A product recommendation screenshot is a record of one encounter. For a brand assessing AI shopping visibility, the harder question is whether the same name, recommendation and supporting links persist when the request is made again. An independent audit now provides evidence that commercial answers need to be examined as variable observations.
What the audit measured
The preprint by Lucas G. Uberti-Bona Marin and colleagues was submitted September 16, 2026, at 14:29 UTC, or 16:29 CEST. ConsumerQ contains 2,528 commercial-advice queries; the audit used 117 selected physical-product queries, repeated three times across five conditions, producing 1,536 responses from 1,755 observations.
ChatGPT expressed first-person preference in 79% of product-recommending interface responses, versus Gemini’s 7% and AI Overviews’ 2%. ChatGPT–Gemini displayed-domain Jaccard overlap averaged 5.4%; 76.7% of comparisons shared none. Within-interface repeat overlaps were 26.0%, 29.8% and 45.9%, respectively. Interface–API overlap averaged 12.0% for ChatGPT and 14.8% for Gemini.
Cross-platform differences and repeat instability are separate measurements. Visible sources do not reveal the complete internal evidence set. The full paper describes logged-out, Netherlands-based collection, predominantly English queries and limited repetitions. These findings are a bounded snapshot, not universal rates. First-person phrasing does not establish actual personal experience, paid influence or accuracy.
What should count as a visibility result?
Our editorial interpretation is that a useful brand report should distinguish several outcomes. A product can be named without its manufacturer being cited. A retailer can be cited without the retailer’s own products being recommended. A review publisher can supply a visible link while the final answer favours a different product. Combining these into a single “AI visibility” score would hide the commercial question being answered.
Consider a hypothetical retailer whose domain appears in four of twenty recorded answers. That observation would justify reporting an appearance rate of 20% within those twenty answers. It would not establish a 20% share of all shoppers, a 20% chance of purchase or a stable position for the following month. Keeping the denominator visible makes the result easier to assess and compare.
A reporting protocol for commercial teams
For a practical monitoring programme, we propose defining the business outcome before collecting answers. Track product mentions, explicit recommendations, retailer links and publisher citations separately. Save the exact wording and destination URL rather than relying only on a screenshot label or an automatically assigned sentiment score.
Each record should have a timestamp, query identifier and a clear description of the tested environment. Keep a change log for the measurement process itself: if a parser is updated or the query list changes, flag the break in the series. Otherwise, a movement in the dashboard could reflect the collection method rather than a change in brand exposure.
A useful dashboard could show the number of successful observations beside each metric and retain answers that omit the brand. Record collection failures separately. If the collector cannot read the source panel, treat that as missing evidence rather than silently treating it as an answer with no citations. Review a sample manually before presenting an automated report to a client.
From observations to optimisation decisions
The commercial temptation is to change content after a single disappearance. Our recommendation is to first ask whether the proposed intervention has a testable objective. Updating an obsolete specification, clarifying a warranty or adding evidence for a comparison can be evaluated against the underlying information need. Merely adding promotional language because one answer favoured a competitor gives a team little basis for judging success.
Separate content changes from measurement changes, keep dated copies of important pages and write down the expected result before testing. That creates a record a future reviewer can inspect even when the outcome disappoints. Visibility research becomes more useful when it can explain its own uncertainty as clearly as its strongest results.
For NetContentSEO, the actionable question is how to build commercial reporting that remains interpretable when individual answers change. The audit is evidence for taking that question seriously; it is not a formula guaranteeing recommendations, citations or sales.