AI Search for Karaoke in Bangkok (2026):the market with nothing to cite
TL;DR: Every study in this series so far has asked which sources an engine reaches for. Bangkok karaoke asks a blunter question, because 32 of 497 venues (6.4%) own a domain. 146 list a “website” and 78.1% of those list somebody else’s — a Facebook page, a LINE account, a Google Site, a bit.ly, one web designer’s demo host. Four engines, 24 questions, English and Thai, 384 of 384 captures, every cell filled. They answer confidently anyway: Karaoke Manekineko is named in 249 of 384 answers. But the citation-counted leaderboard this series normally publishes is arithmetically unavailable to 465 of the 497 venues, and the most-recommended brand in the city is linked to by exactly one engine.
Executive Summary
Prague was the experiment at the top of the scale: a vertical where almost every business runs a real website. Bangkok karaoke is what the bottom looks like, and the bottom turns out to be a different question rather than a smaller number.
A karaoke room in Bangkok does not need a website. It needs a LINE account, a Facebook page with tonight’s prices, and a pin on Google Maps. So that is what the market has: of 497 real venues, 146 filled in the website field on their Maps listing, and 114 of those 146 filled it with something they do not control. 84 rows point at Facebook. Three point at link shorteners. One points at the web agency that presumably pitched them.
That leaves 32 rows across 24 domains — 6.4% — as the entire citable surface of Bangkok karaoke. And it has a consequence this series has not had to confront before: the citation-counted leaderboard we normally publish alongside the mention counts is not a ranking of the market. It is a ranking of 6.4% of it. The other 465 venues are not losing that contest; they are not entered in it.
The engines, meanwhile, are perfectly willing to recommend them. The top eleven venues are named by all four. Karaoke Manekineko appears in 249 of 384 answers. What changes in a market with no websites is not whether you get recommended — it is what the recommendation is built out of, and whether anything links back to you when it happens.
1. What the website field actually holds
Every number in this study rests on one judgement, so it goes first. Google Maps does not verify the website field; it stores the string the business typed. Read it literally and Bangkok karaoke looks moderately webbed, because 29.4% of venues have something in there. Read what the strings actually point at and the picture inverts.
The single largest category is the empty one: 351 of 497 venues left the field blank. The second largest is Facebook, at 84 rows once you count the mobile and Thai-locale hosts separately listed. An own domain is the third thing a Bangkok karaoke bar is likely to put there.
I am counting a domain as a venue’s identity only when the venue controls it. That bars 28 hosts across 114 rows, and a few of those calls are arguable, so all 28 are published with a one-word reason in summary.csv along with the 24 kept. Overrule me with the file open.
Two I want to flag by name. Teeshot Bar lists a marriott.com restaurant page; it is corporate-owned and genuinely the venue’s page, but counting it would attribute every Marriott citation in the study to one karaoke bar. And บ้านเต๋า lists websitedesignth.com, which belongs to the agency rather than the bar.
One number this study will not give you
Five earlier cities in this series are plotted on a curve relating registry own-domain density to ChatGPT’s own-website citation share, and Bangkok at 6.4% would be a tempting new left-hand point. It is not plotted, because those five densities were computed under the looser rule and this one is not. Run the loose rule over this registry and Bangkok reads 45 domains across 63 rows — 12.7%, which is not a floor at all but level with Mexico City’s 12.5% and Seoul’s 13.0%. Same city, same day, same database: 6.4% or 12.7% depending entirely on whether a bit.ly counts as a website. That is the whole argument for the strict rule, and the whole reason this point is not on the curve until the other five are recomputed.
2. The engines agree anyway
A market with no websites might reasonably produce four engines wandering off in four directions. It does the opposite. The top eleven venues below are named by all 4 engines, and the top two are named in roughly half of every answer given.
Karaoke Manekineko is a Japanese chain with three Bangkok branches and 249 of 384 answers — 64.8%, third in this series behind Prague’s SmíchOff (82.1%) and Helsinki’s Löyly (79.8%), and comfortably ahead of Buenos Aires’ La Viruta (60.4%). R&B Karaoke follows at 190. Both are chains, both own a domain, and between them they take the top of every engine’s list.
Then the pattern breaks. Woodball is third on 119 captures and its web presence is a page on a free developer host. Club GEMS is fourth on a Google Site. Waterside is sixth with no website field at all. Being citable and being recommended turn out to be only loosely related.
3. A leaderboard that excludes 95% of the market
This series usually publishes two rankings: how often a venue is named in prose, and how often an engine cites that venue’s own domain. In Seoul and Mexico City the two disagreed about who was first, which was the finding. Here they barely overlap, and the reason is structural rather than interesting: you cannot be cited on a domain you do not have.
