AI Search for Karaoke in Osaka (2026):the chains flip the law
TL;DR: I asked AI engines where to sing in Osaka 24 different ways, in English and Japanese, from a US address and a Japanese one. This series has spent six cities establishing that web supply predicts ChatGPT’s citations to venue-owned sites, and Osaka is the perfect natural experiment, because its registry density (37.7%) matches Buenos Aires’ to the decimal. BA earned a 25.3% share. Osaka earned 48.7% (the series record), and the difference is ownership: 24.4 of Osaka’s 37.7 density points are eleven karaoke-chain corporate domains, which collect 38.9% of ChatGPT’s citations on their own. 86.2% of all answers name at least one big-box chain, and Jankara is the consensus pick, named in 149 of 240 answers. Meanwhile the languages nearly stop talking to each other: EN and JA share 11% of their control top-5, the series’ lowest outside the two zero cities.
Executive Summary
Every vertical this series had measured so far was a crowd of small businesses. Karaoke in Osaka is an industry, and the industry’s structure, more than its web presence, decides what the engines cite.
Two weeks ago, Rome produced the first break in what six cities of data had made look like a law: own-domain density predicting ChatGPT’s own-website citation share, in strict rank order. Rome sat far below its predicted share because a professional gelato press absorbed the citations. Osaka breaks the ordering in the other direction. Its registry measures 37.7% own-domain density (the same figure Buenos Aires measured, to the decimal), and where BA’s tango scene collected 25.3% of ChatGPT’s citations, Osaka’s karaoke scene collects 48.7% (154 of 316), the highest share measured in any city.
The decomposition explains it. Of Osaka’s 37.7 density points, 24.4 belong to 114 branch listings of eleven karaoke chains: Jankara alone has 47 branches in the registry, Big Echo 34, Manekineko 16, all pointing at a handful of corporate domains. Those domains take 38.9% of ChatGPT’s citations, 1.6× their density. The 62 independents that own a domain (13.3 density points) take 9.8%, right where the series’ low-density cities sit. A chain domain is a single retrieval target with hundreds of branch pages, opening hours and price tables behind it; a snack bar’s one-page site competes with an Instagram profile. Concentration, it turns out, is a ranking factor.
By the numbers: 240 captures across three engines with data (ChatGPT and Gemini complete 96-capture grids, Copilot a disclosed 48-capture US-only half), carrying 683 cited URLs and 1,122 extracted venue mentions, 89% resolved against a 515-row registry. Google AI Mode and Perplexity were fired and delivered nothing usable; the methodology and the Dataset schema carry the batch-by-batch account.
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.
Same density as Buenos Aires, double the payout
Seven measured cities. BA and Osaka share one grey bar (37.7%) and sit 23 points apart on the purple one. Rome undershot its density because a guide press took the citations; Osaka overshoots because chains concentrate them.
The registry behind the grey bar: 467 real karaoke venues from a karaoke-specific Google Maps pull, of which 273 (58.5%) carry some form of web presence, 176 of them (37.7%) a site on the venue’s own domain. That second figure ties Buenos Aires’ 37.7% exactly: a coincidence, but a useful one, because it removes the usual excuse. When two cities put the same supply on the table and one gets nearly double the citations, the law as originally stated has stopped explaining things.
What changed is who owns the supply. BA’s 37.7% was hundreds of tango schools each running their own small site. Osaka’s 37.7% is mostly ten corporate domains wearing 114 branch listings. ChatGPT’s retrieval lands on jankara.ne.jp, big-echo.jp and karaokemanekineko.jp over and over (branch pages, price tables, opening hours), and each hit counts as a venue-owned citation, because it is one.
Chains: 114 registry rows on eleven corporate domains → 123 of ChatGPT’s 316 citations. Independents: 62 rows → 31 citations. The independent ratio (9.8% share on 13.3% density) sits where CDMX and Seoul sat.
