AI Search for Temples in Taipei (2026):the government is the top source
TL;DR: Every city vertical in this series has run on some private publishing layer: OTAs for hotels, gelato guides in Rome, Time Out in Bangkok. Ask about temples and that layer is replaced by the state. 41.5% of ChatGPT’s citations are government domains, travel.taipei is the most-cited real publisher for all 3 engines, and Reddit, the series’ perennial top domain, appears 0 times in 694 citations. The engines agree on the canon (Lungshan Temple in 166 of 285 answers) while linking the temples themselves almost never: 32 citations study-wide reach a temple’s own domain.
Executive Summary
A temple is not a business, and the engines treat it accordingly. Where every commercial vertical in this series ran on guides, review platforms and booking sites, temple answers run on the tourism bureau.
Buenos Aires gave the first hint: the city government out-cited every tango venue’s own website combined. Taipei temples complete the shift. Government domains take 76 of ChatGPT’s 183 citations (41.5%), more than its next three source classes together, and travel.taipei, the city’s official travel portal, is the most-cited real publisher in the study at 67 citations across all three engines. Behind it: the Ministry of the Interior’s religious-culture portal, the national Tourism Administration, district offices. 118 of 694 citations (17%) are state domains of some kind.
The temples themselves barely figure as sources. 17.1% of the 504-temple registry controls a domain, and the engines cite one 32 times in the whole study (4.6%). The most any single temple earns is 9 citations. A citation-counted leaderboard exists here, but it ranks noise.
The recommendations, meanwhile, are the most uniform this series has measured: Lungshan Temple is named in 166 of 285 answers, at 57 to 60% on each engine separately, and every one of the twelve most-named temples is named by all three engines. The canon is settled; the engines merely recite it, and when they do, the link goes to the state.
1. The state replaces the guide layer
Each engine splits its citations differently, but the shape is new to the series. In Rome, 70% of ChatGPT’s citations went to gelato guides. In Bangkok, Time Out and two listicle farms carried the answers. Here the largest single bucket on ChatGPT is government, and no private guide comes close.
ChatGPT leans hardest: 76 government citations against 15 to temple sites and 0 to Reddit. Gemini spreads across travel blogs (37), OTAs and Wikipedia, with government at 8.6%. AI Mode’s measurable layer (the 175 citations that name a publisher) is 14.9% government, led by the same city portal.
The religion layer is its own small stack: the Ministry of the Interior’s taiwangods portal, temple-culture reference sites, festival explainers. Add it to the government bucket and the institutional share of ChatGPT’s citations passes half.
2. A settled canon, recited three times
All twelve of the most-named temples are named by all three engines, and the top of the list is remarkably stable across engine, language and vantage. Lungshan Temple appears in 60% of ChatGPT’s answers, 57% of Gemini’s and 57.6% of AI Mode’s. The engines did not need to agree this closely; in Bangkok and Prague they diverged well below the top five.
The top five map onto intent. Lungshan Temple (1738, Guanyin) is the default answer to almost everything. Dalongdong Baoan Temple carries the architecture and heritage prompts on its UNESCO award. Songshan Ciyou and Xia Hai City God own the night-market and matchmaking niches, and Xingtian Temple the business-fortune ones. Wealth prompts are the one place the canon cracks: several answers route to Earth God temples in New Taipei, outside the city registry entirely.
Reach is the narrowest yet measured: 62 of 504 registry temples (12.3%) are ever named, and 442 never appear in any of the 285 answers. Taipei has roughly one temple per city block; the engines work from a guidebook of a few dozen.
3. Thirty-two links to the temples themselves
Under the strict identity rule this series adopted in Bangkok, a domain counts as a temple’s own only when the temple controls it. 198 of the 504 real registry rows (39.3%) carry something in the Google Maps website field; 112 of those 198 (56.6%) carry a host the temple does not control, mostly social pages (87 rows, Facebook ahead of YouTube and Instagram), plus research GIS directories, worship platforms and one LINE link. That leaves 86 rows across 77 domains: 17.1% of the market.
The whole study produces 32 citations to temple-owned domains (4.6% of 694), and the best-cited temple site, Baoan’s, gets 9. ChatGPT’s own-site share is 8.2%, continuing the slide the series has tracked as registry density falls: Helsinki 45.5%, Buenos Aires 25.3%, Rome 6.0% on a 41.1% density. Taipei temples sit at 17.1% density and 8.2%.
No density-curve point is claimed: the five plotted cities were measured under the earlier, looser identity rule, and this study (like Bangkok) uses the strict one. Under the loose rule the registry would read 21.8% (110 rows). Both readings ship in summary.csv, with every barred host and its reason in website-field-rows.csv.
4. ChatGPT answers temples from memory
Two ChatGPT behaviours that defined the commercial verticals are absent here. Its map widget, which returned place carousels on 95 of 96 Buenos Aires captures and 1,319 map entities in Bangkok, fired on 0 of 100 temple captures. And its web_search_triggered flag, true on 95 of 96 Bangkok captures, reads true on 0 of 100 here, while the same captures carry 183 citations (see the instrument notes for what that flag can and cannot support).
