Briefing · Research
Most places were never named by AI. Here’s what went with being named.
A study listed 4,776 cafes, restaurants and bars in two parts of Bali and asked four AI systems where to eat and drink. At least 85.6% were never recommended by any of them. Getting named, coming first and having correct facts turned out to be separate problems.
Published Updated
Why you might be missing
One short answer can hide four different problems
Picture a visitor two streets away asking an AI assistant where to eat. It names four places. If yours isn’t one of them, nothing tells you: no lost click, no complaint. And one answer can’t tell you which of these went wrong.
Looking it up
Did it search first?
Some answers are written after a live web search; others only from what the AI already knew, which can be out of date.
Getting named
Were you named at all?
The answer has room for a handful of names. Most places aren’t among them.
Coming first
Where did you appear?
Named first, or fourth? In this study, different things went with each.
Being right
Were your facts still true?
Hours, address, prices, whether you’re still open. Old facts travel too.
One answer is one draw: one system, one question, one place, one language, one moment.
Bali, one study
Most places were never named
A preprint, a paper not yet peer reviewed, listed every cafe, restaurant and bar in greater Canggu and Ubud, then asked ChatGPT, Claude, Gemini and Perplexity where to eat and drink there.
Share of venues that none of the four AI systems ever recommended
All 4,776 venuesat least 85.6%
Venues with 50 or more ratings72.6%
Even among places with 50 or more ratings, nearly three in four were never named.
Why these numbers don’t transfer to your city
They describe these venues, in these two areas, asked in English, through versions of the four systems that search the web live while answering, available to developers, not the consumer apps people open. They are not a benchmark for another city, language, type of business or app.
Two different steps
Getting named and coming first went with different things
In this study, the places that got named and the places that came first stood out for different reasons.
Getting named went with
- Having its own website
- More reviews
- A listed price
- Being mentioned elsewhere on the web
- Not star rating: it showed no link at this step
Coming first went with
- A higher star rating, among places already named
- More reviews, which still mattered here
These are associations, not causes. None of them shows that changing a detail gets you recommended.
The numbers behind this comparison
The study reports odds ratios from a fitted model. For getting named: review volume 1.64, a website of its own 1.92, a listed price 1.54, mentions elsewhere on the web 1.44, and rating 0.89, which the author reads as no association. Among named places, rating went with first position at 1.17.
An odds ratio above 1 means higher odds inside that model. It is not a percentage-point gain, and it is not evidence of cause.
Wrong facts
Out-of-date places came up more often than invented ones
Across 12,439 valid mentions of venues, the study checked which places were real and still open.
Norly’s preprint, one week of answers
93
recommendations of places that had closed for good
They involved 14 venues confirmed as permanently closed.
Norly’s preprint, one week of answers
10
mentions of one name still judged likely invented
That is 0.08% of the 12,439 valid mentions, after checking.
Keep your hours, address, prices and closures right because customers rely on them. This study doesn’t show that doing so gets you named or brings anyone in.
How far these counts go
Both counts are small and come from one seven-day window. Comparing them is not a ranking of harm. The study recorded what the systems said, not what anyone did next, so it cannot show what either kind of error costs a business.
A second study, United States
Without a web search, two AI models made up most of their recommendations for doctors and advisers
Another preprint checked AI recommendations for doctors, care facilities and financial advisers in the 100 largest US cities against official government registers of providers, with and without live web search.
Without web search
- Only 4% of one model’s recommended doctors, and 11% of the other’s, matched a real clinician in the city asked about.
- Recommended advisory firms had misconduct records at 3.6 times the register’s usual rate, even after adjusting for firm size.
- Real recommendations clustered in the largest cities.
With web search
- 64–71% of recommendations in the same fields matched a real, registered provider.
- Recommended advisory firms had misconduct records below the register’s usual rate.
- Match rates were similar for the largest and the smaller of the 100 biggest US cities.
Whether the assistant searched depends on the run, not on your business. Note it every time you check.