86.2% of answers name a big-box chain
207 of the 240 captured answers recommend at least one of the national karaoke-box chains, and 788 of the 999 resolved venue mentions (78.9%) are chain brands. No earlier vertical in this series had anything like this layer; the closest was a single eight-branch gelato chain in Rome that managed to confuse the citation board all by itself.
Columns are largest-remainder rounded to sum to exactly 100. Copilot’s column is its US-proxy half-grid (48 captures); it could not be fired from a Japanese IP.
48 of ChatGPT’s citations go to Japan’s listing platforms: Hot Pepper, Gurunavi, Jalan, EPARK. This is the layer I expected to dominate, on the Tokyo precedent of a platform-heavy Japanese web. It runs a distant second to the chains’ own sites. When the business itself publishes structured pages, ChatGPT goes straight to the source.
Gemini barely touches the corporate layer (13.7% entity share). It spends 66 citations on Osaka editorial and tourism pages (osakatourism.org alone drew 48, its top domain), and another 20.9% on personal blogs, the habit Rome documented at scale. Two engines, one city, two completely different webs.
Not one citation in the study (any engine, any language, any proxy) went to reddit.com. Rome still had 15; Osaka has none. For a platform that carried 17–20% of ChatGPT’s citations in this series’ early cities, that is a remarkable disappearance, measured here for the first time at exactly zero everywhere.
Jankara vs the field
The Kansai-born discount chain is the consensus answer to karaoke in Osaka: named in 149 of 240 answers (62.1%), by every engine, in both languages; Gemini names it in 74% of its answers. The top five are all chains. The back half of the board is where it gets interesting: expat-facing karaoke bars with strong English-language footprints (Moonshine, ZINO, The Drunken Clam) outrank thousands of Japanese independents.
Which engine sings where
Each cell: how often that engine’s answers name the brand, with the raw count underneath. The chains own the top; the split below them is stark: Copilot, across its 48 US captures, favours the English-web bars (Moonshine 33%, Sakura 33%, Laughtale 27%), which Gemini, reading the Japanese editorial web, names almost never.
| # | Brand | ChatGPT96 captures | Gemini96 captures | Copilot48 captures (US only) |
|---|---|---|---|---|
| 1 | Jankara | 53.1%(51) | 74%(71) | 56.2%(27) |
| 2 | Karaoke Manekineko | 62.5%(60) | 42.7%(41) | 39.6%(19) |
| 3 | Big Echo | 49%(47) | 55.2%(53) | 35.4%(17) |
| 4 | Karaoke Pasela | 34.4%(33) | 26%(25) | 41.7%(20) |
| 5 | Karaoke Kan (カラオケ館) | 11.5%(11) | 26%(25) | 16.7%(8) |
| 6 | Moonshine Karaoke Bar | 8.3%(8) | 9.4%(9) | 33.3%(16) |
| 7 | Karaoke JOYSOUND | 18.8%(18) | 5.2%(5) | 6.2%(3) |
| 8 | ZINO | 14.6%(14) | 2.1%(2) | 16.7%(8) |
| 9 | The Drunken Clam | 11.5%(11) | 9.4%(9) | 6.2%(3) |
| 10 | Sakura namba karaoke | 5.2%(5) | 0(0) | 33.3%(16) |
| 11 | Karaoke Bar Laughtale | 6.2%(6) | 0(0) | 27.1%(13) |
| 12 | Onekara (ワンカラ) | 8.3%(8) | 9.4%(9) | 4.2%(2) |
ChatGPT and Gemini columns are full 96-capture grids and compare directly; Copilot’s 48-capture US-only column is shown for what it is.
Top 12 karaoke brands by answer mentions (click a marker for per-engine counts)
The marker carpet is the finding: rank 1 alone is 47 Jankara branches. A chain does not have a location; it has coverage.
The answers say Jankara; the citations say Manekineko
This series tracks two ways of being #1. Count which venue the answers name, and Jankara wins (149 answers, citation score 90). Count which venue’s web pages get cited, and Karaoke Manekineko wins with a score of 235 (nearly double anyone else) while sitting second on mentions (120).