The citations themselves point the same way. 35 of ChatGPT’s 183 citations are google.com links, and every one is a per-temple /maps/search pin attached to a named temple: pins, not search results. The model names the temple it already knows, then decorates it with a map link and a tourism-bureau page.
5. The canon is bilingual; the sources are not
Holding engine and vantage fixed and switching the prompt from English to Traditional Chinese keeps a median 50% of the top-5 (Jaccard) on every engine. The iconic-temples template returns effectively the same list in both languages on all three engines; the one systematic divergence is the wealth template, where Chinese answers go deeper into Earth God and Wealth God temples the English canon skips.
What the language does move is the sources. Citations to .tw domains rise on every engine when the prompt is Chinese:
| Engine | .tw share, EN prompts | .tw share, zh prompts | Coupling |
|---|---|---|---|
| ChatGPT | 23.1% | 35.2% | 1.5x; the highest .tw share in both languages |
| Gemini | 3.9% | 18.1% | 4.6x |
| Google AI Mode | 2.9% | 13.5% | 4.7x; share of all 326 citation slots incl. wrappers |
The ratios (1.5x to 4.7x) bracket the series baseline: Tokyo measured 5x on Japanese prompts, Rome 2.0x on Italian ones. The temple-specific part is the floor. ChatGPT already gives .tw domains 23.1% of its citations in English, because the government portals it leans on are .tw either way. When the state is your source, the usual English-web default weakens before the prompt language does anything at all.
6. The twelve most-cited domains
Counted across all three engines. The engines column says how many of the three cite the domain at all.
| Domain | Engines | Citations | What it is |
|---|---|---|---|
| google.com | 2/3 | 186 | Not a publisher: 151 AI Mode searchviewer wrappers + 35 ChatGPT Maps pins |
| travel.taipei | 3/3 | 67 | Taipei City's official travel portal, cited by all three engines |
| taipeitourism.org | 3/3 | 27 | A privately run Taipei tourism guide |
| zh.wikipedia.org | 3/3 | 23 | Chinese Wikipedia |
| en.wikipedia.org | 2/3 | 23 | English Wikipedia |
| taiwangods.moi.gov.tw | 2/3 | 15 | The Ministry of the Interior's own religious-culture portal |
| eng.taiwan.net.tw | 2/3 | 14 | The national Tourism Administration's English site |
| taipeitravelgeek.com | 2/3 | 12 | An independent Taipei travel blog |
| tripadvisor.com | 2/3 | 12 | The only classic OTA/review platform in the top ten |
| taiwanese-gods.org | 3/3 | 11 | A temple-culture reference site |
| fr.hotels.com | 2/3 | 10 | hotels.com's French locale, cited for temple queries |
| baoan.org.tw | 3/3 | 9 | Dalongdong Baoan Temple's own site, the most-cited temple domain |
Three of the eleven real publishers are state-run, one of them the Interior Ministry’s own religion portal, and a dedicated temple-culture reference site sits just behind them. The classic commercial layer is reduced to TripAdvisor at 12 and a hotels.com locale page at 10.
7. Instrument notes
Five caveats belong to the instruments, and a reader skimming the charts could mistake any of them for a result. Here they are, with what each one does and does not support.
AI Mode ran 85 of 100, recovered from late snapshots
Both AI Mode batches outlived the capture runner’s wait window on 2026-10-21, settled at the vendor one and two days later, and were inserted through the standard path on 2026-10-28. 15 of 100 prompts failed with browser-disconnect errors inside the vendor’s own run and are simply absent. All AI Mode figures use denominators of 85 captures or 326 citations, never 100.
Two engines failed at the vendor and are absent
Perplexity returned 0 records on probe batches on 2026-10-21 and again on 2026-10-28 (both crawl errors). Copilot’s dataset has been paused at the vendor since 2026-10-13, and its trigger now rejects outright. Neither engine appears in any figure; neither was imputed. This costs the series two of its recurring metrics for this city: Copilot’s entity share and Perplexity’s aggregator layer.
Half of AI Mode’s citations name no publisher
151 of its 326 citations (46.3%) are google.com/searchviewer?svid= wrappers naming no publisher, the instrument change our searchviewer study dated to 2026-08-31. Its source mix is measured on the 175 citations that name one. Notably the wrapper share here (46%) is far below Bangkok’s 80.6%: on temple answers AI Mode still carries sidebar links with real URLs.
The web-search flag cannot carry the no-search claim alone
web_search_triggered reads false on all 285 captures of all three engines, including ChatGPT, where it read true on 95 of 96 captures in the Bangkok run five weeks earlier on the same pipeline. Either ChatGPT genuinely did not search for temples, or the vendor’s flag changed between runs. Section 4 therefore rests on the converging evidence (zero map entities, Maps-pin citations, canon uniformity), with the flag as corroboration only.
The NER extractor was swapped, same rules
Temple mentions were extracted from the prose answers by a Claude extractor rather than the pipeline’s usual Gemini NER (whose key was unreachable from the run sandbox), following the same prompt contract, and every extracted name was mechanically verified to appear verbatim in its source answer: 1,324 mentions, zero verbatim failures. The same swap shipped in the Buenos Aires and Osaka runs.
