Most AEO programmes run their visibility checks in English. A brand tests a set of prompts, records which AI surfaces cite them, and reports back to the business. It feels rigorous. It probably isn't.
Research published by Search Engine Journal, drawing on 1,650 AI responses, found that the brand an AI recommends changes depending on the language in which the query is asked. Same intent, same product category, different recommendation. That is not a minor quirk. For any brand operating across more than one language market, it means their AI visibility data may be systematically incomplete.
Why AI Responses Differ by Language
AI answers are not retrieved from a single universal index. Tools like ChatGPT, Perplexity, and Gemini are trained on data that skews towards English-language sources. When a query arrives in French, German, Spanish, or Japanese, the model draws on a different mix of training data and retrieval signals. The sources that dominate in English - high-authority English-language publications, English Wikipedia, English-first review platforms - carry less weight when the query context shifts.
This means a brand that has invested heavily in English-language digital PR, structured content, and entity signals may be well-cited in English responses but effectively invisible when the same query is posed in another language. The AI is not being inconsistent. It is doing exactly what it is designed to do: drawing on contextually relevant sources. If your brand's footprint in those languages is thin, the AI finds someone else.
Google AI Overviews compounds this further, because the index underlying any given response is itself influenced by the language and locale of the search. A user in France searching in French is drawing on a different slice of crawled, indexed content than a user in the UK searching in English. The recommendation layer is built on top of that.
What This Means for How You Measure Visibility
If you are running an AEO programme and your prompt testing is English-only, you have a measurement gap. You are not measuring your visibility. You are measuring your English visibility. For a UK-only brand with no international ambitions, that might be fine. For anyone else, it is not.
The fix is straightforward in principle, though it requires more resource. Prompt testing should be replicated in each priority language. The prompts cannot simply be machine-translated versions of the English set - they need to reflect how users in that language actually phrase their queries. A Spanish-speaking user in Spain and an English-speaking user in the UK may share the same purchase intent but ask very differently shaped questions.
Tracking AI-referred traffic in GA4 and your CRM also needs to be segmented by language or locale. If you are seeing AI referral sessions from Perplexity or ChatGPT but not attributing them by market, you cannot tell whether your AEO work is performing in the markets where it matters most to the business.
The Citation Infrastructure Problem
AI recommendations are built, in part, on citation signals - the sources an AI model treats as authoritative when constructing an answer. In English, a brand might have strong coverage across respected publications, industry directories, and review platforms. In other languages, that same brand might have almost no equivalent coverage. The AI fills the gap with whoever does.
This is a digital PR and entity signal problem as much as a content problem. Localised digital PR - coverage in respected local publications, mentions in language-specific directories, reviews on platforms that carry weight in that market - feeds the citation layer that AI models draw on. Schema markup in the local language, hreflang implementation, and localised entity associations in Google's Knowledge Graph all contribute. None of this happens automatically when you translate your website.
Brands often assume that because they have a localised website, they have localised authority. Those are different things. A translated product page with no inbound citations in that language does not make you citable. It makes you present. There is a significant difference when it comes to what AI chooses to recommend.
Where to Start If You Operate in Multiple Languages
The first step is an honest audit of where your AI visibility actually sits. Run your core prompts - the questions your target customers ask when they are evaluating your category - in each priority language. Do this across ChatGPT, Perplexity, and Gemini at minimum. Record not just whether you appear, but who appears instead of you. That tells you which competitors have done the work in those languages.
From there, prioritise by commercial impact. If German-language AI responses are recommending a competitor in a market that represents 30% of your revenue target, that is a materially different problem to a gap in a small secondary market. The AEO workstream for each language should be sized accordingly.
On the content side, structured content written for AI extraction - clear answer formats, FAQ sections, well-labelled product and service information - needs to exist in the local language, not just in English with a translation layer. On the citation side, localised digital PR outreach to publications that AI models demonstrably draw on in that language is the highest-leverage activity. This is slower work, but it is the work that actually moves the dial.
The Attribution Question You Cannot Ignore
There is a downstream measurement problem too. If your CRM and analytics setup does not capture the language or locale of AI-referred sessions, you cannot attribute revenue to the right market. You may be successfully visible in, say, French-language Perplexity responses, and generating leads from French-speaking users, but reporting that as undifferentiated AI referral traffic.
Segmenting AI referral traffic by language in GA4 is achievable if your site is correctly structured with locale-specific URLs or subdomains. Connecting that through to the CRM requires that the session source, medium, and landing page locale are passed into the lead record. This is standard practice for international paid search. It should be standard for AI search attribution too.
The research finding is simple but its implications are significant: AI does not produce a universal answer. It produces a contextually shaped one. If your brand's context is thin outside English, your AI visibility is thin too - you just may not have measured it yet.