AI Search for Bistros in Paris (2026):The guidebook city where restaurant websites disappear
TL;DR: We asked five AI engines 26 bistro questions in two languages and checked where every one of their 3,406 citations pointed. ChatGPT linked to a bistro’s own website 0.9% of the time — the lowest share in the six cases of this series, in the city with the best-equipped restaurants we’ve tested (387 of 579 venues have a site). What replaced them wasn’t TheFork either: the reservation layer stayed near 2%. It was the guide layer — parisjetaime.com, the city’s official tourism site, tops the cross-engine consensus (55 citations, all five engines), with Time Out and the Michelin Guide close behind. On the answer side, Bistrot des Tournelles is the consensus recommendation (named by all 5 engines, 108 captures), and Paris gave us the first real neighbourhood test of the series: 94% of located recommendations sit in the right arrondissement — except on Perplexity, which drops to 47%.
Executive Summary
This study completes the Restaurants/Paris slot we staked out when the series began — scoped to bistros so the reference set stays a real universe. The question we brought to it: do reservation platforms capture AI restaurant search the way booking marketplaces captured Berlin yoga?
The answer is no — and what wins instead is more interesting. ChatGPT’s citations for Paris bistros put the restaurant’s own website at 0.9% (3 URLs out of 331). Tokyo bookstores (8%) and Marseille coffee (10%) had already broken the ~32% own-website share we measured across Paris yoga, Berlin yoga and Amsterdam bikes, but both had an excuse: thin owned webs, scenes documented on social. Paris bistros remove the excuse. Two thirds of the venue registry has a website — the highest of any study in this series — and ChatGPT still routes essentially none of its grounding through them.
Where does it ground instead? In the guide layer Paris has been accumulating for a century, now in web form: 38% local editorial (parisjetaime.com — the official tourism board — Time Out, Sortir à Paris), 14% global food & travel press, 10% restaurant guides, almost all of it the Michelin Guide. TheFork — the platform we expected to dominate a French restaurant vertical — shows up on every engine but never rises above a few percent. Booking a table is what you do after the answer; the engines cite whoever wrote the recommendation down first.
One distinction throughout: citations are the URLs an engine used as support; mentions are the bistros actually named in the visible answer. Source-mix numbers count citations; the leaderboard counts mentions. Keeping them separate is what lets us see that a bistro can dominate answers while its website is never cited.
Source mix by platform
Every cited URL, bucketed into an 11-category taxonomy built for this vertical (reservation platforms and restaurant guides get their own buckets — they’re the hypothesis). Percentages are of each engine’s own citation pool.
Columns sum to 100% (largest-remainder rounding). “Other” holds the unbucketed long tail — mostly one-off blogs and venue sites outside the registry.
Website-first, at 74%
Still the engine that grounds recommendations in the business’s own site — but the share keeps sliding down the series: 95–97% in yoga and bikes, 89% in Tokyo, 83% in Marseille, 74% here. Even for Copilot, Paris’s editorial gravity pulls a quarter of the citations away from the restaurants themselves.
The second entity engine — 35%
A role reversal: in Berlin, Perplexity was the booking-marketplace engine (31% of its URLs). For Paris bistros it cites more restaurant websites than any engine except Copilot — its top domains are literally levieuxbistrot-paris.fr (32 cites) and bistrot-cafe-de-paris.com (28). One caveat lands in Section 8: the venues are right, the neighbourhoods often aren’t.
70% Google self-citations
Google AI Mode keeps citing Google: 70% of its URLs are google.com, in line with Marseille (80%) and Tokyo (61%). The rest scatters thinly across the same guide layer the other engines use.
The 0.9% floor
Six replications in, the own-website share has a floor, and Paris bistros just set it. What makes this one different from Tokyo’s 8% or Marseille’s 10% is that the usual explanation is unavailable.