Only four of the top ten by mentions appear in the top ten by citations. The number one flips: Karaoke Manekineko leads on prose and R&B Karaoke leads on citations, 71 to 33. Two venues invert the usual shape — Tenpercent is third on citations with 19 and joint-fourteenth on mentions with 12, because when ChatGPT does name it, it links it.
And five of the top eight by mentions cannot appear on the citation board at all. Woodball, Club GEMS, Waterside, Karaoke City and DRAKE KTV have no domain between them. Their web presence is cited — Woodball’s developer-host page 39 times, Club GEMS’s Google Site 28, DRAKE’s entry on a lifestyle guide 24 — but none of that is a citation to the venue, so none of it counts, and counting it would make the measure meaningless.
What a visibility score would say about this market
465 of 497 venues (93.6%) score a structural zero on any citation-based metric, including the eleven venues that all four engines recommend. A tool that ranked Bangkok karaoke by citations would report that almost nothing in the city is visible, while four engines were busy recommending it. The metric would not be wrong about the citations. It would be wrong about what it was measuring.
4. One engine links the venue. The others just talk about it.
The clearest result in the study is also the simplest to state. Of 166 citations to a venue’s own domain across all four engines, Copilot accounts for 104.
| Engine | Own-domain citations | Share | Note |
|---|---|---|---|
| Copilot | 104 of 428 | 24.3% | The only engine that routinely links the venue itself |
| ChatGPT | 46 of 368 | 12.5% | Mostly R&B and Tenpercent; directories take 42% of its budget |
| Gemini | 6 of 99 | 6.1% | Cites almost nothing at all - 99 citations across 12 of 96 captures |
| Google AI Mode | 10 of 175 | 5.7% | Denominator excludes its 729 searchviewer wrappers |
Karaoke Manekineko is the sharpest illustration. Its own site is cited 33 times in this study, and all 33 are Copilot’s. AI Mode names the brand in 55 captures, ChatGPT in 72, Gemini in 61 — and all three link to it exactly zero times. The most-recommended karaoke brand in Bangkok is recommended, almost everywhere, without a link to itself.
Copilot is also the engine most willing to cite the substitutes. Its citations include 43 LINE deep links, 28 hits on Club GEMS’s Google Site and 24 on Woodball’s developer host. Where the others treat a non-domain as not worth linking, Copilot links whatever the venue actually has — which in this market is the more useful behaviour, even though it scores identically to zero on any own-domain metric.
5. Asking in Thai gets you a different bar
Every prompt was fired in English and in Thai, from a US address and a Thai one. Holding the engine and the vantage fixed and varying only the language gives a clean read on how much the question’s language decides the answer — and the engines disagree sharply about that.
| Engine | Vantage | Median EN↔TH top-5 overlap | Detail |
|---|---|---|---|
| ChatGPT | US | 3 of 5 | median over 23 comparable templates; no pair shares zero venues |
| ChatGPT | TH | 2 of 5 | 22 comparable templates; no zero-overlap pair either |
| Copilot | TH | 2 of 5 | 20 templates - but 6 of them share no venue at all |
| Copilot | US | 2 of 5 | 19 templates, 1 zero-overlap pair |
| Google AI Mode | TH | 1 of 5 | 19 templates, 1 zero-overlap pair |
| Google AI Mode | US | 1 of 5 | 19 templates, 6 zero-overlap pairs |
ChatGPT on the US vantage is the sturdiest cell — 23 comparable templates, a median of 3 of 5, and not a single template where the English and Thai answers share no venue. On AI Mode and Gemini the median is 1 of 5: switch language and four of the five places you are shown change.
The ordering is ChatGPT, then Copilot, then AI Mode and Gemini level at the bottom. Worth noting against the obvious guess: AI Mode on the US vantage and Copilot on the Thai one each have six templates where the two languages share nothing at all, so the disagreement is not a tidy function of where you ask from.
I would not read these as eight comparable measurements. Several “top-5” lists have fewer than five members, because some captures name only two venues — an overlap of 2 against a 2-item list is not the same signal as 2 against a 5-item list. The medians are the publishable part; the individual cells are in summary.csv with their list lengths.
.th citations, English prompts vs Thai ones, both vantages pooled — a 4.3× coupling on 368 citations..th share of any engine (15%) whatever language you ask in. AI Mode and Gemini have three and two Thai-domain citations in total, so neither gets a ratio here.6. Two closed bars, and the engines split on which one
The resolver deliberately failed to match two names, because the venues behind them are gone. Both are still being recommended, and the split between engines is oddly clean.
