AI restaurant discovery is mostly “invisible”—here’s why

AI restaurant discovery is mostly “invisible”—here’s why

Introduction: When your café never shows up in AI

If you’ve ever asked an AI assistant for “a great coffee place near me” and only got a short list of a few names, you’ve already sensed the problem: AI doesn’t just rank—it selects. And that selection can directly affect who gets customers.

New research from Vladimir Pitenin, published as “Invisible to the Machine: Auditing AI Restaurant, Cafe, and Bar Recommendation Against a Complete Market Census,” tackles a surprisingly under-measured question: which cafés, restaurants, and bars do AI systems actually surface—and which ones never make it into the answer at all? The study runs a true “denominator” audit: instead of sampling only a shortlist, it builds a near-complete market list and then checks what four major AI systems recommend against it.

The headline is blunt. In two Bali markets (Canggu and Ubud), 85.6% of venues were never recommended by any system in 2,208 AI search-grounded runs. Even among established venues with 50+ ratings, 72.6% were still never recommended. Worse (for businesses), the “failure mode” isn’t usually creative hallucination—it’s staleness, with systems repeatedly recommending places that are permanently closed.

Why this matters: AI discovery is now a revenue channel—with a measurable blind spot

This research lands right when local discovery is shifting from maps and search results into conversational interfaces. People still use Google Maps and Yelp—but they increasingly ask an assistant for “three places for date night” or “somewhere quiet to work with good Wi‑Fi.” Those answers are often the entire decision surface. No page two. No infinite scroll. Just a small list—and the rest effectively disappears.

Picture a real-world scenario today: you run a small independent café in a tourist area. You have strong reviews, but your web footprint is inconsistent, your pricing info isn’t clearly listed, or your online mentions are thin. You might assume “high rating = more visibility.” This paper suggests a more uncomfortable reality: rating matters mainly after the venue is already retrieved and included in the model’s candidate set. Whether you enter the answer at all is more tied to documentation signals like having an own website, clear price information, and enough review volume to populate the system’s retrieval context.

This builds on—and also clarifies—previous AI auditing approaches:
- Some earlier audits test ranking among a fixed candidate set (you can’t win if the venue wasn’t included).
- Others audit visibility across curated brand lists, but without a real population denominator—so they can’t say what “never recommended” means in rate terms.
- Here, the study enumerates the full market in two bounded areas (4,776 venues) and then measures population-level invisibility. That’s the missing piece.

If you care about GEO (generative engine optimization) or fairness in AI recommendations, this paper gives you something rare: not just “bias exists,” but how selection vs ranking behaves differently, and which concrete business signals correlate with each stage.

What the researchers actually measured: a census-based AI recommendation audit, not a sampled brand test

The study focuses on two adjacent but bounded submarkets in Bali:
- Greater Canggu (including areas like Berawa, Batu Bolong, Pererenan)
- Greater Ubud (including Penestanan, Sayan, Pengosekan)

Building a real denominator: 4,776 food-and-drink venues

The researchers built a census of 4,776 operating cafés, restaurants, and bars across these two polygons at the time of collection (using Google Places as the grounding frame). They used the Places Nearby Search API in an adaptive grid to avoid truncation in dense areas, then added additional venues discovered from AI-suggested names that didn’t initially match the food-type enumeration.

They also did important validation: the census isn’t “perfect,” but it’s close enough that the invisibility rates reported are treated as floors, not guesses. The paper includes capture–recapture-style checks against the open POI dataset Foursquare and other audits to estimate how much might be missing.

Auditing what AI assistants surface: 2,208 search-grounded runs

They queried four production AI systems:
- ChatGPT via gpt-5.2-2025-12-11 using the Responses API with a web-search tool
- Claude via claude-sonnet-5 with a server-side web-search tool
- Gemini via gemini-3.5-flash with Google Search grounding
- Perplexity via sonar

They used:
- 8 personas (e.g., digital nomad seeking work coffee, couple planning a date, vegan/diet-constrained diner, late-night group, etc.)
- 6 first-person templates per persona
- 2 areas (Canggu vs Ubud)
That creates 96 unique queries, each repeated across multiple runs per system.

