ratings, review counts, prices or phone numbers the assistants quoted are correct. The rating comparison measures agreement between assistants, not accuracy.
• The extraction model sometimes misses a business named only in passing, which can make an answer look as if it named no one. Google AI Overviews is most affected, because its business names often appear in link cards that our record does not keep.
• In California, Google AI Overviews, Microsoft Copilot and Alexa returned fewer usable answers in this edition, a gap in our collection.
• Each question is asked once a month, so a single answer can vary. The figures are averages over thousands of answers, not predictions for any one.

What stays private. The public index publishes each market's rankings and the assistants that named each business. The individual answers are not published.

Statistical notes

• Unit of measurement. Shares of picks pool every pick by the 12 assistants across all markets. Shares of answers pool every answer. A market is one metro area and one type of business.
• Ranges. 95% ranges come from resampling the 1,000 markets 1,000 times. For example, picks with a review or rating as a reason: 30.7% (29.9–31.6%); ratings quoted at 4.5 or higher: 90.1% (88.8–91.3%); answers naming no business: 13.2% (12.7–13.9%).
• Missing answers. A question that returned no answer is left out, not counted as naming no one.
• Star ratings. A rating counts when a number from 1 to 5 appears with "stars", "/5", "rating", "rated" or a review site nearby. When a description quotes two ratings, the first is used.
• Comparing ratings. Two assistants rated the same business when the business names matched after ignoring case, punctuation, "the", "and", "of" and legal suffixes, in the same market. Where one assistant rated a business more than once, we used its median; the web-search group's rating is the median of all its ratings for that business.
• List position. Position effects use answers that named five or more businesses, so that every position is filled.
• Typical values. "Typical" means the median.

How to cite this report

Quote any figure in this report with credit to AI Syndicate Research and a link to this page. You may quote up to 50 words of text, or share a single chart with its source line intact. Reproducing the report or large parts of it needs written permission. A suggested citation:

A PDF copy and a plain-text Markdown version are available. The rankings behind the report are published under a CC BY 4.0 licence at aisyndicate.com/ai-local/, with machine-readable copies at /ai-local/data.json and /ai-local/all.md. Journalists and researchers who need a cut of the data that is not here can write to support@aisyndicate.com.

Key figures

Every key figure in this report, in one table. The same figures are in findings.json and findings.csv, free to reuse under a CC BY 4.0 licence. Ranges are 95% intervals from resampling markets.

Sources

1. AI Syndicate Local: the October 2026 edition (1,000 markets) — AI Syndicate. Read Oct 8, 2026.
2. AI Syndicate Local methodology: how the index is measured — AI Syndicate. Read Oct 8, 2026.
3. AI Syndicate Local, machine-readable rankings with the assistants that named each business (CC BY 4.0) — AI Syndicate. Read Oct 8, 2026.