Four of the six are Thai-language prompts. Six distinct templates — it is not one question misfiring repeatedly.
The exact inverse: the only engine that recommends Major is the only engine that never recommends Yes!! R&B.
It is a tidy pattern and I want to be careful with it, because six captures and two captures are not a rate. What this shows is that stale recommendations survive in these systems and that different engines carry different stale records — not how often that happens, and not how long a closed venue lingers. There is no closure date in this data, so there is no lag to report.
A third string, Yes Ratchada Karaoke, turned up once on Copilot. It may be the same chain’s Ratchada branch or an unrelated bar; the database cannot settle it, so it is not folded into the six.
And what the other 8.8% is not
127 of 1438 mentions (8.8%) did not resolve to a registry venue, spread over 105 distinct strings of which 94 appear exactly once. That is a long tail, and it is mostly our seed’s fault rather than the engines’: these are real Bangkok bars missing from a 497-row pull. The Golden Giraffe Bar has its own page on a karaoke directory this study cites. A few are spelling variants our resolver should have caught — Maneki-Neko hyphenated failed to match the chain that the unhyphenated spelling matches 294 times. Calling this figure a hallucination rate would be wrong in both directions.
7. Where the answers come from instead
If 9.2% of citations go to venue domains, the other 90.8% goes somewhere. Here is the top of it, with what each domain actually is.
| Domain | Engines | Citations | What it is |
|---|---|---|---|
| google.com | 1/4 | 729 | Not a source - AI Mode searchviewer wrappers, 729 of 729 |
| timeout.com | 4/4 | 97 | Time Out Bangkok - the only guide all four engines lean on |
| bestofthailand.org | 2/4 | 77 | A programmatic "290 venues by area" listicle farm |
| karaoke.co.th | 4/4 | 71 | R&B Karaoke chain - a venue-owned domain, and the most-cited one |
| thairanked.com | 2/4 | 52 | Parallel EN and TH ranking pages |
| facebook.com | 2/4 | 45 | Venue Pages. ChatGPT and Gemini cite it zero times |
Time Out Bangkok is the only source all four engines lean on, at 97 citations. Below it sit two listicle farms — bestofthailand.org, which runs a programmatic “290 venues by area” page, and thairanked.com, which publishes parallel English and Thai ranking pages. Between them they take 129 citations, more than every venue domain in the city except R&B’s.
The genuinely local layer is social and messaging: Facebook Pages at 45, LINE deep links at 44, and Lemon8 — ByteDance’s Thai-heavy social app — at 30 from all four engines. This is the first study in the series where a messaging platform outranks every venue’s website but one.
Four of the top twelve are cited by exactly one engine, and three of those four are Copilot’s: the LINE links, Manekineko’s own site and Club GEMS’s Google Site. The fourth is google.com, which is AI Mode citing its own wrapper 729 times.
8. Instrument notes
Four things in this run measure the tooling rather than the engines. They are listed because each one is a figure a reader might otherwise take from the data, and each one would be wrong.
AI Mode cites a wrapper, not a publisher
729 of AI Mode’s 904 citations (80.6%) are google.com/searchviewer?svid=… links, which name no publisher — the change our searchviewer study dated to 2026-08-31, a month before this capture. AI Mode’s source mix here rests on 175 citations, not 904, and its own-website share is quoted against that denominator.
Copilot’s citation rows are incomplete
Its captures count 587 citations; the flattened table holds 428. 159 rows (27.1%) are missing, 134 of them from the Thailand vantage against 25 from the US. Shares of what landed are reported; volumes are not compared.
The web-search flag only works on one engine
web_search_triggered reads true on 95 of ChatGPT’s 96 captures and false on all 96 of every other engine — while those three returned 904, 428 and 99 citations between them. The field is simply not implemented for them. “Only ChatGPT searched the web” would be a statement about our pipeline.
A keyword rule was scoring directories as venues
Our own classifier treated any domain containing “karaoke” or “ktv” as a venue’s website, which quietly promoted a karaoke-ranking directory (7 citations), an international venue directory (2), a karaoke licensing company (1) and tucsonkaraoke.com — a blog in Tucson, Arizona, cited for Bangkok queries (2). Twelve citations, fixed before these numbers were computed. It is the reason the strict rule exists rather than a keyword.
If you run a karaoke bar in Bangkok
A domain is cheap and almost nobody has one
32 venues out of 497 own one. Both chains at the top of every engine’s list do. That is not proof a domain put them there — they are also the biggest brands in the city — but it is the only lever in this data that is entirely within your control and currently unused by 93.6% of the market.