What this changes, and for whom
For official tourism publishers
On heritage and religion queries, the engines treat a city’s own portal as the primary source. travel.taipei earns 67 citations here without any visible AI-optimisation effort, which suggests the institutional-authority signal is doing the work. A DMO that keeps its temple and festival pages accurate is, on this evidence, writing the answers AI gives about its city.
For anyone generalising from commercial verticals
The engine personalities this series tracks (Reddit-leaning ChatGPT, aggregator-leaning Perplexity, guide layers everywhere) are properties of commercial markets, and they do not survive contact with a vertical that has institutions instead of competitors. Measure a non-commercial vertical before assuming the playbook transfers.
For the 442 temples never named
The canon is old and the engines recite it, apparently without searching. A temple outside the canon is unlikely to enter it through its own website (32 own-site citations study-wide say that channel barely exists). The routes the data does show are the state’s own portals and the festival calendar: the answers that break from the canon are the ones about specific rituals, festivals and districts.
Conclusion
For the first time in this series, the question was about something that is not for sale, and the answer machinery reorganised itself around that fact. The guide layer gave way to the tourism bureau, Reddit vanished, the map widget stayed closed, and ChatGPT recited a canon it shows no sign of having retrieved.
The result is the most institutional answer set this series has measured: 41.5% of ChatGPT’s citations on government domains, a city portal as the study’s top publisher, and 32 links to the 504 subjects of the answers themselves.
For the series, Taipei temples mark out the far end of a spectrum the commercial cities only hinted at. Buenos Aires showed a city government out-citing the venues; here the state simply owns the sources. Whether that is deference to authority or an artefact of what the web contains for temples, the measurement is the same: when AI answers questions about worship, it cites the government.
Methodology
The grid
25 prompt templates (iconic, first-visit, districts, deities and styles, festivals, fortune sticks, etiquette, personas, photo spots, late opening, plus day-trip and Japanese-shrine bleed controls) × 2 languages (English and Traditional Chinese, translated competently but not native-reviewed) × 2 proxy countries (US, TW) × 3 engines: ChatGPT, Gemini, Google AI Mode. ChatGPT and Gemini: 100/100 each, 2026-10-21. AI Mode: 85/100, from two vendor batches that settled late and were recovered on 2026-10-28; the 15 missing prompts errored inside the vendor’s run. Perplexity (0 records on two probe batches) and Copilot (vendor-paused since 2026-10-13) are absent and appear in no figure. Nothing is imputed.
The registry
Broad Google Maps pulls on temple, shrine, monastery and 廟/寺/宮/觀 terms inside Taipei City yielded 607 rows after an address-derived 12-district whitelist (the raw city column is village-level there). An observed category map keeps 504 real places of worship and classifies out temple-named bus stops, parks, trails and festival POIs. Several answer-surfaced temples were added by name-targeted recovery searches, the same two-pass design as the Bangkok and Rome registries.
What counts as a temple’s own website
The Bangkok strict rule: a domain is a temple’s identity only when the temple controls it. 112 of the 198 with-website rows are barred, led by social pages (87), Academia Sinica’s research GIS directories (4), free hosts (5) and worship-platform directories (2); one row lists a government page and one a LINE link. Every row’s verdict and reason ships in website-field-rows.csv. The loose reading (21.8%) is published alongside, and no density-curve comparison is made across the two rules.
Extraction and resolution
Temple mentions were extracted from each prose answer by a Claude extractor under the pipeline’s NER contract (named places only, verbatim names, rank by order of first naming), with every name mechanically verified to appear in its source answer: 1,324 mentions, zero verbatim failures across three waves. Resolution precedence: duplicate- listing merge, exact normalised match, a hand-checked cross-script alias table (half the mentions are in Han script while the seed carries Google’s romanised titles for the major temples, and vice versa for the neighbourhood ones; no mechanical fold bridges that), then unique token containment. 89.3% resolved; 130 mentions are deliberately unresolved out-of-scope names from the two bleed probes (New Taipei day-trip temples, defunct Japanese-era shrines), and 12 single-mention in-scope strings remain as an honest seed-gap tail.
One temple, several listings
Google Maps lists big temple complexes more than once (main hall, side halls, branch palaces). Listings sharing an identity domain merge into one brand, which is how Xingtian Temple’s Zhongshan temple and its Beitou branch listings count as one row here. Three display names were set by hand after the generator labelled a brand with one listing’s title (Xingtian Temple, Zhinan Temple, and Nung Chan Monastery’s English name); the counts are untouched.
What this study does not measure
A single capture window, so no change over time. No Copilot or Perplexity, so no entity-engine or aggregator-layer comparison. No fan-out (the field reads zero everywhere, indistinguishable from unpopulated). No density-curve point, pending a recompute of the earlier cities under the strict rule. No liveness check on the 77 kept domains. And no claim about why ChatGPT’s search flag reads false, only the converging evidence around it.