Which AI Recommended That Brand? One Answer Is Enough to Tell

Businesses increasingly pay for 'AI visibility' reports — studies of what ChatGPT, Gemini, Grok and others say when someone asks for a gift idea or a company recommendation. The idea is that if an AI mentions your brand often, you're winning. A researcher at arXiv tested whether those reports actually describe the AI, or just the topic it was asked about.

The first finding was striking. Take a single answer, cut it to its first 800 characters, and feed it to a simple pattern-spotting program (one that just counts letter combinations, like a spell-checker spotting habits). It correctly identified which of five AI systems wrote it 97.84% of the time — even after brand names were blacked out. Style, not content, gives the AI away. Length alone was useless by comparison.

But when answers were grouped into brand-level profiles across topics, the picture flipped. A model trained on one topic failed completely on another: all 22 gift-recommendation profiles were matched to the wrong AI. One likely reason is that the chatbots recommend very different numbers of brands by topic (8.4 per answer for gifts, versus under 1 elsewhere), so the numbers drift. Single answers transferred well; summed-up brand behaviour did not.

The practical lesson is double-edged. If you run a brand, do not trust a competitor's tidy 'AI visibility score' — it may be measuring the question, not the chatbot. If you're a regular user, note this: AI answers can be fingerprinted from their writing style alone, so 'anonymous' AI output isn't really anonymous. The study also caught a real-world flaw: 83% of one AI's answers were cut off mid-sentence by a length limit, quietly corrupting the data.