Earlier cities split these boards between a beloved independent and a well-domained chain. Osaka’s split is chain-vs-chain, and the mechanism is page architecture. Manekineko’s site gives every branch its own indexed page under /locations/osaka/, and the engines cite those branch pages constantly: karaokemanekineko.jp drew 69 citations against jankara.ne.jp’s 65 with sixteen branches to Jankara’s forty-seven. Jankara is what the engines say; Manekineko is what they read.
For anyone optimizing a multi-location business for AI search, this is the clearest evidence yet in this series that per-branch pages on the corporate domain are what the retrieval step actually consumes.
Ask in Japanese and the independents change completely
The control prompt (“best karaoke in Osaka” vs 「大阪のおすすめのカラオケ」) shares 11% of its top-5 between languages (one venue in nine distinct), and the median across 18 comparable templates is 11%. Only Istanbul and Mexico City, at literal zero, measured lower. But pin the geography and the split closes: the Namba/Dotonbori district prompt returns an identical top-5 in both languages, this run’s version of the probe convergence BA and Rome measured.
| Prompt template | EN↔JA top-5 overlap |
|---|---|
| district: Namba/Dotonbori | 100% (5/5) |
| solo karaoke (hitokara) | 67% (4/6) |
| tourist near Dotonbori | 67% (4/6) |
| all-night karaoke | 43% (3/7) |
| nightlife (bleed control) | 25% (2/8) |
| English songs | 25% (2/8) |
| big group party | 25% (2/8) |
| control (best karaoke) | 11% (1/9) |
| district: Shinsaibashi | 11% (1/9) |
| family with kids | 11% (1/9) |
| first-timer how-to | 0% (0/10) |
| district: Umeda | 0% (0/10) |
| karaoke bars (sing to a crowd) | 0% (0/10) |
| daytime free-time rates | 0% (0/10) |
| karaoke snack bars | 0% (0/10) |
| most famous | 0% (0/10) |
| where locals go | 0% (0/10) |
| cheap but good | 0% (0/10) |
The overlap that remains is almost entirely the chains. Both languages recommend Jankara, Big Echo and Manekineko; what they disagree on is everything else. English answers route to the expat karaoke-bar circuit (Moonshine, The Drunken Clam, ZINO), venues with English websites and English reviews. Japanese answers surface snack bars and 昭和歌謡 joints (スナックバー美亜, うた吉, ザ・ベストテン大阪) that the English-language web has simply never written about. Eight of the eighteen comparable templates overlap at exactly zero, including “where do locals actually go” in both directions.
The language→TLD coupling collapses for the same reason. Tokyo bookstores measured the series high: Japanese prompts citing .jp domains 5× more than English ones. Osaka karaoke measures 1.05× (58 JA-prompt citations to .jp entity domains vs 55 EN), statistical parity, because the chains’ .jp corporate sites are what BOTH languages cite. When the dominant web presence is corporate, the usual language-to-domain geography disappears.
District note: the registry’s Osaka rows carry chōme-level neighborhood strings, so the sub-city district-accuracy test is partially a seed-granularity artifact here (Umeda venues carry Sonezaki/Doyamacho addresses that never contain the string “Umeda”), disclosed as in Amsterdam, Berlin and Rome rather than charted.
What the engines actually read
The most-cited domains across all 683 citations, with how many of the three delivering engines cite each. Five of the top six are corporate karaoke sites.
