Google has spent August expanding its generative AI performance report across Search Console, a rollout trade press has tracked as broadening rather than complete — as recently as this month, Google’s John Mueller confirmed it still wasn’t live for every domain. It carries a single metric: impressions inside AI Overviews and AI Mode, split by page, country and device, with no clicks, no queries and no position. Google’s guidance for those same features holds, in substance, that optimising for generative AI search is still SEO, not a separate discipline requiring new tactics. The retailer is left with a line that moves and, in that console, nothing to move it with.

The timing matters because AI referral has stopped being marginal for apparel. Adobe Analytics figures reported by Digital Commerce 360 put AI-referred traffic to US retail sites up 62% year over year in July. Growth alone would not force a budget line, since a small base can grow quickly and remain rounding error. Quality forces it: those visits converted 60% better than non-AI traffic, the eleventh consecutive month AI referrals have out-converted the rest of the site. A channel that converts better than everything around it gets a target next year, and a target needs an instrument.

Google’s guide is unusually direct about what does not work, and the list covers most of what is currently being sold to retailers. It says llms.txt files may be crawled but are ignored, and that adding one won’t affect a site’s visibility or rankings. It says structured data isn’t required for generative AI search, with no special schema.org markup needed. Content need not be chunked, and there is no separate register to write in for machines. Each claim is probably true, which is the difficulty, because a document that removes every tactic leaves the retailer holding a metric and an instruction to carry on as before.

An impressions count with no query dimension cannot tell a retailer whether the line moved because their catalogue improved or because Google widened where AI Mode fires.

Google’s own August numbers make that point more efficiently than any argument does. During the same rollout window, a logging error depressed impression counts between 13 and 17 August, as Google’s data anomalies log records. Sites checking their numbers in that window would have seen a dip they hadn’t caused. That Google disclosed and bounded the fault is to its credit. The lesson is narrower and permanent: the baseline moves for reasons internal to Google, and the site owner watching the line cannot separate those movements from their own.

The lever fashion retailers do have is real, and Google reports it in a different console. The guide’s one retail-specific instruction is to use “Merchant Center (such as Merchant Center feeds) and Google Business Profiles” so that products appear in AI responses. At Google Marketing Live in May, the company launched conversational attributes in Merchant Center: structured product fields and rewritten descriptions whose stated job is to help Google’s AI systems match products to conversational shopping queries across AI Mode and Gemini. Writing for machines is precisely what the guide told website owners they did not need to do.

Google can reconcile the two positions, and the reconciliation is the problem. The guide governs your website; Merchant Center governs a feed you hand Google directly, and the rules for the two were never the same. But AI performance insights, the report that measures the feed side and carries what Search Console lacks, remains a limited US pilot with Australia, Canada, India and New Zealand promised. Its headline number is share of voice against a named competitor set, broken out by query type and by stage of the shopping journey. Most retailers therefore hold the number they cannot act on and lack the number they could.

Apparel is the category with the most to gain from that feed-side work, and its position in Adobe’s readability rankings has moved fast. The company’s content visibility checker, which scores how much of a page a language model can read, put apparel at 51% in May, trailing both cosmetics and electronics. Apparel homepages scored 62% in the same survey, which located the shortfall on the product detail page, the only page a shopping answer needs.

By July, the same tracker had apparel at 76%, ahead of electronics and cosmetics alike. Nothing in the public data explains the swing: a rewritten catalogue, a seasonal photography refresh, and a shift in how much of AI Mode’s traffic touches apparel at all would each move the number, and Adobe’s report cannot separate them. That is the same attribution gap the impressions line carries, on a smaller scale and over a shorter stretch of time.

The strongest case against this reading is that meter and lever are joined after all. Better product data should make more of a page quotable, quotable pages should appear more often in AI answers, and the Search Console impressions line should therefore climb as the catalogue improves, which would make it exactly the before-and-after instrument retailers need. That is the condition on which this argument fails, and it deserves stating at full strength. It does not survive the missing query dimension. A retailer who rewrites forty thousand product pages across two quarters and watches impressions rise cannot separate the rewrite from seasonal demand, from a core update, or from Google expanding the share of searches that trigger AI Mode at all. Attribution requires holding demand constant, and the report offers nothing to hold it with.

The missing metrics are scheduled, which changes the timetable rather than the problem. The CMA’s publisher conduct requirement, which we examined in July, obliges Google to give publishers clearer, more detailed engagement metrics. If Google extends those metrics beyond UK publishers, the measurement gap closes and the impressions line becomes attributable. The control gap does not close with it, because conceding a lever would concede a separate optimisation channel, and denying that channel is what the guide exists to do. Read the line as an audience estimate rather than a performance metric, useful to know and impossible to manage. The decision this quarter is not which AI tactic to buy. It is whether to fund catalogue legibility on argument alone, because the number that would settle it sits in a console most retailers have not been given.