One line spread through marketing circles. Its source is a podcast with no numbers.
Profound, the US startup that shows brands how they appear inside AI answers, closed a $180M Series D in mid-September at a $1.8B valuation — under seven months after its previous round, nearly double. It reports more than a thousand enterprise customers, covering a third of the Fortune 100.
Travelling with the funding news was a line from founder James Cadwallader: think of the major models as different species. Gemini leans heavily on YouTube, because YouTube belongs to Google. ChatGPT mostly cites Reddit on consumer questions and shifts to LinkedIn on business-to-business ones. Claude used to answer more from pretraining, and has recently grown more sensitive to real-time information, going to the web more often.
The line is memorable and easy to repeat, and it spread through marketing discussion within a week. We went looking for the source: a Sequoia Capital podcast episode published on 17 September 2026. All four statements are in the transcript, and none of them is backed by a number or a third-party study anywhere in the episode.
That makes it one founder’s observation. An observation is worth hearing — but once it gets rewritten as news, it reads like something already verified. So we checked the four claims one at a time.
Four claims, checked one at a time
Claude shifting toward live retrieval — the direction holds, the magnitude doesn’t. Anthropic began opening up web search for Claude in March 2025, which is a verifiable first-hand fact, and the mechanism is sound. But there is no public quantitative study on how much more sensitive to real-time information it became. Use the direction; don’t quote a magnitude.
Gemini leaning on YouTube — small-sample support for the direction, a mess underneath the numbers. One independent study points the same way (Weglot, August 2026, 180 prompts per model). But we found four mutually contradictory figures for the same thing, ranging from 0.2% to 29.5%. The reason is not hard to find: plenty of reports fold “Gemini the chatbot” and “Google AI Overviews the search summary” into one number, and those are two different products. The direction is usable. None of the figures are.
ChatGPT citing Reddit on consumer queries and LinkedIn on B2B — we suggest dropping this one. Outside the founder himself, there is no second source. Awkwardly, Profound’s own published research (Reddit at 2.4%, LinkedIn at 0.39%) never splits consumer from B2B queries at all, so it cannot support the claim. And the one independent study that did test B2B scenarios (Ten Speed, September 2026) points the other way: 88% of citations in B2B decision contexts came from brand-owned content, with forums and video together accounting for 4.2%.
Four claims: one holds directionally, one is down to its direction alone, and two have stronger evidence against them than for them.
Even if all four held, they describe English-language markets
That is the part worth your attention here.
Japan has been measured. In the Japanese-market breakdown Ahrefs published in October 2025, ChatGPT cited Reddit 228,388 times, Google AI Mode cited YouTube 91,954 times, Copilot cited Japanese Wikipedia 125,842 times, and Perplexity cited Yahoo! Chiebukuro more than 518,000 times. Four engines, four completely different source lists.
Korea has been measured too. The AI Citation Index Report published in July 2026 ran 1,616 prompts across 30 industries and collected 4,848 answers containing 29,085 citations. Social content accounted for 28.2% of all citations and media for 17.4%; within the social share, Tistory blogs took 38.1%, YouTube 30.5%, and Naver blogs 15.9%.
The Japanese and Korean lists do not resemble each other, and neither resembles the Reddit-dominated English-market picture.
Profound itself ran a cross-language comparison: 3.25 billion citations, native-language prompts across 14 countries. Change the language and the engine changes the platforms it reaches for — 65% of Portuguese-language citations came from YouTube, 29% of Arabic-language ones from Instagram, and TikTok’s share in Spanish ran five times the English baseline.
The same study holds one detail that is easy to skim past: the overall social share of citations falls across every non-English market, and yet within that social share Reddit still takes 51% to 76% — in every country. Those are two different layers. “Social matters less here” and “Reddit still rules social” can both be true at once, and collapsing the two layers into one number gets you the opposite advice.
There is no Chinese in that list of 14 countries
For Traditional Chinese and the Taiwan market, we could not find a single comparable study with a stated sample size and method. Academic sources, industry media, SEO communities — all came up empty. The Taiwan-specific versions circulating out there (which local forum is the Reddit equivalent, which platform is better positioned) are all English-market conclusions carried over by analogy; at least one of them states in its own text that the data it cites comes from a global English-language study.
Worth noting in passing: the door is open. PTT has no robots.txt, Dcard blocks only one account-activation URL, Bahamut serves Allow: /, and Pixnet currently deploys no robots.txt — none of the four sets any rule for the major AI crawlers. (Mobile01’s robots.txt was blocked by its CDN from the network we checked from, so we could not retrieve it and have left it out.) But permitting crawling and actually being cited are two separate things, and nobody has measured the second one in Chinese-language markets.