| Domain | Engines | Citations | What it is |
|---|---|---|---|
| jankara.ne.jp | 3/3 | 65 | Jankara’s corporate site: 47 Osaka branches behind one domain |
| karaokemanekineko.jp | 2/3 | 69 | Manekineko’s corporate site, with a page per branch (the citation board’s #1) |
| big-echo.jp | 3/3 | 54 | Big Echo’s corporate site |
| osakatourism.org | 3/3 | 48 | An Osaka tourism guide, Gemini’s single favourite domain this run |
| pasela.co.jp | 3/3 | 28 | Pasela Resorts: one Dotonbori branch, outsized citation presence |
| theclamjp.com | 3/3 | 13 | The Drunken Clam, an expat karaoke bar: the biggest independent venue domain |
| hotpepper.jp | 2/3 | 12 | Hot Pepper, the reservation platform half of Japan’s nightlife runs on |
| shop.joysound.com | 2/3 | 11 | JOYSOUND’s directly-run boxes: the machine-maker selling rooms |
| japannook.com | 1/3 | 12 | A Japan travel blog only Gemini reads |
| r.gnavi.co.jp | 2/3 | 9 | Gurunavi listing pages |
| prtimes.jp | 2/3 | 9 | Corporate press releases: chains announcing new branches |
| karaokekan.jp | 3/3 | 8 | Karaoke Kan, the chain Lost in Translation made famous |
| jkba.or.jp | 1/3 | 6 | The Japan Karaoke Box Association’s member list, cited as a directory |
Domains normalized (www stripped, subdomains kept). Engine count is out of the three engines with captures in the study.
467 venues, two nightlife poles
The registry maps the Osaka everyone knows: a northern cluster around Umeda/Kita and a southern one from Shinsaibashi through Namba to Tennoji, plus a long residential tail of neighborhood snack bars the engines almost never name in English.
All 467 mapped registry karaoke venues across Osaka
Instrument notes
Two engines fired, zero usable data
Google AI Mode was fired twice per proxy (four collection batches), and every one was still “running” on the provider’s side hours after the 40-minute polling windows closed. Its google.com self-citation metric, tracked since the start of this series, simply has no Osaka data point; a third attempt would have pushed the run past its spend cap. Perplexity landed 10 captures of 96 with zero citations attached, after 8/96 in Buenos Aires and 14/96 in Rome, its third consecutive collapse, and this time not even a caveat-laden count is publishable. Copilot ran its US half perfectly and had its JP half rejected at trigger (“no snapshot_id”), the same per-scraper proxy constraint Prague’s run hit on ChatGPT × CZ. All three are disclosed in the Dataset schema; nothing is imputed.
The ghost chain and the fossil chain
Engines recommend コート・ダジュール (Cote d’Azur), a real national karaoke chain, but a name-targeted Google Maps pass finds no live karaoke listing for it anywhere in Osaka, only an apartment building that happens to carry the name. Those mentions stay unresolved rather than being force-matched. One JA answer also recommends シダックス (Shidax), a chain that exited the karaoke business in 2018. An engine recommending it in 2026 is quoting training data, not the live web. Same class: カラワン (flagged as closed inside the very answer that recommends it) and the vinyl snack bar カンコドリ, which engines describe convincingly but no Maps listing supports.
Fifteen answers recommended nothing
15 captures name no venue at all, mostly the how-does-karaoke-work advice prompts answered with pricing mechanics, plus a handful of nightlife bleed-control answers that recommended dance clubs instead (correctly: the probe exists to measure whether karaoke leaks into general nightlife answers, and it mostly doesn’t). ChatGPT skipped web search in 8 of its 96 captures and answered from memory.
If you run a karaoke venue (or any multi-location business)
1. Branch pages on your own domain are the unit of AI visibility
Manekineko out-cites every rival with a third of Jankara’s branch count because each branch has its own indexed page with hours, prices and access. The engines cite those pages directly. An aggregator listing did not substitute for this anywhere in the data: Hot Pepper drew 12 citations against the chains’ combined 235.
2. Independents: your competition in AI answers is the chains, in every language
86.2% of answers include a chain; the residual slots are what independents fight over. The ones that win them (Moonshine, The Drunken Clam, ZINO, Sakura) all have a real website in the language of the asker. An Instagram profile (how 78 of this registry’s independents present themselves online) earned effectively nothing.
3. The two language markets barely overlap: treat them as separate channels
Outside the chains, EN and JA answers recommend different venues almost everywhere (11% median overlap). A venue courting tourists needs English-web presence specifically; ranking in the Japanese answers does not transfer.
