Translated Content and AI Citations: The Evidence

01Three independent lines of evidence

Most writing about multilingual AI visibility is speculative. This topic is unusual in that three separate kinds of evidence exist, from three unrelated sources, and they agree.

Each has weaknesses, and we name them as we go. Together they make a reasonably strong case that language coverage is a bigger lever in AI search than it was in conventional search, which is a claim worth testing rather than assuming.

02The 1.3 million citation study, and its limits

The largest dataset available on this question was published by Weglot, which sells translation software. That is a real conflict and we flag it up front. It is also unusually transparent about method, which is more than most vendor research offers.

The design: two phases focused on Spanish-language markets. Phase one analysed 153 high-traffic sites with no English translation, 98 from Spain and 55 from Mexico. Phase two added a comparison group of 83 sites that had both Spanish and English versions. The top 50 non-branded keywords per site were converted into natural-language queries and translated between languages, generating 22,854 queries in phase one and 12,138 in phase two.

The shape of that table is the finding. Untranslated sites are close to invisible for queries in a language they do not serve. Translated sites perform nearly as well in the second language as the first, which the study summarises as a 327% visibility increase for the previously unserved language.

A secondary result is worth noting because it is less obvious: translated sites received 24% more total citations per prompt overall, with a 33% increase in English and a 16% increase in Spanish. The Spanish increase is the interesting one, since translating into English should not directly help Spanish queries. The study’s interpretation is that translation signals authority across all languages. That is a hypothesis rather than a demonstrated mechanism, and we would treat it as the least reliable claim in the set.

That is the study author’s proposed mechanism, and it is worth reading as a hypothesis rather than a finding. It is plausible and it is also exactly what a translation vendor would conclude, so the honest position is that the correlation is well evidenced and the causal story is not.

03Why the engines differ, and what it means for you

The most decision-relevant finding in the study is not the headline number. It is that the effect is almost entirely a Google AI Overviews phenomenon.

The likely explanation is structural rather than mysterious. AI Overviews draw heavily on conventional search ranking, and conventional search is strongly partitioned by language. ChatGPT works more from model knowledge that crosses languages, so a concept it learned from Spanish content remains available when answering in English.

The practical consequence is a targeting decision. If your buyers reach you through Google surfaces, language coverage is a substantial lever. If your buyers are predominantly in assistants like ChatGPT, translation is a much weaker play and your effort is better spent on the citation-source work we set out in the citation share benchmarks.

04Where AI crawlers actually spend their budget

The second line of evidence is independent of the first and comes from server logs rather than citation tracking.

A 48-day log analysis reported that AI crawlers spend the majority of their crawl budget on language variants: roughly 80% for Meta-WebIndexer, 62% for GPTBot and 60% for Bingbot. If that generalises, translated pages are receiving most of the AI crawler attention on multilingual sites.

That is a striking allocation, and it is corroborating rather than duplicative evidence, because crawling and citing are different events measured by different means. A crawler heavily fetching translated pages is consistent with those pages mattering to retrieval, without depending on any citation study being right.

Two caveats. It is one site over seven weeks, so treat the exact percentages as indicative. And heavy crawling of a page category does not by itself prove those pages get cited more, only that systems are investing effort in reading them. The full log analysis is in what server logs reveal about AI crawlers.

05Grounding queries arrive localised, which we have not seen documented

The third line of evidence is our own, and it points at something the other two do not capture: language coverage matters for machine retrieval, not only for human search.

In our own Search Console and Bing Webmaster data we observe long, templated multi-brand comparison prompts that no human would type, carrying trailing instruction text of the kind found in a prompt rather than a search box. Those are AI systems grounding an answer by running a search.

The relevant part for this article: the same template appears in localised variants. We see Spanish versions of the identical English prompt, including a Spanish rendering of the same trailing instruction, alongside Spanish equivalents of the comparison phrasing and of words like reviews and ranking.

Two honest limits. This is observation on one property, not a controlled study, and we cannot see which system generates those queries or what it does with the results. What we can say is that the localisation is real, systematic and template-identical, which is hard to explain other than as the same prompt set being run in multiple languages. We described the wider phenomenon, and the reporting trap it creates, in the ChatGPT tracker analysis.

06The translation proxy problem

There is a second cost to being untranslated that is easy to miss because it looks like visibility.

When an untranslated site does surface for a foreign-language query, the link frequently points at a machine-translated proxy rather than at the site itself. The study cites Ahrefs data from June 2025 indicating Google’s translation proxy pulled 377 million monthly visits from untranslated websites, with the traffic staying on Google rather than reaching the source.

Their worked example makes the cost concrete. A Spanish book retailer that stocks English titles and ships worldwide appears 64% less often in English queries, surfacing 36 times for every 100 Spanish appearances, and in those 36 cases the link goes to the translation proxy so the retailer does not capture the visit.

So the choice is not visible versus invisible. It is capturing the visit versus having someone else capture it while you supply the content, which is a familiar shape in AI search generally.

There is a measurement consequence too. If a translation proxy is intercepting those visits, your analytics will show foreign-language demand as close to zero, which makes the market look uninteresting and discourages the very investment that would capture it. Anyone deciding against translation because there is no existing traffic from a market should check whether that traffic is being intercepted rather than genuinely absent.

