Study finds user context can shift AI brand recommendations

A nine-day study of 8,609 AI responses found that what ChatGPT, Gemini, Claude and Perplexity know about a user can change which brands they recommend. The findings point to a new challenge for marketers trying to measure AI visibility, especially when geography and account history shape results.

Why it matters: - Brand visibility in AI search may not be fixed. User context can change which brands appear, which complicates measurement for marketers tracking generative AI recommendations. - Geography stood out as the clearest factor in the UK-based study, with implications for brands competing in local markets against larger international rivals. - The findings suggest generative engine optimization needs to account for audience, region and measurement method, not just content and authority.

What happened: - Researchers analyzed 8,609 AI responses over nine days in the UK. - The study, The Personalization Gap: How a Model's Knowledge of the User Reshapes Brand Recommendations in Generative AI, was co-authored by Cassie Wilson Clark, an AI search visibility consultant and host of Found in AI, and Joao da Silva, co-founder of friction AI. - The team tested ChatGPT, Gemini, Claude and Perplexity across six personas, three consumer categories and 30 prompts. - The tests compared history-free APIs, temporary logged-in sessions and accounts primed with user histories.

The details: - After accounting for normal response variation, ChatGPT showed 16.0 percentage points of history-associated brand-set divergence above the study's Volatility Floor. - Gemini showed 33.3 percentage points of history-associated divergence, the largest effect among the platforms tested. - Perplexity showed a smaller effect of 7.2 percentage points. - Claude showed no detectable history-associated change in its recurring brand set beyond normal volatility, although its search behavior changed when user context was introduced. - The researchers call the measurable difference between brands recommended with and without persistent user context the Personalization Gap. - In history-primed accounts, researchers found a 6.6 percentage-point increase in UK/EU-origin brand share. - The same accounts showed a 4.9 percentage-point decrease in US-origin brand share. - History-primed accounts also produced a 6.0 percentage-point increase in citations from .uk domains. - Across 173 matched comparisons, API sampling recovered 79% of recurring brands observed in history-primed sessions. - Temporary logged-in sessions recovered 95% of recurring brands at the same sampling depth. - The study says API-based AI visibility tools remain useful, but results depend on sampling depth, collection methods and user context. - Clark also published a marketer-focused analysis, The Personalization Gap: What 8,609 AI Responses Tell Us About Brand Visibility, on the implications for AI visibility measurement, localization and audience research. - Clark and da Silva discussed the experiment in the Sept. 22 episode of the Found in AI podcast. - The full research is available on Zenodo.

Between the lines: - The study challenges the idea that AI brand visibility can be measured from a single universal response set. - The results suggest that a screenshot or one-off answer may capture an outcome, but not the full measurement environment. - Geography appearing as the strongest personalization signal hints that regional relevance may matter more for some brands than broad global authority. - The findings also show that missing brand mentions may reflect the prompt environment or account history, not just weak content.

What's next: - Marketers are likely to test AI visibility using deeper sampling, more local context and more persona-based measurement. - Brands targeting UK and European audiences may need to strengthen regional relevance and domain signals alongside broader SEO and content work. - Further research may show whether the same personalization effects appear in other markets and languages.

The bottom line: - AI recommendations are not static. User history and geography can reshape brand visibility, making measurement more complicated and localization more important.