Your LINE account is working harder than you think
44 citations in this study point at LINE deep links, essentially all from Copilot, and 45 point at Facebook Pages. If an engine is going to link something of yours, in this market it is usually that. Keep the hours and the room prices on it.
Being in the guide matters more than being on the web
Time Out Bangkok, bestofthailand.org and thairanked.com take 226 citations between them — more than every venue domain in the city combined. Those lists are editable by someone. Being on them is the closest thing to a distribution channel this vertical has.
Fix your Maps listing before anything else
351 venues have no website field at all, and three of the fourteen most-recommended places in the city are among them. The engines resolve you through Maps whether or not you have a site, so the pin, the name and the category are doing the work.
Conclusion
This series has spent twelve cities asking a question that assumed an answer: given that businesses have websites, which ones do the engines cite? Bangkok karaoke removes the assumption. 6.4% of the market has a domain, and the engines go on producing confident, consistent, largely identical recommendations — eleven venues named by all four of them, a clear favourite in two answers out of three.
What breaks is not the recommending. It is the measuring. The citation-counted leaderboard, which in Seoul and Mexico City was the interesting half of the story, here describes 6.4% of the venues and silently scores the rest at zero — including every one of the eleven the engines agree on. Run a citation-based visibility score over this city and it reports a ghost town.
The useful finding is the one engine that behaves differently. Copilot links what venues actually have: the LINE account, the Google Site, the page on a free developer host. The other three name those same venues and link nothing. If you are a bar with no website, three of four engines will recommend you and give the reader nowhere to go, and one will hand them your LINE account. Which is, if you think about what a karaoke bar needs from a search engine at eleven at night, probably the right answer.
Methodology
The grid
24 prompt templates (control, styles, formats, group sizes, personas, districts, price, plus nightlife and live-music entity-bleed controls) × 2 languages (English, Thai — translated competently but not native-reviewed) × 4 engines (ChatGPT, Gemini, Copilot, Google AI Mode) × 2 proxy countries (US, TH), fired 2026-09-23 through the pipeline’s own capture runner. All 384 captures landed: 96 per engine and exactly 24 in each of the sixteen engine × country × language cells. Nothing missing, nothing imputed, no engine excluded.
The registry
Broad Google Maps pulls on Thai and English karaoke terms, plus 36 name-targeted recovery searches for venues the answers surfaced, yielded 545 rows. An observed category map keeps 497 real — karaoke bars, KTV lounges, karaoke restaurants and one bowling-karaoke venue — and classifies out the live-music bars, nightclubs and restaurants that carry the keyword without having a room. One registry row is one venue location, so a five-branch chain contributes five rows to the 497.
What counts as a venue’s own website
The judgement this whole study turns on. Google Maps stores the string the business typed, unverified. A domain counts as a venue’s identity only if the venue controls it, which bars 28 hosts across 114 rows: social (84 rows), third-party directories and guides (14), messaging deep links (5), free developer and site hosts (7), link shorteners (3) and one hotel parent brand. All 28 barred hosts and all 24 kept domains are listed with a reason in summary.csv. The karaoke song-catalog sites (KaraFun, Smule, karaoke-version) appear in 0 of 497 registry rows — that layer shows up in citations, never in the registry.
Extraction and resolution
Venues named in each prose answer were extracted by the pipeline’s language-model NER pass, yielding 1,438 mention rows. Resolution precedence: exact normalized match, fuzzy similarity, unique token containment, then a hand-checked Thai-to-Latin alias table — the gap in this vertical is cross-script, and no mechanical fold bridges Thai script to a seed carrying Google’s English titles. 91.2% resolved. Every extracted name was mechanically verified to appear in its source answer.
Chains merge unevenly, and it is worth knowing why
Branches are merged into one brand when they share an identity domain, which is how Manekineko’s three branches and R&B’s five (including Echo House) become single rows. Venues with no domain have nothing to merge on, so Woodball’s branches stay separate. In a vertical this web-poor, “is this a chain?” is partly an artefact of whether anyone bought a domain — which matters when the subject of the study is exactly that. The two merged brands are labelled by hand in _mapdata.ts; the generator had labelled each with one branch’s name.
What this study does not measure
One capture date, so nothing over time. No fan-out: the field reads zero on all 384 captures for all four engines, which cannot be told apart from never having been populated. No closure dates for the two closed venues, so no staleness rate and no lag. No decoding of AI Mode’s wrapper IDs. And no check that the 24 kept domains are currently live — they are classified from the URL and the venue name, not fetched.