In Marseille we could explain the low share with the scene itself: only 41% of coffee venues had any website, so the engines had little to cite. Here the bistros did their homework — sites, menus, reservation widgets — and it made no measurable difference to ChatGPT. The competition isn’t absence; it’s abundance. Paris is plausibly the most guidebook-documented restaurant city on earth, and a retrieval system choosing between lebistrotduperigord.fr and a Time Out roundup of “the 23 best bistros in Paris” picks the roundup, which answers the whole question in one document.
The guide layer that took the citations
The domains cited across the most engines, with volume. This is the infrastructure an AI answer about Paris bistros is actually built on.
| Platforms | Cites | Bucket | Domain |
|---|---|---|---|
| 5 / 5 | 55 | editorial — local | parisjetaime.com (official tourism board) |
| 5 / 5 | 41 | editorial — local | sortiraparis.com |
| 5 / 5 | 26 | booking | thefork.fr |
| 4 / 5 | 59 | editorial — local | timeout.fr |
| 4 / 5 | 58 | restaurant guide | guide.michelin.com |
| 4 / 5 | 36 | press — global | pariseater.com |
The city’s own tourism board is the #1 consensus source. Only three domains reach all five engines — parisjetaime.com (55 cites), sortiraparis.com (41) and thefork.fr (26) — and the tourism board leads them. That’s a first in this series, where the cross-engine consensus slot has always been held by Reddit or a private guide. If you want one placement that reaches every AI engine answering bistro questions about Paris, the official city guide is that placement — ahead of any booking platform, and far ahead of your homepage.
The TheFork test: a hypothesis that failed
We came into this study with a specific prediction, made when the Restaurants/Paris slot was staked out: engines would route restaurant answers through reservation platforms — TheFork above all, in its home market — the way Berlin’s yoga answers ran through booking marketplaces. Worth stating clearly, because the data said otherwise.
Present everywhere, dominant nowhere
TheFork does something no restaurant website manages: it appears in the citations of all five engines (Perplexity 22, ChatGPT 16, AI Mode 6, Copilot 3, Gemini 2). But the volume never follows. Compare Berlin yoga, where Urban Sports Club, Eversports and ClassPass together took 31% of Perplexity’s citations — here the equivalent reservation layer peaks at 3% of Perplexity. The Michelin Guide alone out-cites TheFork 65 to 49.
Our best reading of the difference: in Berlin, the marketplaces are the discovery layer — class schedules, studio pages and comparison content live there. In Paris restaurants, discovery editorial already existed for decades before online booking arrived, so TheFork functions as checkout, and checkout pages answer no one’s “where should I eat” question.
Reddit narrows to one engine; Instagram stays home
Two social findings, both departures from the earlier cities.
Reddit citations that belong to ChatGPT
Reddit is still ChatGPT’s single favourite domain here (61 cites, 18% of its pool, more than any other single source). But the habit stops there: AI Mode adds 9, Gemini 4, and Copilot and Perplexity cite Reddit exactly zero times for Paris bistros. In earlier cities Reddit was the cross-engine consensus source; in Paris that role passed to the tourism board, and Reddit became a ChatGPT quirk.
Instagram citations, Paris bistros vs Marseille coffee
Marseille’s standout signal — Instagram as a primary AI citation source, 237 cites — does not travel to Paris. 32 citations across the whole study, despite plenty of bistros here also living on Instagram. Instagram-as-citation looks like a property of scenes whose only documentation is Instagram — it evidently isn’t a French thing or a food thing.
Gemini skips Michelin’s city and cites list sites
Gemini’s pattern in this series is to anchor on one editorial vein per city. Its Paris pick is the strangest yet.
restaurantsforkings.com on GeminiIn the world’s most guide-covered restaurant city, Gemini’s top sources are low-recognition English list sites
Its top domains: restaurantsforkings.com (45 — a Florida-based, occasion-themed restaurant-ranking platform), everydayparisian.com (31 — an American expat lifestyle blog), then the Michelin Guide (18), Paris by Mouth (15) and topratedplaces.ai (10 — a generic “top places” rankings site on a .ai domain). The list-site bucket takes 22% of Gemini’s citations — no other engine gives it more than 2%.
