The Maze: AI shopping answers can draw on very different evidence, even when the question stays the same. AirOps, an AI content-workflow company, compared Meta’s Muse with ChatGPT, Google’s Gemini and Google AI Mode on buying questions during October 23–27, 2026. Muse cited more material but returned much shorter answers. That combination matters for brands: the places an engine consults and the space it gives a recommendation can both change how a product gets explained.
Muse consulted a broader pool of websites. Its average answer contained 25.1 citations, meaning references to source material, versus 6.9 for ChatGPT, 6.2 for Gemini and 7.3 for AI Mode. Across the test, Muse cited 10,171 distinct website domains, compared with 2,405, 2,372 and 2,660 respectively. The domain totals describe the test as a whole, not the number of websites behind each reply. More references create more opportunities for outside evidence to enter an answer; they do not establish that the resulting recommendation is more accurate.
Reviews had a larger place in Muse’s source mix. Review and user-generated-content sites accounted for 17.6% of its citations, against 3.4% for ChatGPT, 3.2% for Gemini and 9.1% for AI Mode. User-generated content means material contributed by users, such as reviews; it does not necessarily mean verified purchases. These percentages measure shares of citations, not shoppers or sales. The commercial implication is conditional: where independent customer accounts influence the material an engine reads, a brand’s own product page may explain only part of how it is represented.
More sourcing came with less room to speak. Muse averaged 777 characters per answer, versus 5,192 for ChatGPT, 4,010 for Gemini and 3,023 for AI Mode. Only 11.4% of Muse answers used a list or table; the other engines did so in 95.5%, 98.8% and 98.8% of answers. This was a different presentation pattern, rather than merely a shorter version of the same comparison table. A compact recommendation can leave less space for several products and their qualifications. The test does not show which brands gained or lost that space. For example, an engine could consult many product discussions but return one brief recommendation; that is an illustration of the mechanism, not a reported purchase.
Muse rarely handed the decision back as another question. Just 0.2% of its answers asked a question back, compared with 73.6% for ChatGPT, 39.5% for Gemini and 91.5% for AI Mode. That supports AirOps researcher Jason Saltzman’s interpretation of a more decision-oriented response style. It does not prove better decisions or completed purchases. The post does not disclose sample size, geography, model settings or repetitions, so this remains a five-day comparison of the tested engines, rather than a permanent ranking of their capabilities.
Why it matters: Muse combined a wider evidence pool with a narrower answer. If that pattern persists, brands face two distinct competitions: entering the material an engine consults and surviving its compression into a brief recommendation. Strong product information and credible external customer evidence could serve different roles in that process. Counting citations alone would miss the second contest: what the shopper actually gets told.
Sources: AirOps Research / Jason Saltzman, October 2026.
