Stellarcast's Somya Goyal: English AI tracking can hide your real competitors

Somya Goyal of Stellarcast explains why English-only AI visibility checks can point marketers at the wrong competitors, and how to test local-language answers instead.

Illustrated portrait of Stellarcast founder Somya Goyal beside English and German AI answer cards

When a German coffee roaster checks its AI visibility in English, it can end up tracking American brands instead of the local businesses competing for its customers. That is the practical warning from Somya Goyal of Stellarcast, who compared AI-generated coffee recommendations across languages and markets.

The company's October 4 study asked ten coffee-shopping questions in English and local languages in Germany, France, Japan and Brazil. Each question was repeated, and the publicly shared dataset contains 640 answers. The German example shows why a marketer's reporting language can change the competitive picture more than expected, although the research covers a single product category and a limited set of prompts.

Table of contents

When the language changes, so does the shortlist

On Perplexity in Germany, two German-language answers to the same coffee question shared 74% of the brands named. Comparing a German answer with an English answer produced only 11% overlap, according to the Stellarcast study.

For a roaster selling primarily to German buyers, the consequences are tangible. Somya said German-language replies named regional roasters, local retailers and store brands, while English-language answers leaned toward American names.

A marketing team using English prompts could therefore pay attention to the wrong rivals and spend time pitching sites that barely feature in German responses. Somya recommends starting with the buying questions customers would actually type, using the local market and language, then recording both the brands mentioned and the sites cited.

First measure how much the answers drift

The 74% figure is as important as the 11%. AI answers already vary when someone asks the same question twice. A language comparison only becomes convincing when its difference is substantially larger than that normal variation.

Somya calls the repeat result a baseline against which to judge the language gap. For Germany on ChatGPT, the study recorded just 28% brand overlap between repeated answers in the same language, compared with 21% across German and English. That smaller gap leaves little basis for claiming a strong language effect on that engine from this sample.

For monthly reporting, she advises keeping the prompts, location and engines consistent, repeating each question and reporting a range of results. She suggested comparing new readings with the middle 50% of earlier results rather than declaring victory from a single change.

The underlying lesson is about how carefully marketers interpret dashboards. A new brand in one answer might reflect random movement in the model rather than a campaign gaining ground, which is why teams rethinking what success looks like in AI search need a baseline before they need a trend line.

Follow the sources that customers actually see

The study also examined the domains cited alongside recommendations. Somya reported that 85% of sites cited by Perplexity's German-language replies in Germany used the .de domain, versus 5% of those cited in English. For Google AI Mode in Germany, the comparison was 79% against 0%.

Those figures are shares of cited domains within the observed answers, not proof of consumer behavior or evidence that any particular website placement will increase sales.

A team could still put them to work. If locally relevant retailers or review sites repeatedly appear in the answers to customer questions, it may make more sense to improve accurate product information and genuine third-party coverage there than to produce another generic page on its own website. Somya says brands should change that priority when the cited-domain evidence changes.

A repeatable test for a small marketing team

For a business operating in two countries, Somya proposes a routine that can fit into a monthly reporting cycle:

  1. Write 15 to 20 realistic buyer questions in the language of each market.
  2. Test each question on the two or three AI engines the audience is most likely to use, with the location set appropriately.
  3. Repeat each test three times and log the brands named and the sources cited.
  4. Compare results with earlier same-language runs before judging whether visibility improved or declined.

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 changes the brief sent to an agency or PR team. Instead of asking for more mentions everywhere, a brand can identify which sources appear in the answers its actual buyers are receiving and investigate why.

Where the findings stop

Stellarcast tested coffee-buying questions, rather than a broad sample of product categories. Japan was the only APAC market in the study. Somya expects local-language testing to matter elsewhere, but the size of that effect is still untested for most markets and categories.

The research was published by the company that sells AI-visibility monitoring services. Its available raw answers and methodology make the reported results inspectable, but they do not turn one company's limited test into an industry-wide benchmark. The gap also differed by engine: ChatGPT's German-language comparison did not stand out from its repeat variation.

Somya's interest in how AI changes its brand recommendations began with her quality-engineering work, including testing language-model outputs. That curiosity eventually grew into Stellarcast, and she says the shifting answers keep giving her another question to investigate.

This article is produced by ContentGrow. We're building branded media outlets for B2B companies. Interested in learning more? Learn more.