In Marseille, Gemini’s concentration at least landed on trade press (Barista Magazine). Here the same behaviour — find one authoritative-looking editorial vein and drink deeply — selects sites whose main qualification appears to be ranking-shaped content in English. For a city where Le Fooding, Le Monde and fifty years of guide culture are a query away, that’s a retrieval-quality finding in itself.
The Paris Bistro AI Leaderboard
Ranked by how many captures actually name the bistro in the visible answer (chains and sibling venues merged into brands). The citation-counted score sits alongside for contrast — the two metrics tell usefully different stories here.
| Rank | Bistro | Text mentions | Cite score | Engines |
|---|---|---|---|---|
| #1 | Bistrot Des Tournelles | 108 | 58 | 5 / 5 |
| #2 | Bistrot Paul Bert | 75 | 14 | 5 / 5 |
| #3 | Le Vieux Bistrot | 71 | 75 | 4 / 5 |
| #4 | Bistro des Livres | 64 | 119 | 4 / 5 |
| #5 | Brasserie Martin | 56 | 31 | 5 / 5 |
| #6 | Bistrotters | 50 | 26 | 5 / 5 |
Why the cite-score ranking disagrees
By citations, #1 would be Bistro des Livres (score 119): its family of three sibling bistros — des Livres, des Lettres, des Poèmes — shares one domain, bistrodeslettres.com, which engines cite heavily (66 cites across 3 engines). By what the answers actually say, it’s #4. Same lesson as Marseille’s cite-#1 artifact, milder form: citation counts reward domain arrangements; text mentions reward being recommended.
Two rows to read carefully
#5 “Brasserie Martin” is really the Nouvelle Garde group — its restaurants (Martin, des Prés, Dubillot, Bellanger) share lanouvellegarde.com, so brand aggregation merges them; read that row as the group’s combined AI presence. #9 Bouillon Pigalle likewise merges Bouillon République via bouillonlesite.com. Both are honest merges of sibling venues, flagged so nobody reads them as single dining rooms.
The same leaderboard, split by engine
Cell = share of that engine’s captures whose visible answer names the brand (raw count in parentheses). Capture counts differ per engine — ChatGPT, Gemini and Copilot ran both proxies (104 each), AI Mode and Perplexity US-only (52) — so rates, not raw counts, are the comparable unit.
Bistrot des Tournelles is the only brand every engine agrees on at volume — 33% of AI Mode’s captures, 36% of Copilot’s, about a fifth of everything else. Gemini’s zeros on Le Vieux Bistrot, Bistro des Livres and Bistrot Richelieu are real: its answers simply favour a different set (Paul Bert 31%, Le Bon Georges 20%) — consistent with it reading different sources (Section 6).
| # | Bistro | AI Mode52 prompts | ChatGPT104 prompts | Perplexity52 prompts | Gemini104 prompts | Copilot104 prompts |
|---|---|---|---|---|---|---|
| 1 | Bistrot Des Tournelles | 32.7%(17) | 20.2%(21) | 19.2%(10) | 22.1%(23) | 35.6%(37) |
| 2 | Bistrot Paul Bert | 11.5%(6) | 18.3%(19) | 15.4%(8) | 30.8%(32) | 9.6%(10) |
| 3 | Le Vieux Bistrot | 19.2%(10) | 10.6%(11) | 50%(26) | 0(0) | 23.1%(24) |
| 4 | Bistro des Livres | 19.2%(10) | 11.5%(12) | 21.2%(11) | 0(0) | 29.8%(31) |
| 5 | Brasserie Martin | 30.8%(16) | 9.6%(10) | 1.9%(1) | 10.6%(11) | 17.3%(18) |
| 6 | Bistrotters | 15.4%(8) | 26.9%(28) | 1.9%(1) | 1.9%(2) | 10.6%(11) |
| 7 | Le P'tit Bistrot | 21.2%(11) | 7.7%(8) | 1.9%(1) | 0(0) | 23.1%(24) |
| 8 | Bistrot Richelieu | 19.2%(10) | 6.7%(7) | 17.3%(9) | 0(0) | 16.3%(17) |
| 9 | Bouillon Pigalle | 23.1%(12) | 4.8%(5) | 3.8%(2) | 18.3%(19) | 3.8%(4) |
| 10 | Le Bon Georges | 7.7%(4) | 10.6%(11) | 5.8%(3) | 20.2%(21) | 0(0) |
| 11 | Bistrot de l’Oulette | 11.5%(6) | 15.4%(16) | 3.8%(2) | 0(0) | 13.5%(14) |
| 12 | Bistrot Rougemont | 1.9%(1) | 20.2%(21) | 1.9%(1) | 0(0) | 9.6%(10) |
Rows ordered by total text mentions; colour scales to the table maximum (Le Vieux Bistrot on Perplexity, 50%), zero cells greyed out. Hovering shows the raw counts behind each rate.
