Channel sheet · CH-28 · gain 4 min · logged November 17, 2026
Content & SEO in the AI EraDirect input
English AI Tracking Can Hide Your Real Competitors, Study Finds
Stellarcast's 640-answer study found 11% brand overlap between German and English AI replies on Perplexity, versus 74% within German — English tracking may watch the wrong rivals.
By Nathan Brooks4 min read724 words
Signal notes
- Stellarcast's October 4 study covers 640 AI answers to ten coffee questions in Germany, France, Japan and Brazil.
- Perplexity German answers shared 74% of brands when repeated, but only 11% when compared with English answers.
- 85% of domains cited by Perplexity's German replies used .de, versus 5% for English replies; Google AI Mode showed 79% versus 0%.
- ChatGPT's German repeat overlap was 28% versus 21% across languages — no clear language effect on that engine.

Two German-language answers to the same coffee question on Perplexity shared 74% of the brands named. A German answer compared with an English answer shared only 11%. That gap is the central finding of an October 4 study by Stellarcast, and it carries a blunt warning for marketers: check your AI visibility in English, and you may be tracking the wrong competitors entirely.
The study asked ten coffee-shopping questions in English and local languages across Germany, France, Japan and Brazil, repeating each question. The publicly shared dataset contains 640 answers. Somya Goyal, who founded Stellarcast after quality-engineering work that included testing language-model outputs, led the research.
Her running example is a German coffee roaster. German-language replies named regional roasters, local retailers and store brands, she found, while English-language answers leaned toward American names. A roaster selling primarily to German buyers could therefore watch the wrong rivals and pitch sites that barely appear in the answers its customers actually receive.
Why the baseline matters more than the headline
The 74% repeat figure is as important as the 11% language gap. AI answers vary when someone asks the same question twice, so a language comparison only convinces when its difference clearly exceeds that normal drift. Goyal treats the repeat result as the baseline for judging everything else.
That caution is not academic. On ChatGPT in Germany, repeated answers in the same language shared just 28% of brands, while German and English answers shared 21%. That narrow difference leaves little basis for claiming a strong language effect on that engine from this sample. The gap, in other words, differed by engine, and the study says so.
For monthly reporting, Goyal advises keeping prompts, location and engines consistent, repeating each question, and reporting a range rather than a single number. She suggests comparing new readings against the middle 50% of earlier results instead of declaring victory from one favorable answer. A new brand appearing in one response may reflect random model movement, not a campaign gaining ground.
What do the cited sources show?
The study also tracked the domains cited alongside recommendations. The numbers are sharp:
- Perplexity, German-language replies in Germany: 85% of cited sites used a .de domain, versus 5% for English replies.
- Google AI Mode in Germany: 79% versus 0%.
Those are shares of cited domains within the observed answers — not proof of consumer behavior, and not evidence that a particular placement will lift sales. Goyal said so plainly.
The operational takeaway still stands. If local retailers or review sites repeatedly appear in answers to customer questions, improving accurate product information and genuine third-party coverage there may beat publishing another generic page on your own site. She adds a condition: change that priority when the cited-domain evidence changes.
A test a small team can run monthly
For a business operating in two countries, Goyal proposes a routine that fits a monthly reporting cycle:
- Write 15 to 20 realistic buyer questions in each market's language.
- Test them on the two or three AI engines your audience most likely uses, with location set correctly.
- Repeat each test three times; log both brands named and sources cited.
- Compare against earlier same-language runs before judging whether visibility moved.
She cautions against judging a campaign from a screenshot of one AI answer. A single response cannot establish a durable ranking, a competitor's rise, or a campaign's effect. That discipline also reshapes agency briefs: instead of asking for more mentions everywhere, a brand can identify which sources appear in the answers its actual buyers receive and investigate why.
Where the findings stop
The limits deserve equal billing. Stellarcast tested coffee-buying questions, not a broad sample of categories. Japan was the only APAC market. Goyal expects local-language testing to matter elsewhere, but the size of the effect remains untested for most markets and products.
One more caveat the study itself acknowledges: the publisher sells AI-visibility monitoring services. The raw answers and methodology are publicly available, which makes the results inspectable — but an inspectable test by one vendor is not an industry benchmark.
Goyal's interest began in quality engineering, testing what language models produce. The shifting answers, she said, keep handing her another question to investigate. Marketers reading her study would do well to borrow that habit.
via contentgrip.com (Source)
More from Nathan Brooks
Show full bio
Correspondent covering regulation, policy and their impact on marketing professionals running AI in the stack at Mart Signal.
155 articles