AI Search for Climbing Gyms in Prague (2026):the density law finds its ceiling
TL;DR: Four AI engines, 22 climbing questions, English and Czech — 308 of 308 captures landed — every capture the platforms allow. Prague’s gyms are the best-webbed vertical this series has measured (84.6% run an own-domain site), and ChatGPT put 47.7% of its citations on those sites — the highest share the series has recorded, on a batch that is US-proxy only because the CZ country is refused for that scraper. Every engine agrees on the city’s #1 (253 of 308 answers name SmíchOff, the strongest consensus yet), three engines keep recommending a bouldering gym under a brand that no longer exists, and ChatGPT gave Reddit zero citations in the most Reddit-native hobby we could find.
Executive Summary
We went looking for the one vertical where every business runs a real website — because climbers need this week’s route-setting, today’s opening hours and a day-pass price, and Instagram carries none of that. The question was whether ChatGPT’s citations would follow the supply all the way up.
They followed it — and then stopped climbing. Of the 39 gyms and walls in the registry, 33 (84.6%) run an own-domain site, the densest owned-web layer this series has measured. ChatGPT placed 47.7% of its 199 citations on those domains (95 of them) — a series record, ahead of Helsinki’s 45.5% — but only just. Six points from CDMX’s 12.5%→0% to here, the rank order has never once broken, and the slope has fallen with every step. Roughly half of what ChatGPT cites is reserved for the layer that talks about businesses — guides, directories, the crag database varp.cz, the expat press — no matter how good the businesses’ own web gets.
The city itself produced the cleanest agreement the series has seen: Climbing Center SmíchOff is named in 253 of 308 answers (82.1%), by all four engines, in both languages — and it also tops the citation-counted board, so for once the two scoring methods crown the same winner. Further down the answers, something stranger: 12 captures across three engines recommend “Boulder Bar”, the pre-rebrand name of a Holešovice bouldering gym that Google Maps no longer lists. ChatGPT — the engine that searched the live web in 30 of its 44 captures — never once does.
By the numbers: 308 of 308 captures, 1,384 cited URLs, 1,453 extracted venue mentions, 93.3% resolved against a 122-row registry. One platform limit is disclosed rather than filled: the ChatGPT scraper refuses a CZ proxy, so ChatGPT ran US-proxy only — 44 of the 44 it is allowed. A fifth engine, Perplexity, landed 37 of 88 after error-item batches and is excluded from the study rather than reported.
Terminology, as on every page in this series: citations are the URLs an engine attaches under its answer, mentions are the venue names inside the answer text. Source analysis uses the former, the leaderboard the latter.
Supply keeps rising; the share barely moves
The density law says ChatGPT cites businesses’ own websites in proportion to how many of them have one. Prague was chosen to test the end of that claim — a vertical whose product (route resets, opening hours, day passes) forces every gym onto its own domain.
Prague’s density bar is the business-venue reading: 33 of 39 gyms and walls. The registry also holds 11 outdoor crags, which cannot own a domain (their “website” is a varp.cz database page); counting them puts the full-registry density at 66%, below Helsinki — a reading under which Prague’s record share would break the rank order instead of extending it. Every prior registry in the series contained only businesses, so the business-only bar is the like-for-like one. One asymmetry to hold in mind on the share bar: Prague’s 47.7% is measured on 44 US-proxy captures, where every earlier city pooled two proxy countries — the language-and-IP study found the query’s country barely moves ChatGPT’s source set, but this run has no CZ-proxy ChatGPT captures to check that against. Both density numbers ship in summary.csv.
Read the six measured density points as a sequence of slopes. Istanbul returned almost exactly what its market supplied (22.1% density, 21.3% share). Buenos Aires got two-thirds of its density back. Helsinki got 58% of its. Prague, with six points more density than Helsinki, gets a share two points higher — a marginal return of about 40%. The ordering has survived all six measured cities; the proportionality has not.
What occupies the rest of ChatGPT’s citation budget here is instructive, because it can’t be blamed on missing gym websites: 51 citations (25.6%) go to local editorial and expat guides (theexpatcity.com alone out-drew every individual gym domain in ChatGPT’s citations, 26 to SmíchOff’s 23), and 33 more to the climbing web — the varp.cz crag database, the national federation, climbing press. Even talking to the best-webbed market we could find, ChatGPT spends half its evidence on third parties.