Where the winners are
The 12 leaderboard brands on the map — multi-venue brands show every location. Click a marker for the per-engine mention breakdown.
Top 12 most-recommended brands — popups show per-engine text mentions
The first real neighbourhood test — and Perplexity fails it
Every previous city rendered the neighbourhood-accuracy question unanswerable: the seeds stored “Amsterdam” or “Marseille” with no district field. Paris postal codes encode the arrondissement, so 494 of our 579 venues carry one — and for the five neighbourhood prompts (11th arrondissement, Le Marais, Saint-Germain-des-Prés, Montmartre, Canal Saint-Martin) we can finally score every located recommendation.
| Engine | In target arrondissement | Accuracy |
|---|---|---|
| ChatGPT | 52 / 52 | 100% |
| Copilot | 47 / 47 | 100% |
| AI Mode | 24 / 24 | 100% |
| Gemini | 36 / 37 | 97% |
| Perplexity | 9 / 19 | 47% |
Three engines are literally perfect, Gemini misses once — and then there’s Perplexity at 47% (9 of 19; a small sample, flagged as such, but every other engine’s sample of similar size scored 97–100%). The failure shape is specific: ask Perplexity for bistros in the Marais and it answers with excellent bistros… in the 5th, the 9th, the 11th. Its website-heavy retrieval seems to reward venues whose sites rank for “bistro Paris” regardless of where they stand. A user following its Marais answer would eat well, two métro lines away.
The supply side, mapped
Every geocoded venue in the registry. Hover a dot for the name. The engines’ favourites cluster where the seed clusters — the 11th, the Marais, Saint-Germain — which is part of why the accuracy test is passable at all.
All 579 geocoded seed venues (438 classified as bistros/brasseries)
EN vs FR: back to the 25% baseline
Same question, both languages, ChatGPT on the FR proxy: how much of the top 5 survives translation? After Marseille’s 11%, Paris bistros land exactly where Berlin yoga and Paris yoga did.
| Template | EN vs FR top-5 overlap |
|---|---|
| dist_marais / dist_canal / dist_saint_germain | 67% |
| bleed_brasserie / booking / persona_tourist / dist_montmartre | 43% |
| control (best bistros in Paris) | 25% |
| persona_date / style_traditional / style_wine / vibe_historic / vibe_aesthetic | 25% |
| michelin / price_cheap / price_premium / time_lunch / time_late / dist_11e | 11% |
| persona_local / persona_family / style_neo / style_seasonal / time_sunday / vibe_terrace / bleed_winebar | 0% |
The control prompt overlaps 25% — matching Berlin yoga (25%) and Paris yoga (25%), double Marseille’s 11%. That fits the guide-layer story: Paris is thoroughly documented in both languages, so neither language starves. But the split by template is the sharper finding. Neighbourhood prompts converge (67% — geography anchors both languages to the same famous spots), while identity prompts diverge to zero: “bistros where locals actually eat,” neo-bistros, Sunday opening, family dining — the FR answers and EN answers share nothing at all in their top fives.
