On 7 August, an audit titled Invisible to the Machine was published on arXiv. Its method is unusual, and that is what makes it interesting: instead of looking at who shows up in AI answers, it started from the complete census of all 4,776 restaurants, cafes and bars in two Bali markets, then counted how many of them were recommended at least once by ChatGPT, Claude, Gemini or Perplexity across 2,208 responses. The result: 85.6% never were. Not ranked badly. Never mentioned. And the factor that best separates the two groups is neither the rating nor the quality of the food. It is owning a website.
What this study does differently
Most numbers circulating about AI visibility are produced backwards. You ask ChatGPT some questions, look at who comes out, and infer a leaderboard of winners. The flaw is obvious as soon as you say it out loud: that method can say nothing about the absent, because it has no idea who is missing.
Vladimir Pitenin's audit (Norly Research) takes the problem from the other end. It starts by building the exhaustive list of venues that exist in two bounded markets, Canggu and Ubud, using Google Places data. That gives a denominator: 4,776. Only then does it query the AI systems and tick off who appears. What is left unticked is the invisible set, and it can finally be counted.
| Parameter | Value |
|---|---|
| Venues censused | 4,776 (restaurants, cafes, bars) |
| Markets | Canggu and Ubud, Bali |
| Systems tested | ChatGPT, Claude, Gemini, Perplexity |
| Responses collected | 2,208, from 96 persona-conditioned queries |
| Duration | 7 days, pre-registered protocol, stability retest 2 weeks later |
| Never recommended | 85.6% |
Two caveats before going further, because they matter. First: the study was funded and run by a company that sells AI visibility tools to hospitality businesses. That does not disqualify the data, since the protocol is published and so are the queries, but it does invite some distance from the commercial conclusions. Second: Canggu and Ubud are extremely dense tourist markets. The exact 85.6% figure does not transfer to a Paris side street. What transfers is the mechanism underneath.
The number that should worry you is not 85.6%
Across a complete census, a high invisibility rate is almost expected: the population includes food stalls, pop-ups, places with no online presence at all. It is easy to reassure yourself that you are not in that group.
So the author reran the model keeping only genuinely established venues, those with 50 or more Google reviews. That leaves 2,173. And within that subset, 72.6% are still never recommended.
That is the real finding. Three in four venues with visible local reputation, dozens of reviews, often years of history, do not exist inside assistant answers. The most widespread belief among business owners, "I have good reviews, so I will surface", has just been contradicted with numbers.
What gets you into the answer
The most useful part of the audit is not the headline, it is the model. By crossing each venue's characteristics with its probability of appearing, the author isolates four significant factors, and a fifth that does nothing at all.
| Factor | Effect on entry | What it means |
|---|---|---|
| Owning your own website | × 1.92 | The strongest factor of all |
| Review volume | × 1.64 | Per standard deviation, not per review |
| Listed prices | × 1.54 | A readable menu, not "contact us" |
| Third-party web mentions | × 1.44 | Guides, blogs, press, directories |
| Star rating out of 5 | × 0.89 (null) | No measurable effect on appearing at all |
| Foursquare presence | No effect | Neither at entry nor at ranking |
Spend a moment on the first row and the fifth. Owning a website counts twice as much as five stars counts for nothing. That is counter-intuitive after a decade of being told reputation is everything, and yet the logic is clean once you sit in the machine's seat.
An assistant composing an answer has to do two things: find candidates, then verify they are real and current. A 4.7 rating helps with neither, because every competitor is also at 4.7, so it carries no discriminating power. A website, on the other hand, is an independent, verifiable source the system can read, cross-check and cite. That is exactly the mechanism behind being cited by an AI: you are not picked because you are good, you are picked because you are checkable.
The same goes for listed prices (× 1.54). Plenty of restaurants publish a menu as a PDF or an image, or no menu at all. A menu in text, with readable prices, turns a guess into a usable fact, and makes you eligible for the whole family of questions that mention a budget. Technically it is the same reasoning as structured data: give machines facts they do not have to infer.
Getting in and ranking first are two different doors
The subtlest point in the study sits here, and it reconciles everyone. Star rating is null at entry, but becomes significant at ranking: among venues already recommended, it predicts the first position (× 1.17), alongside review volume (× 1.30).
Two floors, two levers. The first floor decides whether you are in the candidate set at all: it turns on the existence of sources (website, prices, mentions, review mass). The second decides your rank among those selected: it turns on reputation. Polishing your rating while you are still stuck on the first floor is polishing a shopfront on a street the machine never walks down. Order matters: you become a candidate first, you become first afterwards.
The real risk is not hallucination, it is staleness
We hear constantly that AI makes things up. On this corpus, that is wrong: only 0.08% of mentions matched fabricated venues. Statistically, noise.
On the other hand, 93 recommendations pointed at 14 permanently closed venues. The failure mode is not invention, it is lag. A place that shuts down leaves behind reviews, blog posts, "top 10" listicles, and that footprint keeps feeding answers long after the doors close.
Flip the sentence around and you have the strategy: these systems answer from footprints, and footprints have inertia. A competitor who has published for three years will keep being recommended for a while even if they slow down. And if you start today, you will not be visible tomorrow, but you are building a stock that will have inertia too. It is delayed-effect work, which is a reason to start early, not a reason to postpone.