Five engines, one answer: SmíchOff
A converted Smíchov factory hall with the country’s best-known lead walls is named in 253 of 308 answers — 82.1%, the highest agreement on any venue in twelve studies (Helsinki’s Löyly held the record at 79.8%). It leads the citation-counted board too (152), so the mention canon and the citation economy point the same way for once — Buenos Aires, Seoul and Mexico City each split the two crowns.
All four engines recommend ten of the twelve board venues — exactly the count Buenos Aires produced on the same four engines. What separates Prague is the depth of the agreement, not its breadth: SmíchOff sits in 82.1% of answers where La Viruta topped out at 60.4%. With ~39 climbing venues and a web layer this dense, the canon is short, famous and identically retrievable everywhere.
The pecking order tells a Prague story. BigWall (215 mentions) is the rope-wall counterweight to SmíchOff; HUDY Boulder Karlín rides the country’s biggest outdoor-retail brand; and the three Jungle branches split what would otherwise be a podium finish — counted as one brand they’d sit near BigWall. The smallest board member, Cibulka, earns most of its presence from a single engine (Copilot, 29 of its 39 mentions).
Who recommends whom — the per-engine matrix
Each cell: how often that engine’s answers name the venue. Copilot names SmíchOff in 90.9% of its answers; Gemini is the JamJam enthusiast; AI Mode leans hardest on Jungle’s Letňany branch.
| # | Venue | AI Mode88 captures | ChatGPT44 captures | Gemini88 captures | Copilot88 captures |
|---|---|---|---|---|---|
| 1 | Climbing Center SmíchOff | 79.5%(70) | 77.3%(34) | 78.4%(69) | 90.9%(80) |
| 2 | BigWall | 65.9%(58) | 50%(22) | 75%(66) | 78.4%(69) |
| 3 | HUDY Boulder Karlín | 36.4%(32) | 52.3%(23) | 54.5%(48) | 44.3%(39) |
| 4 | Jungle Sport park | 48.9%(43) | 40.9%(18) | 28.4%(25) | 18.2%(16) |
| 5 | Lezecká stěna Třináctka | 17%(15) | 38.6%(17) | 21.6%(19) | 40.9%(36) |
| 6 | Climbing wall Ruzyne | 12.5%(11) | 31.8%(14) | 22.7%(20) | 46.6%(41) |
| 7 | Jungle Holešovice | 15.9%(14) | 31.8%(14) | 13.6%(12) | 50%(44) |
| 8 | LOKAL BLOK | 17%(15) | 18.2%(8) | 19.3%(17) | 28.4%(25) |
| 9 | UltraAnt Climbing Club | 11.4%(10) | 20.5%(9) | 2.3%(2) | 35.2%(31) |
| 10 | JamJam | 13.6%(12) | 27.3%(12) | 26.1%(23) | 0(0) |
| 11 | Boulder V Síti | 11.4%(10) | 22.7%(10) | 9.1%(8) | 10.2%(9) |
| 12 | Climbing Centre Cibulka | 0(0) | 6.8%(3) | 4.5%(4) | 33%(29) |
ChatGPT’s column runs on its 44 US-proxy captures; the others on the 37 of 88 that landed (section 8).
Top 12 venues by answer mentions — click a marker for per-engine counts
Four engines recommend a gym that isn’t there
“Boulder Bar” was for years the name in Prague bouldering — and its Holešovice hall now operates under the Jungle brand, with no Google Maps listing left under the old name. Search Google Maps for “Boulder Bar Praha” and it hands you the Jungle gyms. Yet 12 captures across 3 engines — AI Mode (9), Gemini (2), Copilot (1) — recommend Boulder Bar by name, as a place you could walk into tomorrow.
The exception is the engine this series usually catches in a lie: ChatGPT recommended the ghost exactly zero times in 44 answers. It also searched the live web in 30 of those 44. The engines that skipped or half-ran retrieval are the ones reciting a market map that expired years ago — a failure mode this series hadn’t isolated before: not a fabricated name, but a real, once-beloved brand served past its death.
The rest of the unresolved pile is honest geography rather than error: asked for outdoor climbing “near Prague”, the engines correctly leave the city — Srbsko, Alkazar, Svatý Jan pod Skalou, Roviště, the Prachov Rocks — all real crags, all outside a Prague registry by construction. After a hand-checked Czech-alias pass, 97 of 1,453 mentions stay unresolved, and the ghost plus the out-of-town crags account for most of them.