In the confirmatory wave they ran 2,208 runs total (over 7 days). After rule-based filtering, they extracted 12,439 valid venue mentions from the assistants’ short lists (the kind of “3–8 venues” output people usually see).

Extracting and matching venue mentions reliably

Two technical details matter for credibility:

1. Mention extraction: They used an LLM extractor (with strict JSON schema) to label each venue mention as:
   • recommended
   • mentioned-neutral
   • advised-against
2. Entity matching: They matched assistant output names back to the census using a fuzzy matcher with core-token similarity thresholds and additional safeguards (including cross-area rules).

They validated this pipeline using:
- independent double annotation on a sampled set of runs (mention-level precision 97.6%, recall 99.1%, and strict run-level agreement 91.5% after remediation)
- a 100-mention audit for entity matching (population-weighted correct-entity rate around 98–99%, with remaining errors concentrated in hard near-name cases)

Why include this detail? Because if you mess up matching, you can accidentally “discover” patterns that are really just errors in the audit pipeline—something the paper explicitly warns about and corrects for during validation.

The core result: most venues never appear—visibility is a long-tailed gate, not a popularity contest

Here’s the most dramatic number in the whole paper:

• 85.6% of all 4,776 venues were never recommended by any system in any run
  That’s 4,087 venues never surfaced.
• Even among “established” venues with 50+ Google ratings, 72.6% were never recommended.

And this isn’t just one system. It’s across all four systems in the multi-run protocol.

Long-tailed “winner” effects—but no winner-take-all

Among the venues that did get recommended, the distribution isn’t concentrated on one or two famous places. The leading venue holds only:
- 1.9% of all recommendations
- pooled concentration metrics show the visibility curve is long-tailed (e.g., Gini 0.668 across recommended venues)

So the problem isn’t “AI only recommends the top few by quality.” It’s that the gate to being included at all is extremely hard to pass, and then attention is distributed across many visible options—but the visible set is tiny relative to the market.

Two margins: documentation determines entry, rating determines ordering after entry

One of the paper’s most useful contributions is a clear conceptual split:

• Margin 1 (entry): does a venue appear in the assistant’s recommended list at all?
• Margin 2 (rank/order): if it appears, does it land at the top?

The study finds a dissociation: the signals that affect entry are not the same signals that affect rank.

What predicts whether AI includes a venue (entry margin)?

In their primary model, the probability of being recommended in eligible runs is associated with:

• Review volume: OR 1.64 (per SD of log review count)
• Own website: OR 1.92
• Listed price information: OR 1.54
• Web mentions volume: OR 1.44

Meanwhile:
- Star rating at entry is basically null: OR 0.89 (not significant after correction)

So: you can have a great average rating and still not get retrieved well enough to be recommended. That’s a key “invisible to the machine” dynamic.

What predicts first position among venues that were recommended (ordering margin)?

When conditioning on venues that are already in the answer lists, star rating does start to matter:

• Rating predicts first position: OR 1.17
• Additionally, review volume remains positive for ordering (OR around 1.30 in the conditional model)

But documentation signals weaken for the rank-first question once you restrict to the candidate set that made it through entry.

A concise way to picture the mechanism

Think of the assistant as doing two steps:

1. Retrieval gate: “Which venues do we even know about from our sources?”
2. Generation/ranking: “Given this short list, which ones sound best to say first?”

In this model, documentation and volume are the “door keys” for retrieval. Rating is the “tiebreaker” for how the assistant orders the options it already retrieved.

The paper explicitly interprets this as consistent with a retrieval-then-generation architecture—and it also explains why earlier studies can appear contradictory:
- Conjoint-style audits that always include candidates find rating matters (because entry is guaranteed).
- This census-denominated audit finds rating matters only after entry (because entry is not guaranteed).

Why the “POI dataset visibility hack” doesn’t hold up (at least here)

A common industry assumption is that being present in open point-of-interest databases (like Foursquare) boosts AI visibility.

The study tested that directly:
- Foursquare presence shows no positive association with recommendation entry.
- Among Foursquare-present venues, the “ladder” variables (quality tier, tips, popularity) are also not significant for either margin once the key documentation/volume controls are included.

So the takeaway is pretty direct: this particular “visibility hack” doesn’t appear to work in the audited markets and conditions.