07Community content, translation and citations

One angle deserves separate treatment because community sources are the most cited category in AI answers, and because most brands have no control over them.

Research into citation sources places community platforms first across major AI engines. Those platforms are overwhelmingly English for most technical and B2B categories, and they are not translated in the way a corporate site is, which creates an asymmetry: your brand may be discussed extensively in English community threads and barely at all in another language.

The practical implications are modest but real. If you are pursuing visibility in a non-English market, the community layer that helps you in English may simply not exist there, which shifts the balance toward your own translated pages and toward local review platforms and forums. It also means auditing your citation sources per language rather than once, because the source mix genuinely differs.

A cheap version of that audit: run five of your most important category prompts in the target language and record which domains are cited, then compare that list against the English one. The overlap is usually smaller than teams expect, and the non-overlapping half is the actual work.

We would caution against the tactic this observation tempts people toward. Translating and reposting community discussion you did not write, in order to create citable material, produces low-quality duplicates and is the kind of scaled publishing that carries real risk, which we set out in the scaled content analysis.

08Partial translation is the documented failure mode

If you translate, the evidence is specific about how it goes wrong, and it is not translation quality.

The failure is incompleteness. Partial translations create coverage gaps where AI systems do not see your content for those queries at all. The illustrative case is a site with translated homepage and product pages but an untranslated blog, which surfaces in some AI responses and not others depending on which part of the site would have answered.

That matters because partial translation is the default outcome of most translation projects. Marketing pages get translated because they are visible to executives; the documentation, the comparison pages and the long-tail blog do not, and those are frequently the pages that would have been cited.

09A note on translated community pages specifically

One pattern deserves flagging because it is becoming common and is frequently misread as an opportunity.

Community threads get scraped, machine-translated and republished on aggregator sites, which then sometimes appear as citations in non-English AI answers. If you search your own brand in another language you may find your product discussed in a translated copy of a thread you have already read in English.

Two things follow. The first is defensive: those copies are frequently poor translations of already-imperfect summaries, and errors compound. A complaint that was specific and fair in the original can become a flatly wrong factual claim about your product two hops later, and that version may be what an engine cites in that market. Worth checking annually if you sell internationally.

The second is that this is not a tactic to copy. Republishing translated community content you did not write is scraped-content republishing with a translation step, and it carries both the quality-signal risk described in our scaled content analysis and a straightforward attribution problem. The legitimate version of this is participating in the communities that already exist in that language, which is slower and does not scale, and is the only approach that survives scrutiny.

If you find damaging mistranslations of genuine feedback, the practical response is the same as for any inaccurate third-party source: publish a clear, current, well-structured answer in that language on a property you control, so there is something correct for an engine to prefer.

10Whether any of this applies to you

Translation is expensive to do properly and permanently expensive to maintain, so the honest question is whether your situation matches the one the evidence describes.

Run the cheap test first. Look in your own Search Console for queries in languages you do not publish in, and check your analytics for sessions from non-English locales. Most teams have never looked, and the answer is usually decisive in one direction or the other. It costs ten minutes.

Weight by engine. The effect is concentrated in Google surfaces. If your category’s buyers are predominantly in assistants where the measured language bias is a few percent, translation is a weak lever regardless of how impressive the headline figure is.

Be honest about maintenance. The documented failure mode is partial translation, and partial is what happens when a translation project ships and then the English site keeps moving. If you cannot commit to keeping versions in sync, you are buying the failure case.

Sequence it after the basics. If AI crawlers cannot reach your pages, or your key content is only linked through JavaScript, translation multiplies a number that is currently zero. Crawlability first, which we cover in the server logs piece.

11What this evidence does not show

Stating the limits plainly, because the headline number is the kind that gets repeated without them.

The main study is vendor research. Weglot sells translation software and the finding supports its product. The methodology is published and the sample is large, which is better than most vendor studies offer, and the conflict remains.

It is one language pair and two markets. Spanish and English across Spain and Mexico. Whether the same gaps appear for languages with different script systems, or in markets where English proficiency is very high, is untested here.

Correlation, not mechanism. Sites that translate are plausibly different from sites that do not in ways that also affect citations: bigger budgets, more mature content operations, international ambition. The study does not control for that, and the claim that translation improves citations in the original language is where we would be most cautious.

Our own observation is uncontrolled. The localised grounding queries are real and systematic in our data, and we cannot attribute them to a specific system or measure what they produce. We report it because we have not seen it documented elsewhere, not because we have quantified its effect.

What we would treat as durable across all three lines: AI Overviews strongly prefer content in the query language, AI crawlers invest heavily in language variants, and machine grounding runs across languages. What we would re-test before acting on: any specific percentage above.

One last note on how to use research like this generally. A single vendor study with a striking percentage is the most quoted and least reliable artefact in this whole field. What made this one worth writing about is not the 327% figure, it is that two unrelated datasets collected by different means point the same direction. When you next encounter a headline number in AI search, the useful question is not whether the study looks rigorous, it is whether anything else independently agrees with it.

12Frequently asked questions