“For AI, Just Keep The Wikipedia Page Tidy”
Rebuild the PR media list and somebody in the room says it. Doesn’t that one line more or less settle the question?
I’ve done more than nod along. I printed the top ten “sites the AI cites most” out of an English-market study and pasted it straight into an internal deck; the first three rows were all Wikipedia properties, so it looked like there was nothing left to argue about.
But a tally that recounted the citations across 13 languages says that ranking doesn’t have to hold where you are.
An English-language ranking pasted onto another market’s list is another city’s route map. Every line on it is drawn properly, and every one of them runs through somewhere else.
Dmitrij Żatuchin’s preprint, out in June 2026, counts 167,551 citations across 12 markets, 13 languages and 128 brands (arXiv:2606.25787).
Where Does An AI Pull Its Sources From When It Writes About A Brand?
In 11 of those 12 languages, Wikipedia is the most-cited source. So far the English-market ranking holds up.
Then there’s Lithuanian. The top source there is vz.lt, a business title, and the ranking I had printed out stopped being right there — it never said which of the 12 would be the exception.
What’s being counted is grounding: the pages the assistant actually opened as evidence before writing, rather than the wording of the answer that came out. Which, for a media list, is the useful half — the pages a model opens are the ones worth asking whether you can get onto.
In Poland, A Job Board Outranked Polish Wikipedia
The bigger swap is in Poland, among its home-grown brands: there the most-cited source is YouTube. Video.
And four HR and careers portals combined drew twice as many citations as Polish-language Wikipedia. Careers pages, of all things, were being counted as something that builds a brand’s visibility.
How many companies have the careers site on the media list?
Mine didn’t.
The paper doesn’t measure why the lineup swaps. That each language market has a different kind of outlet doing the heavy lifting is a guess — a reasonable one, I think, and still a guess. So the cause column is still blank.
How Far Does This Carry To Your Own Market?
Not as far as the citation counts make it look. The 12 markets sit in Central and Eastern Europe, the Nordics and the Baltics. Neither Japan nor the Japanese language is in the count.
If your own market isn’t one of the 12 either, the lineup in this paper belongs to those markets, and the part that carries over to yours is the shape of the finding.
Since this is a preprint, nobody outside has been over the method yet. And one of the author’s affiliations, Rankfor.AI, sells AI-visibility ranking measurement — which doesn’t make the numbers wrong, but it belongs next to them.
The paper gives no measurement date, and the author lists the dependence on particular model versions and a particular collection window among the limits. The assistants are Perplexity, Gemini and GPT (three in all, and Claude isn’t one of them). Three datasets were merged to build the count, and the difference in granularity that came with merging sits on the author’s own list of limits too.
85.7% of the citations went to third-party domains — a second count landing where an earlier one already landed. The thread this piece is pulling runs elsewhere.
Maybe I’m leaning on the swap too hard. 11 of 12 languages landing on Wikipedia is agreement, not chaos, and someone could fairly read Lithuanian and the Polish job board as the fringe around a stable list. Still, that reading leaves the question I actually want answered untouched: which side my own market’s language falls on.
Add A Column For “Cited In This Language” To Your Media List
The English-market ranking isn’t the thing to throw away. What’s missing from it is a column.
Down the right-hand side of the list you already have, add one: is this outlet actually cited in the AI answers in your language? The rows that come back empty are the places you haven’t measured.
There’s no answer here for Japanese, or for any market outside the 12. Somebody did count the domains assistants cite for Japanese-language questions, and four of the top ten slots turned over in two months: Let’s Count Who The AI Quotes When The Question Is Japanese.
So make a second table, one for your own language, on the assumption that the names in it move. I’d keep the first one too — with 11 of 12 languages landing on Wikipedia, it still shows where most of the category agrees, and the new column shows which of those outlets your own language actually cites. From there on, it’s yours to measure.
Sources
- Żatuchin, D. (Estonian Entrepreneurship University of Applied Sciences / Rankfor.AI), “How Large Language Models Source Brand Reputation Across Languages and Markets”, arXiv:2606.25787, 2026-06-24, https://arxiv.org/abs/2606.25787