The authors propose compatible explanations:
- assistants may not rely on Foursquare for venue discovery here, or
- its signal may be redundant with other web presence/documentation metrics already captured.

AI “failures” in local discovery: staleness beats fabrication, and answers churn like sampling noise

A lot of people worry about AI hallucinating restaurants. The paper finds that outright fabrication is rare in this dataset:

• 0.08% of valid venue mentions are categorized as likely invented (by their conservative verification taxonomy)

But there’s a different, more practically serious failure mode:

Permanently closed venues get recommended repeatedly

They find:
- 93 recommendations of permanently closed venues across the audit
- across 14 confirmed-closed establishments

So instead of “the AI invents fake cafés,” it’s more like:

the AI remembers the cafés that used to exist, and that old information keeps being retrieved and repeated.

Because the paper’s invisibility and entry models reward documentation (reviews, website, web mentions), those stale signals can persist even after a venue closes—creating a strong mismatch between what customers need today and what the assistant surfaces.

Results aren’t stable run-to-run, but they don’t drift over time either

Another important finding: answers churn.

• Cross-run venue-set similarity is moderate (Jaccard around 0.22–0.45 depending on comparison)
• Cross-engine agreement for top-20 sets is low (top-20 Jaccard 0.33–0.54)
• Paraphrase changes can perturb results even more than repeating the same query

However, when they did a two-week test–retest, they found that:
- answer similarity across periods is about as stable as same-day reruns
- meaning the churn behaves like sampling stochasticity, not real temporal drift

Comparison snapshot: agreement across engines (top-20)

(The paper reports the range across all pairwise comparisons as 0.33–0.54.)

And only 8 venues appear in all four engines’ top-20 lists, while 15 appear in exactly one. That implies you can’t optimize “AI visibility” once and assume every assistant will reward you similarly.

What this means for cafés, investors, and GEO teams trying to “game” visibility ethically

If you’re a venue operator or marketing team, this study suggests a hierarchy that’s more actionable than “get better reviews”:

1. Be documented before you’re excellent

   • having an own website
   • ensuring clear price information
   • building review volume (which also proxies for many other factors)
   • accumulating web mentions (third-party presence)

2. Understand where ratings matter

   • ratings help with ordering among candidates, but they don’t guarantee you enter the candidate set.

3. Fix the staleness problem

   • while the paper mainly critiques assistants/platforms, venues can still reduce their risk of being recommended incorrectly (e.g., ensure closure notices propagate in major directories and official channels).
   • for platforms, this is a system design lesson: cached retrieval without freshness checks is costly.

4. Measure across engines and paraphrases

   • because churn exists, single-shot “visibility checks” are likely to be noisy.
   • the paper’s results support repetition across phrasing/engines to get a meaningful signal—especially if you sell GEO services.

If you’re interested in how the original paper frames the method and analysis details, it’s all in Pitenin (2026) on arXiv. The design—census denominator + repeated search-grounded audits + validated extraction/matching—is what makes the conclusions harder to dismiss as anecdotal.

Key Takeaways

• AI invisibility is the default: in two Bali markets, 85.6% of 4,776 venues were never recommended by any of four systems across 2,208 search-grounded runs.
• Even “established” venues are frequently invisible: 72.6% of venues with 50+ ratings still never appeared.
• Visibility has two margins:
  • Entry into answers is associated with review volume (OR 1.64), own website (OR 1.92), listed price info (OR 1.54), and web mentions (OR 1.44)—while star rating is null at entry (OR 0.89).
  • Rank within recommended lists: star rating predicts first position (OR 1.17).
• Foursquare presence doesn’t help here: no positive effect at entry or ordering once documentation/volume are controlled.
• The main failure mode is staleness, not hallucination: fabrication is near-zero (0.08% of valid mentions), but assistants recommend 93 permanently closed venues.
• Answers churn, and cross-engine agreement is low: top-20 Jaccard across systems ranges 0.33–0.54, so optimizing for one assistant won’t reliably optimize for “AI” overall.
• For venues and GEO teams: focus first on documentation and retrievability (website, price info, review volume, web mentions). Don’t assume that a higher star rating automatically increases the odds of being recommended at all.