Conclusion
Osaka was picked to stress-test the series’ oldest finding against a market structure it had never seen, and the test delivered: the density law, already broken downward by Rome’s guide press, breaks upward here. Two cities with identical 37.7% supply sit 23 points apart in ChatGPT’s citations because one supply is fragmented across hundreds of small sites and the other is concentrated into eleven corporate domains built like databases.
What survives all seven cities is a simpler statement than the law was: the engines cite whatever the best-structured description of the vertical happens to be. In Helsinki that was the venues’ own sites. In Rome it was a professional press. In Osaka it is the chains’ own infrastructure, and the 11% language overlap shows the corporate layer is also the only part of the answer the two language webs agree on.
Next stress point for the series: a city where the dominant local platform (a Tabelog-class directory) actually owns the vertical’s best pages: Japanese restaurants would be that test.
Methodology
The grid
24 prompt templates (control, formats, prices, personas, occasions, advice, canon, four districts, one nightlife bleed-probe) × 2 languages (EN/JA) × 2 proxy countries (US/JP) × 5 engines fired = 480 theoretical captures on 2026-09-30. 240 are in the study: ChatGPT 96/96, Gemini 96/96, Copilot 48/96 (US half only; the capture service rejects its JP proxy at trigger). Google AI Mode 0/96 (all four snapshots, two attempts per proxy, never became downloadable) and Perplexity 10/96 with zero citations are fired, disclosed and excluded. Percentages are computed only over complete grids, except Copilot’s, which always appears with its 48-capture denominator stated and carries no language claim.
The registry
515 Google Maps rows: a 500-row karaoke-specific Google Maps pull (EN locale) plus two name-targeted recovery passes for canon the cap missed (including Onekara, an entire solo-karaoke chain) and the venues the answers surfaced. 467 rows are real karaoke venues (boxes, karaoke bars, snack bars, lounges); izakaya, hostess bars, plain restaurants and one apartment building named after a karaoke chain are classified out but stay resolvable, so naming them is not counted as hallucination. Registry web layer, recomputed at publish time: 273 of 467 (58.5%) list any web presence; 176 (37.7%) an own-domain site, of which 114 rows sit on eleven chain corporate domains and 62 are independent. Social profiles, Hot Pepper pages and LINE links do not count as own domains.
Extraction and resolution
Venue mentions were extracted from each prose answer by a language model under the same rules as the tango and gelato runs (the series’ sandbox extractor), with every one of the 1,122 extracted names mechanically verified to appear verbatim in its source answer: zero failures. Resolution against the registry runs a precedence chain new to this study: a chain-pattern step first (registry titles are romaji, half the mentions are Japanese script; ジャンカラ 道頓堀店 resolves to Jankara’s canonical row, which is safe because the leaderboard aggregates brands by identity domain), then exact match, fuzzy match, width-folding, a hand-checked alias table and unique token containment: 89% of mention rows resolved. Ambiguous names (カラオケバー ラフテル matches two different registry bars) stay deliberately unresolved.
Counting rules
The leaderboard counts capture-level presence: an answer naming a brand once or five times counts once. Chain branches aggregate to one brand by identity domain. Citation classification runs on normalized domains with an identity guard: a venue listing its Instagram, Hot Pepper page or the karaoke industry association as its “website” does not make those domains venue-owned. About 8 citations to venue-owned domains of venues outside the registry (a small operator’s b-kara.jp, two bars’ own sites) sit in “other”, making the own-site share conservative. The full data contract (every number on this page recomputed from the published CSV) runs in this repo’s test suite.
Frequently Asked Questions
Related studies
The first density-law break: downward, to the guide press. Osaka is its mirror image.
The city whose 37.7% density Osaka matches to the decimal, with half the payout.
The first Japanese city, whose 5× language→TLD coupling Osaka’s chains erase.
The hotel study whose measurement methodology this series applies city by city.