Six cases in: what the series now says
One row per replication, on the axis that has become the series’ spine — how much of ChatGPT’s grounding the businesses themselves get, and what takes the rest.
| City + vertical | ChatGPT own-site % | What takes the citations instead |
|---|---|---|
| Paris yoga | 32% | |
| Berlin yoga | 32% | Booking marketplaces (Urban Sports Club) |
| Amsterdam bikes | 42% | |
| Tokyo bookstores | 8% | Local-guide web (whenin.tokyo, Tokyo Weekender) |
| Marseille coffee | 10% | Instagram + Reddit + French micro-guides |
| Paris bistros | 0.9% | The guide layer: tourism board, Time Out, Michelin |
What replicated
- Engine personalities. Copilot most website-first (74%), AI Mode 70% Google self-cites, Gemini concentrated on one editorial vein, ChatGPT blending social with editorial.
- The 25% language overlap on the control prompt, matching Berlin and Paris yoga — which recasts Marseille’s 11% as the series outlier.
- Citation-vs-mention divergence. The cite-counted #1 (a shared-domain family) again differs from the consensus the answers actually name.
What Paris bistros adds
- The own-website floor is not about missing websites. 67% of the registry has one; ChatGPT cites them 0.9%. Editorial abundance suppresses them as effectively as owned-web scarcity did.
- Reservation platforms are not a discovery layer. The Berlin marketplace takeover has a boundary, and it’s here.
- Neighbourhood accuracy is measurable and engine-specific. 94% overall, with a 100%-vs-47% split between ChatGPT-class engines and Perplexity.
Dataset shape and operational notes
What landed, what didn’t, and one infrastructure incident worth recording.
The TLD split is unmeasurable this time, for a new reason. In Marseille the .fr sample was too thin because venues lacked websites. Here the bistros have websites — ChatGPT just doesn’t cite them (3 URLs total), so a .com-vs-.fr comparison over entity domains has no sample to stand on. Recorded as a measurement limit, not a finding.
What this means if you run a Paris restaurant
The citation economy for Paris food runs through the guide layer, so that’s where the work is:
- Chase the guide layer before the platform layer. parisjetaime.com, Sortir à Paris and Time Out are what all engines cite; a Bib Gourmand listing feeds the 10% guide bucket. Your TheFork profile converts diners — it just doesn’t create AI visibility.
- Keep the website — its job has changed. It anchors your name to a place and menu for every resolution system — ours included — even while earning almost no citations directly.
- Reddit reaches exactly one engine. Threads about you feed ChatGPT (18% of its grounding) and nothing else. Price the effort accordingly.
- Check what Gemini reads. Its top Paris sources are English list sites most restaurateurs have never heard of — being absent from restaurantsforkings.com costs Gemini visibility specifically.
- Audit both languages when it’s a taste question. Locals, natural wine and Sunday-open prompts have zero EN/FR overlap; neighbourhood prompts mostly converge.
A century of guidebooks, machine-readable at last
We expected TheFork to own this vertical and it doesn’t. We assumed a scene where two thirds of businesses maintain websites would lift the own-site share and it set a record low instead. Both surprises have the same cause: Paris restaurants are the most editorially documented local businesses we’ve tested, and retrieval systems prefer a document that answers the whole question to five documents that each answer a fifth of it.
That preference has history on its side here. Paris invented the modern restaurant guide; what the engines are doing is reading a hundred years of that tradition — tourism-board lists, Michelin entries, Time Out roundups — and treating the restaurants’ own voices as raw material rather than sources. The bistros wrote the menus. The guides got the citations.
For the series, Paris closes a question Tokyo and Marseille opened. Low own-website shares aren’t a symptom of scenes too scrappy to build websites; they happen wherever some other layer documents the scene better than the businesses do. What varies by city is only which layer that is. Six cases in, predicting an AI engine’s sources for a city means asking one question first: who already wrote the best guide?