The engine personalities, on their best behavior
Give the engines a market where the venues’ own web is actually good and the structural personalities snap back to their extremes: Copilot returns to the 95% entity band it held in the series’ early cities, and Gemini puts a majority of its citations on venue sites for the first time, and no booking-platform layer exists to distract anyone — climbing has no ClassPass.
261 of 274 citations on venue-owned sites — back in the 95–97% band it held in Amsterdam, Paris and Berlin, after sliding as low as 25.7% in Seoul where there was nothing to cite. Copilot’s share has always been the market’s own-web layer read back at maximum gain; Prague confirms it from the supply-rich end.
Climbing is one of Reddit’s loudest hobbies — and ChatGPT, which put 17–20% of its citations on Reddit for four straight studies in the spring, cited it 0 times here. Study-wide Reddit received 12 citations of 1,384, none from ChatGPT. Three cities in a row now (Helsinki, Buenos Aires, Prague): whatever changed in ChatGPT’s retrieval this summer, it wasn’t vertical-specific.
85 of Gemini’s 208 citations are Google properties — overwhelmingly plain google.com/maps links standing where a venue’s website would go. Add its 23.6% entity share and Gemini reads like an engine that knows the right venues but prefers to route you through its own map to reach them.
English and Czech agree on everything except fame
First Slavic language in the series. The control prompt overlaps 43% between English and Czech — Buenos Aires’ number exactly, far above the Istanbul/CDMX floor of 0% — and the tourist prompt returns the identical five gyms in both languages. Then the pattern inverts in one place: ask for the most famous gyms and the two languages agree on nothing at all.
| Prompt template | EN vs CS top-5 overlap |
|---|---|
| tourist persona ("visiting for a few days") | 100% (5/5) |
| day pass, no membership | 67% (4/6) |
| bouldering gyms | 67% (4/6) |
| rope climbing walls | 67% (4/6) |
| serious training | 67% (4/6) |
| open late | 67% (4/6) |
| biggest and most modern | 67% (4/6) |
| control (best climbing gyms) | 43% (3/7) |
| gyms with a climbing wall (bleed control) | 43% (3/7) |
| outdoor rock near Prague (bleed control) | 43% (3/7) |
| district: Žižkov / Karlín | 43% (3/7) |
| for kids | 43% (3/7) |
| rope parks (bleed control) | 25% (2/8) |
| district: Smíchov | 25% (2/8) |
| shoe rental | 25% (2/8) |
| for beginners | 25% (2/8) |
| first time climbing | 25% (2/8) |
| affordable gyms | 25% (2/8) |
| not too crowded | 25% (2/8) |
| adult courses | 11% (1/9) |
| district: city centre | 11% (1/9) |
| most famous gyms | 0% (0/10) |
Overlap = shared venues ÷ distinct venues across both top-5 lists (Jaccard), read from ChatGPT’s map carousel — present in all 44 captures this run. Both language columns come from the US proxy, because the capture tool’s ChatGPT scraper refuses a CZ proxy (section 8).
The agreement sits exactly where a visitor’s and a local’s needs coincide: what’s open late, where to boulder, where to get a day pass — 67% overlap each, with the tourist prompt at 100%. The median across all 22 templates is 43% — Buenos Aires, the previous convergence high, measured a 25% median.
The one zero is the “most famous” prompt, and it’s a telling zero: English fame and Czech fame are different canons. The English answers reach for what expat guides celebrate; the Czech answers for institutions Czech climbers grew up with. Meanwhile the language→TLD coupling is mild — Czech prompts put 83.5% of ChatGPT’s citations on .cz domains, English prompts 61.1% (1.37×, beside Buenos Aires’ 1.33× and nowhere near Tokyo’s 5×) — because in a market this webbed, even the English answers end up on Czech gym sites.