There is no "the AI". There are four, and they disagree
One last, underrated result: agreement between the four systems on their top 20 is weak. The measured Jaccard index sits between 0.33 and 0.54, meaning two assistants asked the same question mostly recommend different addresses.
That kills a convenient illusion: there is no single leaderboard to conquer, therefore no single trick to find. Nobody can sell you "the" method to rank first in AI, because the four do not agree on who is first. What they do share are the entry factors above. And that is good news economically: the work that makes you eligible counts for all four at once, whereas ranking work would have to be redone for each.
What I take from it for a local business
I am not going to paste 85.6% onto Europe, that would be dishonest. Three things strike me as solid and transferable.
One. The entry threshold is higher than people think, and reputation does not clear it. If you never appear, it is probably not a quality problem, it is an available-material problem. I described the same diagnosis for shops in why a local business does not appear in ChatGPT; this study puts numbers behind it.
Two. The website reclaims a role many had written off. For years the line has been that a Google listing and an Instagram account are enough for a restaurant. That was already arguable for conversion; it is measurably false for being recommended by a machine. The listing gives tick boxes, the site gives text, and text is what gets quoted. The full argument is in why a restaurant still needs a website.
Three. Reviews still matter, but for the right reason. Volume counts at entry (× 1.64), rating counts at ranking (× 1.17). So keep collecting and replying, without expecting the rating alone to make you exist. Replying to reviews in ten minutes a week remains one of the best effort-to-effect ratios in this trade, and what Gemini builds from your reviews inside Google Maps depends on it directly.
The checklist, ordered by measured effect
- Own an indexable, up-to-date website (× 1.92). Not a Linktree, not a Facebook page: a domain of your own, with pages in readable text. It is the strongest factor in the study, by some distance.
- Publish your menu and prices as text (× 1.54). Not a PDF, not an image, not "price on request". An HTML page with items and amounts.
- Grow review volume (× 1.64), without fixating on the rating. Ask "what did you enjoy?" rather than "mind leaving a review?": the first produces text, the second produces silent stars.
- Exist somewhere other than your own site (× 1.44). Local guides, neighbourhood press, blogs, serious directories, tourist boards. Three solid mentions beat thirty automated listings.
- Check that AI crawlers can read your site. A site closed to AI crawlers cancels factor number one. That is a five-minute check in your
robots.txt. - Keep it fresh. Hours, exceptional closures, menu changes. Staleness is the dominant failure mode: do not become your own ghost venue.
- Measure your baseline before changing anything, so you can tell what moved. On the Google side, the Search Console AI report is the least-bad starting point.
Frequently asked questions
Why does my restaurant never show up despite great reviews?
Because great reviews are not the entry criterion. In the audit, 72.6% of venues with 50 or more reviews are never recommended, and star rating has no measurable effect on appearing (× 0.89, not significant). What gets you in is owning a website (× 1.92), review volume (× 1.64), listed prices (× 1.54) and mentions elsewhere (× 1.44). Your rating helps you rank first once you are in the shortlist; it does not get you in. If you are absent, look at your sources, not your reputation.
Is a website still worth it if nobody visits it?
Its job has changed, not its necessity. It is no longer only a place that receives visits: it is a verifiable source assistants read, cross-check and cite. That is why it comes out as the top predictor of appearing, ahead of every reputation signal. A Google listing supplies tick boxes; a website supplies text: menu, prices, method, answers to real questions. That text is what makes you quotable, including to people who will never see your homepage. See how a restaurant gets recommended by ChatGPT.
Does a Bali study apply to Paris or London?
The raw numbers, no. The mechanism, yes. Canggu and Ubud are dense, English-speaking tourist markets; the 85.6% figure does not transfer to a European high street, and I would be wary of anyone who claims it does. What transfers is the structure of the explanation: these systems recommend what they can verify from several independent sources, and a venue with no source of its own stays out of scope. That depends on how the systems are built, not on the country or the language.
Do AI systems invent restaurants?
Far less than the discourse suggests: 0.08% of mentions were fabricated. The real problem is staleness: 93 recommendations pointed at 14 closed venues whose online traces outlive them. For a business that is open, that reads as an instruction: the point is not to correct an AI, it is to maintain a recent, coherent footprint. An up-to-date site, fresh reviews, accurate hours.
Do I need a different strategy per assistant?
No, and the study shows it indirectly. Agreement between the four systems on the top 20 ranges from 0.33 to 0.54 on the Jaccard index: they largely recommend different addresses. So there is no leaderboard to conquer and no assistant-specific trick to buy. What the four share are the entry factors: your own source, reviews, prices, mentions. That is where effort pays, because it counts four times. That is the whole principle of GEO.
What if I am not a restaurant?
The study covers hospitality, and I will not pretend it transfers as-is to a tradesperson or a clinic. But the four measured factors are not sector-specific: a source of your own, review volume, readable prices, mentions elsewhere. Swap "menu" for "services and rates" and the checklist holds. If your local basics are not in place, start with local SEO for shops and independents before AI.
The full study is open access: Invisible to the Machine (arXiv, 7 August 2026). If you want to know which side of the line your venue sits on, write to me: I look at your site, your listing and your mentions, and tell you which of the four factors you are missing first. That is the core of my work on AI search optimization, and when the foundation is missing it starts with a website that finally gives machines something to work with.