Finding 01 — The same brand, the same week, four different verdicts.
Nine brands, all well known in their Taiwanese categories: Dcard, LINE Bank, Coupang Taiwan (酷澎), Anden Hud, Dent&Co (牙醫小幫手), FunNow, Rainbow Six Siege, Love and Deepspace (戀與深空), and iNewMe. Every question was asked the way a Taiwanese buyer would ask it, in Chinese.
| Across all nine brands | ChatGPT | Google AIO | Claude | Perplexity |
|---|---|---|---|---|
| Recommended the local brand | 35% | 70% | 48% | 67% |
| Ranked it first | 11% | 43% | 30% | 39% |
The spread is not a rounding error. Anden Hud, a Taiwanese underwear brand, was recommended by Google AI Overviews on 6 of 6 questions — described in its own words as「最知名、最多人買的網購國民內褲品牌」. On ChatGPT, same questions, same week: 0 of 6, with the slot going to brands the company does not compete with. Ask one engine and you are the category leader. Ask another and you do not exist.
Finding 02 — The engines are not reading the same amount. They are not close.
We counted the unique sources each engine disclosed per answer. The difference is an order of magnitude.
| Per answer | ChatGPT | Google AIO | Claude | Perplexity |
|---|---|---|---|---|
| Median sources cited | 0.5 | 4 | 10 | 14 |
| Answers citing nothing | 50% | 11% | 4% | 0% |
Half of ChatGPT's answers to these Chinese buyer questions arrived with no sources attached — written from what the model already knew. Perplexity never once answered without citing something. This was not driven by one unusual brand: zero-source answers appeared for eight of the nine brands we scanned.
Finding 03 — Retrieval is what decides whether the local brand shows up.
Split all 211 answers by whether the engine cited anything, and the gap is stark.
| Answer type | Named the local brand |
|---|---|
| Built from cited sources (n=176) | 61% |
| No sources cited (n=35) | 26% |
The same split holds inside individual engines, which is the part that matters. ChatGPT recommends the local brand in 44% of its sourced answers but only 26% of its unsourced ones. Google AI Overviews: 75% versus 33%. The engine is not changing its opinion of the brand — it is changing whether it goes and looks.
Finding 04 — What this changes about the work.
Two jobs, not one. Win the retrievals: be present in the sources these answers are actually built from — in our scans those were community and comparison sites far more often than official sites. Across the nine brands, the most-cited source in a category was a rival's own domain more than once; in the dental booking category, a competitor's site was cited 26 times, more than the client's own main site.
And be in the memory: for zero-source answers, the only thing that helps is being famous enough in the category that the model already knows you. That is slow, and it is the honest reason a brand new entrant cannot fix this in a quarter. Coupang Taiwan, entering the market recently, was recommended on 50% of checks overall — but on the two questions about reputation and reliability, AI handed the answer to the three domestic incumbents.
Method, and what this does not show
Nine brands across six categories, six buyer questions each, four engines, one pass per cell: 216 cells, 211 scored (5 lost to engine rate limits). Questions were generated from each brand's own site and market, then asked verbatim in Traditional Chinese. Each answer was scored by the same rubric: is the brand named, is it recommended, is it ranked first, and which sources were cited.
Four honest limits. One: we count the sources an engine discloses. An engine may retrieve without showing citations, so "no sources" means "no sources shown", and the comparison between engines is a comparison of disclosed behaviour. Two: Google AI Overviews was queried through a Taiwan-located search, which gives it a structural local advantage the chat engines do not have — some of its lead is a geography setting, not a preference. Three: these nine brands are not equally famous, and a single pass per cell cannot separate a genuine gap from the ~15% run-to-run drift we have measured in three previous studies. Four: Gemini was excluded — its grounded runs did not return within our time budget.
What we are confident in: the retrieval split holds across engines, across categories, and across eight of nine brands. What we are not claiming: that any engine is biased against Taiwan. The measured behaviour is about how often an engine looks things up, and the consequence falls hardest on brands whose fame is regional.
Is your brand in the memory, or only in the lookups?
Paste your site — one real buyer question across five engines, free, no email. The full scan tracks it weekly, engine by engine, with the sources behind every answer.