Brands Hold, Evidence Turns Over: 12 Weeks of Ecommerce Citations Across Three AI Engines

Prior analyses in this series looked at which brands appear in AI answers. This one separates that question into two: which brands the engines name, and which URLs they cite to support it. Those turn out to behave differently enough that measuring them together obscures both.

We used BrightEdge AI Catalyst to track ecommerce and shopping prompts across ChatGPT, Gemini, and Google AI Overviews for 12 weeks, from 10 May 2026 through 26 July 2026. For each engine we measured two things weekly: share of mentions, meaning which brands the engine names, and share of citations, meaning which URLs it cites as evidence. The finding is in the gap between them.

The Brands Hold, the Evidence Turns Over

Comparing like for like at matched volume, share of brand mentions changed an average of 25% week over week. Share of citations changed 39% on ChatGPT and 41% on Gemini.

The turnover figures are sharper. On Gemini, 37% of meaningfully cited URLs were not cited the prior week, against 3% of its mentioned brands. On ChatGPT the same comparison is 15% against 5%.

The practical reading is that brand presence in AI answers is comparatively durable while the evidence layer supporting it is renegotiated weekly. A brand that appeared last week will very likely appear this week. The pages the engine uses to justify that appearance are substantially different.

What We Analyzed

We measured two metrics weekly per engine: share of mentions, the proportion of brand mentions in tracked answers attributable to a given brand, and share of citations, the proportion of cited URLs attributable to a given page or domain. Volatility is reported as average absolute week-over-week change in share, and as the proportion of entities appearing in the current week that did not appear the prior week. Comparisons are reported as shares, rates, and rank positions rather than raw counts, so that differences between engines and between metrics remain the unit of analysis.

Data Collected

Key Finding

At roughly 40% weekly citation turnover, a single week's movement in citations carries little diagnostic value. The measurement discipline that follows from this is a baseline rather than a target.

Without trended data at weekly cadence, a team cannot separate normal churn from a genuine positional change. That produces two failure modes, and both are expensive. The first is reacting to routine turnover as though it were a problem. The second is failing to detect a real loss because it resembles the noise surrounding it.

Four practical reads follow from the observed baselines.

Fluctuation is not the problem. Fluctuation without a baseline to measure it against is.

Google's Own Surfaces Do Not Agree

Gemini models power Google's shopping stack, including AI Mode product panels and the product card review summary. Google's two AI surfaces nonetheless draw on substantially different evidence.

Community and video sources account for 26% of AI Overviews citations and roughly 5% of Gemini's.

The engines converge substantially on which brands belong in an answer and diverge substantially on what to cite in support. A content program built for one engine's evidence preferences should not be assumed to transfer.

AI Overviews Reweighted Its Source Mix

Over 11 weeks, YouTube's share of AI Overviews citations moved from roughly 31% to roughly 65%, while community and marketplace sources declined over the same period. Nothing about those sites changed during the window. The engine's weighting did.

AI Overviews is also considerably more concentrated than the other two engines.

Placement in AI Overviews is closer to winner-take-most, while the other two engines distribute citations across a longer tail.

A brand measuring only its own visibility during this period would have registered movement without identifying the cause, because the cause was a reweighting across the entire source mix rather than anything specific to that brand.

Why Ecommerce Feels This First

Google assembles a product card from multiple pipelines, of which the merchant feed is one. Reviews, video, forum content, and an AI-generated review summary populate the remainder.

The feed is the only component of that card that does not change without merchant action. Attributes left unpopulated are therefore filled from the layer turning over roughly 40% a week, from sources the merchant does not control, in language the merchant never reviews.

Unpopulated fields are not neutral. They are delegated.

What Marketers Need to Know

Measure two things, not one. Share of mentions indicates position. Share of citations indicates the conditions around it. Combining them into a single visibility score obscures both, because they move at materially different rates.

Set a baseline before setting a target. At the observed turnover rates, a category-specific baseline is a precondition for identifying a trend rather than an optional refinement.

Move monitoring to weekly. The churn being measured is weekly. A monthly cadence cannot resolve it, and will report the result of a reweighting rather than the reweighting itself.

Do not assume one content strategy covers all engines. The engines agree far more on which brands to name than on what to cite, and the divergence extends to two surfaces operated by the same company.

Technical Methodology

Key Takeaways