Skip to main content

    Explainers

    85% of local venues were never recommended by AI. What the rest had

    An audit of 4,776 cafés, restaurants and bars found 85.6% were never recommended by ChatGPT, Claude, Gemini or Perplexity. Four things went with better odds.

    Mohamed Ashraf

    Marketing Specialist8 min read

    • What this covers: The first audit that counted every café, restaurant and bar in a market and checked which ones AI recommends, the four things the recommended ones had, why star ratings didn't matter for getting in, and a checklist for your own business
    • Who this is for: Founders of businesses customers find by place: cafés, restaurants, clinics, gyms, studios, coworking spaces, agencies with an office
    • Time to act: An afternoon for the basics, then about 20 minutes a month to check
    • Related pages: What is GEO?, Entity-based SEO, AI Mode "does not drive traffic". So what does being named get you?

    A researcher listed every café, restaurant and bar in two towns in Bali, Canggu and Ubud: 4,776 places. Then the study asked ChatGPT, Claude, Gemini and Perplexity where to eat and drink, 2,208 times over seven days. 85.6% of the venues were never recommended, by any of the four, not once. Even among established places with 50 or more ratings, 72.6% never came up (Pitenin, arXiv preprint, 7 August 2026).

    The places that did get recommended had four things in common. A high star rating wasn't one of them.

    Who ran the study, and how far can you trust it?

    It's a preprint, so it hasn't been peer reviewed yet. And it was funded and run by Norly, a company that sells review and AI-visibility tools to cafés, restaurants and bars. The paper says so on its first page and in a disclosure section. It also published its plan before collecting the main data, and it reports results that go against claims convenient for that business, such as the star-rating finding below.

    Four more limits, all stated in the paper:

    • It shows links, not causes. The numbers say what recommended venues tended to have, not that adding one of those things will get you recommended.
    • One region, one language. Two neighbouring towns in Bali with lots of tourists and remote workers, and the questions were in English.
    • Developer versions, not the apps. The four assistants were tested through their companies' developer access with web search switched on, not through the apps people use.
    • Food and drink only. We found no similar audit for accountants, clinics or agencies, or for Cairo or Dubai.

    What makes it worth reading anyway: most audits only see the businesses AI names. This one counted the whole market first, so it can show who was left out.

    The study measured each thing with an odds ratio, holding the others equal. Above 1 means the odds of being recommended went up when a venue had it. Below 1, they went down. The further from 1, the bigger the link.

    What the venue had Odds of being recommended at all (odds ratio)
    Its own website 1.92
    More reviews (how many, not how good) 1.64
    Price information listed 1.54
    Mentions on other websites 1.44
    A higher star rating 0.89, no clear link

    So the four are a website, review volume, prices and mentions elsewhere. The author's own summary for venues: "be documented before being excellent".

    Our read: every one of the four gives an assistant something to read. A website and listed prices answer the questions people actually ask ("is it expensive?", "do they have a menu?"). Reviews and mentions on other sites are other people saying the place exists and is worth going to. An assistant that searches the web before it answers can only recommend what it finds written down.

    Why didn't star ratings matter?

    They did, but only after a venue was already in the answer. The study found two steps:

    1. Getting into the answer. Here the rating made no clear difference (0.89, not statistically significant). The four things above did.
    2. Being listed first. Among venues that were recommended, a higher rating predicted being named first (odds ratio 1.17).

    Our read: a 4.8 doesn't help if the assistant never finds you. Get documented first. Then the rating starts to work for you.

    What about the "secret" data sources?

    You'll read that AI assistants take their local results from one particular places database. The study tested one of them. Being in Foursquare's open places dataset showed no positive effect on getting in or on being listed first. The paper notes it was partly designed around that test.

    Google's own advice is plainer. Its guide to AI features says that using its Business Profiles "can help your products and services to be visible in both AI responses and other Google Search results" (Google Search Central). And its local ranking help page lists three factors: relevance, distance and prominence, where prominence is "based on info like how many websites link to your business and how many reviews you have" (Google Business Profile Help). Links from other websites and review volume: two of the four things from the audit.