Study design
Data collection
- 26 prompt templates × 2 languages (EN/FR) × 2 proxy countries (US/FR) × 5 AI engines = 520 theoretical; 416 captured (Perplexity × FR download failure, −52; AI Mode × FR not fired, −52)
- Engines: ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode
- Captured 2026-07-08 via Bright Data; a duplicated ChatGPT × US batch (infrastructure retry) was removed before analysis
- 416 captures · 3,406 cited URLs · 3,034 extracted mention rows (70% resolved)
- Seed: 500-row Apify Google Maps pull for “bistro” in Paris (359 classified real bistros/brasseries) + 79 Google Places recoveries → 579 venues, 438 real, 387 with websites
What we measured
- Bistros named per answer (brand-aggregated leaderboard, text mentions primary)
- Cited URLs bucketed into an 11-bucket source taxonomy (booking and restaurant guides broken out)
- Reservation-platform share per engine (the staked-out hypothesis)
- EN vs FR top-5 overlap per prompt template
- Arrondissement accuracy of neighbourhood-targeted recommendations, per engine
- Cross-city comparison vs Paris yoga, Berlin yoga, Amsterdam bikes, Tokyo bookstores, Marseille coffee
How we turned answers into bistros: the NER pipeline
AI answers are free text — “Paul Bert remains the benchmark, though Les Arlots up by the Gare du Nord…” — not a clean list of businesses. To count anything, we first extract the venue mentions. Each answer runs through a deterministic named-entity-recognition pass in four steps:
- Extraction. A Gemini pass at temperature 0, constrained to a strict JSON schema, pulls every venue the answer actually recommends — named places only, no inferred ones — with its position in the answer.
- Normalisation. Candidates are lower-cased and stripped of boilerplate (“Paris,” arrondissement suffixes, punctuation) so “Bistrot Paul Bert Paris” and “Paul Bert” collapse toward the same key.
- Entity resolution. Normalised mentions match the venue registry exactly or by fuzzy similarity (cutoff 0.86). Names that don’t resolve are flagged, and frequently-recommended ones are looked up in Google Places: results verified as real Paris restaurants join the registry and their mentions re-resolve (79 venues entered this way — the classic canon a “bistro” Maps scrape misses); everything else stays uncounted.
- Brand aggregation. Venues sharing a website domain merge into one brand (Nouvelle Garde’s restaurants, the Livres/Lettres/Poèmes family, both Bouillons), with per-location coordinates kept for the map. Social and booking domains are excluded from brand identity so an Instagram-only venue can’t absorb instagram.com citations.
A taxonomy pass then tightened the source bucketing: before it, the residual “other” bucket held 39% of citations (the Paris editorial tail plus recovered venues’ websites the seed didn’t know); after adding the recovered venues’ websites and the Paris-specific editorial needles, “other” sits at 20%.
Caveats
- Perplexity × FR proxy is missing (Bright Data download-stage fetch_error after two runs and five attempts). Perplexity numbers are US-proxy only; both prompt languages are present in that batch.
- AI Mode × FR was not fired. The trigger-level FR-proxy rejection hit Paris yoga and Marseille coffee before this study; combined with the cost cap, we spent the calls elsewhere. AI Mode numbers are US-proxy only.
- The TLD test has no sample. ChatGPT cited 3 entity-website URLs in total; no .com-vs-.fr split can be read from that. Measurement limit, not a finding.
- Perplexity’s arrondissement sample is small (19 resolved mentions vs 24–52 for the others). Its 47% is flagged with its n; the perfect scores around it make the gap meaningful, not conclusive.
- Venues named in fewer than 5 captures were not Places-recovered; their mentions stay unresolved, so leaderboard counts are conservative lower bounds. One known miss: Benoit (Alain Ducasse’s bistro, ~12 mentions) was rejected by the recovery type-gate and is absent from the leaderboard.
- Seed scope is “bistro”-anchored. The registry is built from a Maps search for bistros plus engine-recommended recoveries; it is a bistro universe, not a census of all Paris restaurants.
- Google AI Mode’s google.com self-citations (70%) inflate its citation volume relative to engines that cite the open web.
- Disclosure: no personal affiliation with any Paris restaurant.
Frequently Asked Questions
Summarize with AI
Continue Reading
The rest of the cross-vertical series, and the hotel research it grew out of.
All Research