There is one more piece out of Japan worth a look from Taiwan: across 71,041 AI answers, ChatGPT cited .co.jp domains more than any other TLD (44.8%), while Google AI Overview cited .com most (46.7%). If that pattern holds in Chinese, .tw is more than a suffix on your address. But we could not establish a firm publication date for that study, so it stands as corroboration and not as a basis for anything.
Even with that table, it changes every week
In the six-month follow-up to that Japanese tracking, note.com climbed and both Reddit and Yahoo! Chiebukuro dropped off the list — one reshuffle in half a year. That is the normal state of things, and it has been quantified.
The drift study SISTRIX published in May 2026 measured whether the sources cited this week are still cited next week: Google AI Overviews turned over 5%, Google AI Mode 56%, ChatGPT Search 74%.
Profound ran its own month-over-month version: AI Overviews 59.3%, ChatGPT 54.1%, Copilot 53.4%, Perplexity 40.5% (that one did not cover Gemini or Claude).
Lay the two together and something shows up that neither one reveals alone: AI Overviews drifts only 5% week to week, but 59.3% month to month. It isn’t twitching daily; it replaces its sources in batches, on a cycle. ChatGPT Search, meanwhile, turns over 74% in a single week.
Those two rhythms call for different work. Against the first, you have a window of weeks to get content in place before the next recomputation; against the second, what you need is to be present continuously rather than to bet on one moment. Most GEO discussion in Chinese still treats every engine as one thing.
There is a second, easier confusion here. People say AI answers are probabilistic — ask the same question ten times and the citations differ every time. That sounds like the same phenomenon as the numbers above, and it is not: everything above measures drift across weeks and months, which is what happens after an engine re-crawls and recomputes. Whether ten consecutive asks at the same moment diverge is something we searched hard for and found no study on.
The distinction decides whether your money is worth spending. Cross-period drift means the engine looked at the web again, so content you improve has a shot at being picked up on the next pass. Same-moment randomness would mean you are rolling dice regardless. One you can act on, one you cannot — and those are precisely the two things being discussed as if they were one.
How to tell whether a GEO study is usable
What this piece has been doing throughout is one thing: every time a number appears, ask where it came from. Here are the four tests we actually apply.
Check who benefits. We used Profound’s cross-language study because the sample is large and the method is disclosed — and because it comes from a vendor in this field, we say so when we cite it. Conclusions that flatter the publisher deserve harder checking than the ones that don’t.
Count the independent sources. The reason “B2B queries go to LinkedIn” is unusable is simple: one person has said it. And a second data point that looks like corroboration — Tinuiti measuring 73% growth in Reddit’s share of commercial-category citations — turns out to have been produced with Profound. That is not a second source; it is the same source wearing another face.
Check the sample and its edges. We cited the 88% figure from Ten Speed’s B2B study, but that study ran 220 prompts, covered only B2B SaaS and professional-services clients, and its author states plainly that it does not extend to consumer products. Fine for talking about B2B. Not fine for talking about your e-commerce store.
Check whether the numbers agree with each other. On the single question of how often Gemini cites YouTube, we found four mutually contradictory figures ranging from 0.2% to 29.5%. Follow them back and most come from sites that sell GEO tools or content-marketing services. Four answers a hundred-fold apart usually means nobody is measuring carefully — they are producing things shaped like research.
We are not listing these so you can screen studies yourself; your time belongs in your business. We list them because once you see how few published numbers survive all four, it becomes clear why having someone watch this for you beats reading ten articles a month.
Don’t guess the table. Shape your own pages to be quotable.
Don’t try to guess the Taiwan equivalence table and then put your budget behind one forum. Nobody holds that table right now, and guessing wrong costs you months.
Until someone actually measures citation structure in Chinese-language markets, the spend that won’t be wasted is shaping the content on your own site so it can be quoted. Whatever the engine, whatever the language, however the rankings reshuffle, all of them have to read your pages first. The one independent study that did test B2B scenarios points the same way: 88% from brand-owned content.
As for the road itself — what you have just read is the work we do. Tracing all four claims back to where they came from, setting them against the Japanese, Korean and cross-language studies, pulling the robots.txt of Taiwan’s forums one by one: there is no second piece like it in Chinese. Nobody has walked this road in the Chinese-language market yet. We are the ones out in front.
To see the GEO health of your site as it stands, the audit at geoweb.tw is free, scores in three minutes, and tells you which of the 12 dimensions is sitting at zero.
Further reading: Same question, completely different sources: how ChatGPT, Gemini and Perplexity choose, Being cited isn’t the value — who cites you is.