Details and limits of the second study
Three set-ups were compared: an open-weight model, one anyone can download and run, a proprietary model without web search, and the same proprietary model with search. The registers were Medicare clinician and facility records and SEC adviser disclosures. The open-weight model’s few matches looked like name coincidences: its matched clinicians were no likelier to be primary-care doctors than names drawn at random.
Restaurants, measured separately in the same audit, showed a three- to five-fold premium for review count but a rating premium of at most a tenth of a star.
It is a preprint, covers only the United States, and doesn’t name the proprietary model it tested. It compares set-ups rather than proving cause.
Repeat the question
Ask the same question twice and the list changes
The Bali study asked each system the same questions again. On average, repeated answers shared only part of their names.
Names two answers to the same question had in common, out of all the names either one gave
Claude, the least consistent22%
Gemini, the most consistent45%
One answer is a snapshot. Check more than once, and with more than one system, before drawing a conclusion.
How the study was run
The main, preregistered collection covered 2,208 runs over seven days, using 96 English questions, each phrased as a particular type of visitor. Perplexity was run 10 times per question, ChatGPT and Gemini 5 times, and Claude 3 times, with Claude limited to two searches per run.
The study measures this overlap as mean Jaccard similarity, from 0.22 for Claude to 0.45 for Gemini. After fixes, 91.5% of runs passed the strict validity check, and matching names in answers to real venues was about 98% to 99% accurate. Across different systems, overlap among the top 20 venues ran from 33% to 54%. A retest two weeks later changed about as much as reruns inside the original week, which the author puts down to chance variation between runs rather than change over time.
Norly funded and ran the study, and sells review-management and AI-visibility tools to hospitality businesses. The author discloses this and describes steps meant to insulate the work; they don’t make the conflict disappear.
What to do
Six habits for checking where you stand
Check each step on its own
Look at searching, being named, position and correct facts separately. One combined score hides which one moved.
Write down how you checked
The system, the app or version, whether it searched, the question, language, place and date, plus the answer and the order of names.
Ask more than once
Use several runs, wordings and systems. Treat a handful of answers as a snapshot, not a verdict.
Keep your public facts current
Customers use your hours, address, prices and closures. Keep them right for that reason, not as a ranking trick.
Check the official listings about you
A missing, outdated or duplicate entry in an official register is a separate problem from not being named, and one you can fix yourself.
Measure customers somewhere else
Appearing in an answer isn’t a visit or a sale. Count those in your own bookings and till.
Keep reading
Related Briefings and next steps
Three related Briefings, and where to start with your own business.
- Your business was named in an AI answer. Did anyone walk in?A mention is one event and a customer is another. Count visits, bookings and sales in your own systems.Read the Briefing
- Anyone can suggest a change to your Google listing. Make sure your facts win.Google added new protections for Business Profile edits in April 2026. Keep the facts you author accurate and watch for changes.Read the Signal
- Google says its AI Search is launching in France. That doesn’t mean it shows your business.Eligibility, appearance, citation, recommendation and a customer’s action are separate states. Check each one.Read the Signal
- Check your AI PresenceSee what AI answers say about your business today.See the check
- Merchant InfrastructureThe business facts and commitments AI agents depend on, kept accurate at the source.See how it works
Sources
Where this comes from
Two preprints, papers not yet peer reviewed. Norly funded the Bali study and sells tools in the same category as Obenan, so read both their findings and our summary with care. Obenan has no partnership, endorsement, integration, pilot or special-access relationship with the authors, Norly, arXiv, ChatGPT, Claude, Gemini, Perplexity or their providers.
- Full arXiv HTML paper
arXiv Submitted Checked
Methods, preregistration, results, validation, limits, funding and the competing-interest statement. Checked again on 2 September 2026.
- Understanding AI Provider Recommendations in Local Service Markets
arXiv, Hazem Ibrahim and Yasir Zaki Submitted Checked
The US registry audit comparing answers with and without web search.