The twelve domains that answer climbing questions
For the first time in this series, the most-cited real-world domain is a venue’s own website — in every other city that spot went to a government portal, a ticket seller or a directory. Ten of the twelve are Czech domains, and four of the twelve are cited by every engine in the study — varp.cz, the volunteer-run crag database, alongside two gym sites and the Czech yellow pages.
| Domain | Engines | Citations | What it is |
|---|---|---|---|
| lezeckecentrum.cz | 3/4 | 95 | Climbing Center SmíchOff — the consensus #1 on both boards |
| big-wall.cz | 3/4 | 81 | BigWall Vysočany, the biggest hall in the country |
| varp.cz | 4/4 | 47 | The Prague crag database — cited by every engine in the study |
| stena-ruzyne.com | 4/4 | 39 | Climbing wall Ruzyně |
| hudysteny.cz | 4/4 | 33 | HUDY Boulder Karlín (the outdoor-retail chain’s gym arm) |
| holesovice.jungle.cz | 3/4 | 30 | Jungle Holešovice |
| theexpatcity.com | 2/4 | 29 | An English-language expat guide to Prague |
| jungleletnany.cz | 2/4 | 24 | Jungle Sport park’s alternate domain |
| stenastodulky.cz | 1/4 | 23 | Lezecká stěna Třináctka — one engine cites it, 23 times |
| ultraant.cz | 2/4 | 21 | UltraAnt Climbing Club, Old Town |
| squashpark.cz | 2/4 | 20 | Climbing Centre Cibulka (inside the Squashpark complex) |
| firmy.cz | 4/4 | 20 | Seznam’s business directory — the Czech yellow pages |
google.com itself drew 609 citations — 524 of them AI Mode’s searchviewer wrap, 85 Gemini’s Maps links — and is excluded here as self-reference rather than a source.
A ring of gyms, an empty centre
Prague’s climbing lives where industrial halls are cheap — Smíchov, Holešovice, Vysočany, Letňany, Stodůlky — and the registry map shows a ring with almost nothing inside the tourist core (UltraAnt, on Old Town’s edge, is the exception). The engines know it: asked for gyms “in central Prague”, ChatGPT’s recommendations land in the actual centre 0% of the time, because there is nothing there to recommend — it answers with the ring, which is the right answer to a slightly different question. District prompts for Smíchov and Žižkov/Karlín stay local only 10–17% of the time; the canon outranks the neighborhood even when you name the neighborhood.
All 50 mapped registry venues across Prague
One cross-city bleed surfaced: asked for rope walls, ChatGPT’s map widget twice offered “Climbing wall Kladno” — a real wall, 25 km outside Prague, in a list of Prague gyms.
What the instruments did this run
ChatGPT could not be proxied from Czechia
The ChatGPT scraper rejects the CZ country outright (“Selected country is not available for this scraper”), a per-scraper limitation the series hadn’t hit before. ChatGPT’s 44 captures are complete but US-proxy only, so its language comparison measures prompt language alone, with proxy held constant — arguably a cleaner isolate, and the one disclosed on every overlap table above. The other three engines accepted CZ.
Perplexity landed 37 of 88, and is out of the study
Most of both Perplexity batches returned as error items (29 captures survived); a targeted refire of only the missing prompt×proxy pairs recovered 8 more, leaving 37 of 88. Fewer than half a grid cannot be compared with the engines that ran complete, so it is excluded from the study rather than charted behind a disclosed denominator. Those captures remain in the source database.
AI Mode’s publisher links are partially back
A week ago in Buenos Aires, 100% of AI Mode’s citations were google.com searchviewer self-references — the instrument change decoded in the svid study. This run, 179 of its 703 citations (25.5%) are real publisher URLs again, slightly more via the CZ proxy. The wrap is evidently a rollout dial, not a completed migration; the domain dimension remains untrustworthy either way.
ChatGPT skipped the web in 14 of 44 answers
web_search_triggered was false in 14 ChatGPT captures — the engine answered a local-business question from parametric memory alone, though its map carousel still rendered in all 44 (556 map entities). 8 captures study-wide recommended no venue at all (3 Gemini, 3 Copilot, 1 AI Mode, 1 Copilot).
If you run a climbing gym (in Prague or anywhere)
Your website only gets you to the start line. In a market where 84.6% of gyms have one, the site gets you counted — the separation happens in the third-party half of the citation budget: the guides, the crag database, the expat press. That is where SmíchOff’s 253-answer dominance was built.
Treat a rebrand as an AI-visibility migration. 12 answers in this study recommend a brand that no longer exists. If the old name still outranks the new one inside three engines, you want the old domain redirecting, the old name mentioned once on the new site, and updated mentions in whatever guides the engines cite for your city.
The English canon is written by expat media. The “most famous” prompt overlapped 0% between languages, and theexpatcity.com drew more ChatGPT citations than any single gym site. If tourists matter to your fill rate, one well-placed mention in the city’s English-language guides moves the English answers in a way your own site can’t.