    It adds: "There's no way to request or pay for a better local ranking on Google."

    Is AI making things up about local places?

    Rarely. One likely-invented venue name made up 0.08% of the valid mentions. Being out of date was the bigger problem: the assistants recommended permanently closed venues 93 times, across 14 places that had shut. In the paper's words, "staleness, not hallucination, is the practical failure mode."

    And the four assistants didn't agree much with each other. Their top-20 lists overlapped only partly (a Jaccard score of 0.33 to 0.54, where 1 would mean the same list). Being named by one assistant tells you little about the others.

    A checklist for your business

    Four things from the study, plus one from its failure mode. How to do each one is our read.

    1. Your own website, with the basics in words

    One page per location, with your name, address, opening hours, what you offer and who it's for. Write it as text, not only in images or a PDF menu, so it can be read and quoted.

    2. More reviews, not just better ones

    Ask every happy customer, on the platforms your customers use. Reply to them. Google says: "Positive reviews and helpful replies can help your business stand out." In the audit, it was the number of reviews that went with getting in, not the average.

    3. Your prices, written down

    A menu with prices, a price range, packages, or "from" prices. "Contact us for pricing" leaves an assistant with nothing to say when someone asks whether you're affordable.

    4. Real mentions on other websites

    Local guides, neighbourhood blogs, press, partners and the directories in your category. Real ones only. Google says "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem."

    5. Keep it current

    Update your hours, holiday closures, moves and closed branches everywhere they're listed. The study's 93 recommendations of closed venues show that old information keeps getting served.

    How to check whether AI recommends you

    Once a month:

    1. Write five questions a customer would ask, with the place in them: "best brunch spot in Zamalek for a work meeting", "a dentist in Dubai Marina open on Friday".
    2. Ask each one in ChatGPT, Gemini, Perplexity and Claude, three times each, because the same question can get a different answer twice in a row.
    3. Count how many answers name you. Write down who gets named instead.
    4. Look at what they have that you don't, against the four things above.

    Common mistakes

    • Chasing a higher star rating before you have a website.
    • Thinking one profile on one platform is enough.
    • Hiding prices.
    • Buying reviews or mentions.
    • Leaving old hours or closed branches online.
    • Checking once, in one assistant, and calling it done.

    What we don't know yet

    • Whether the same four things matter outside Bali, in Arabic, or for service businesses like accountants, clinics and agencies. We haven't found an audit of Cairo or Dubai yet.
    • How much adding any one of them actually changes your chances. The study measured links, not causes.
    • Whether the consumer apps behave like the developer versions the study tested.

    We'll update this post as more audits come out.

    Frequently asked questions

    How many local businesses does AI recommend?

    In the first audit to count a whole market, very few. Of 4,776 cafés, restaurants and bars in Canggu and Ubud, Bali, 85.6% were never recommended by ChatGPT, Claude, Gemini or Perplexity over seven days and 2,208 answers. The study is a preprint, funded by a company that sells review and AI-visibility tools.

    Does a high Google rating get you into ChatGPT's answers?

    Not in that audit. Star rating showed no clear link with being recommended at all (odds ratio 0.89, not significant). Among venues that were recommended, a higher rating went with being named first (1.17).

    What makes AI more likely to recommend a local business?

    Four things went with better odds of being recommended: an own website (odds ratio 1.92), more reviews (1.64), listed price information (1.54) and mentions on other websites (1.44). These are associations from one study, not proof of cause.

    Does AI make up local businesses?

    Rarely. One likely-invented name made up 0.08% of valid mentions in the audit. The assistants recommended permanently closed venues 93 times, so out-of-date information is the bigger risk.

    Can you pay Google for a better local ranking?

    No. Google's Business Profile Help says: "There's no way to request or pay for a better local ranking on Google."

    Be the place in the answer

    When someone asks AI where to go, most places don't come up at all. The ones that do are written down: their own site, their prices, their reviews and other people's mentions. If you'd like to know whether AI recommends your business in your city, book a free AI visibility call.

    Keep reading

    Sources

    All sources re-read on 9 November 2026.