Don’t buy Reddit advice for this channel. The spring’s “get discussed on Reddit” playbook is three studies dead: 0 ChatGPT citations here, in Reddit’s own favorite hobby. The layers that replaced it are venue sites and local editorial.
Conclusion
Prague was picked as a boundary experiment: feed the engines the best-documented small-business vertical we could find and see what maxes out. The answer is nearly everything — Copilot at 95.3% entity citations, a consensus #1 named in four answers out of five, both leaderboards agreeing, the two languages converging on every practical question — except the one number the experiment was aimed at. ChatGPT’s own-website share set its record and still stopped short of half, on citations, in the market that made the strongest possible case for venue websites.
So the density law survives its sixth test with an amendment: supply decides how close you get to the ceiling, and the ceiling is structural. The engines keep half their evidence budget for the web that talks about a market — and in Prague that half runs through a volunteer crag database, an expat magazine and the Czech yellow pages. The ghost of Boulder Bar haunting three engines is the same lesson from the other side: what the third-party web remembers about you matters as much as what you publish yourself.
Next in the series: still owed — a continent the series hasn’t visited.
Methodology
The grid
22 prompt templates (control, disciplines, personas, price, access, districts, plus outdoor / generic-gym / rope-park entity-bleed controls) × 2 languages (English, Czech — translations competent but not native-reviewed) × 4 engines (ChatGPT, Gemini, Copilot, Google AI Mode) × 2 proxy countries (US, CZ), fired 2026-09-09 through the pipeline’s own capture runner. All 308 reachable captures landed: Gemini, Copilot and AI Mode each 88/88; ChatGPT 44/44 on the US proxy, because the CZ proxy is refused for that scraper (“Selected country is not available for this scraper”) — a standing per-engine limitation, disclosed on every language table. A fifth engine, Perplexity, landed 37 of 88 after error-item batches and is excluded from the study rather than reported.
The registry
Two broad Google Maps pulls (“climbing gym”, “bouldering gym”, “lezecká stěna”, “boulderová stěna”, then “bouldering”, “horolezecká stěna”, “lezecké centrum”, “boulder”) yielded 121 unique places; one name-targeted recovery pass added Divoká Šárka. Classification by an observed category map keeps 50 rows real — 39 gyms/walls and 11 outdoor areas (Google’s “Observation deck” rows here are the Hlubočepy and Dívčí hrady crags) — and classifies out playground walls, rope parks, fitness centers, gear shops and rope-access firms. Four venues Google files under plain “Gym” (JamJam, Rock Gym Skalka, Horolezecká Hala Chodov, the Gutovka wall) were flipped real by hand; each flip is recorded in the run provenance and flagged in the PR for review.
Extraction and resolution
Venues named in each prose answer were extracted by the pipeline’s language-model NER pass, yielding 1,453 mention rows. Resolution precedence: exact normalized match, fuzzy (≥0.86), unique token containment, then a hand-checked Czech alias pass — the seed carries Google’s English titles while the engines answer in Czech (“Lezecké centrum SmíchOFF” vs “Climbing Center SmíchOff”), so a diacritic-folding needle table, every entry verified unambiguous against the registry, resolves what string distance cannot. 93.3% of mentions resolved; the 97 that remain are dominated by the Boulder Bar ghost and real out-of-town crags, both reported as findings rather than repaired.
Counting rules
Citations are per-capture deduplicated URLs; source buckets are keyword-and-domain based, with venue identity decided by registry website domains plus two verified alternate domains (jungleletnany.cz, stenatrinactka.cz). varp.cz — the crag database eight outdoor rows list as their “website” — is barred from counting as any venue’s identity and buckets as climbing web instead; that guard is why the density figure comes in two disclosed readings (business-only 84.6%, full-registry 66%). The leaderboard counts each venue once per capture; the citation score counts ChatGPT map presence plus citation-domain matches. Source-mix percentages are largest-remainder rounded per engine to sum to exactly 100. All headline stats ship as summary.csv.
Frequently Asked Questions
Related studies
The previous density record-holder — 78.4% supply, 45.5% share
The density curve’s middle point, and the bilingual canon
Where the density hypothesis was first framed — from the bottom end
The searchviewer wrap — at 100% last week, 74.5% this week