71 percent of the citations ChatGPT returned in a three-month study pointed to pages that never appeared on Google's first page, and in the same week three platforms opened new routes for product catalogs to reach AI shopping surfaces. As of November 11, 2026, the useful question for an ecommerce team is narrower than a channel strategy: what does a product listing have to declare before an agent will quote it. None of the three announcements came with merchant numbers, so the work this week is a field-level audit of the catalog rather than a new platform bet.
Three routes opened in one week
The three moves are not the same kind of change. Microsoft is adding catalog listings to an assistant that already answers questions, Amazon is letting stores outside its marketplace submit a feed into its own shopping assistant, and Google is extending checkout inside an answer. What they share is a single dependency: the platform reads a structured feed instead of a web page, and it uses whatever fields that feed carries.
That makes the week less about traffic and more about data quality. A feed with thin titles, missing materials and no delivery window gives an assistant nothing to quote, so the model fills the gap from reviews, forums or a competitor. For a category where consumer demand shifts quickly, that gap is where the sale is lost before anyone visits the store.
A merchant does not have to join all three. The point is that the number of surfaces reading the same catalog went up, and each one has its own eligibility gate. A field that is optional on one surface can be the reason a product never appears on another.
The feed is now the listing
For most of the last decade, the listing was a page and the feed was plumbing: a file that pushed price and stock to a shopping tab. The order has flipped. When an assistant builds a recommendation, the feed decides what the product is called, which variant exists, what it costs and when it arrives, and the page mainly confirms it.
What a declared spec looks like
A declared spec is a fact with a boundary. Not "cotton shirt" but the fabric weight, the fit, the sizes in stock and the return window. Not "fast delivery" but the cutoff hour and the regions covered. Each of those facts needs a source and a date, because an answer assembled in November from a page last updated in March will describe last season's catalog with this season's confidence.
Two failures repeat across catalogs. The first is a specification that exists only inside an image, which a feed cannot read. The second is the same attribute written differently on the page, in the feed and in the marketing copy, which teaches the model that the merchant is unreliable rather than that one field is wrong.
Eligibility is not citation
Being purchasable and being quoted are different states, and teams keep treating them as one. Eligibility is a gate a merchant can pass by accepting terms, keeping policy pages complete and listing catalog-eligible products. Citation is a competition that happens after the gate, against every other listing that passed it.
The 14.8 percent problem
The citation study tracked 7,949 prompts across five engines and 239,841 citations to 27,765 domains between July and September 2026. Alongside the finding that 71 percent of ChatGPT citations went to sites outside Google's first page, it reported that Google AI Mode kept only 14.8 percent of the domains it had cited before. Visibility is not a position a merchant holds; it is a sample that moves.
That number changes how a team should measure. A single afternoon snapshot proves almost nothing. What proves something is the same question set, run on a fixed cadence, with the cited URLs stored next to the version of the listing that produced them. Without that pairing, a drop in mentions cannot be separated from a model update.
The European comparison box tells a similar story from the retail side. BrightEdge data across Spain, Italy, France and Germany found Google's own shopping service appearing 16 to 35 times more often than its closest rival inside the box that replaced free product listings, with 71 percent of comparison-site appearances in the top four positions. The box is premium space, and most of it is not available to a merchant who assumes listing everywhere is enough.
Refusals happen before the shopper arrives
There is a quieter failure mode that no feed audit catches. An agency sent an automated shopping agent to 200 retailers in the UK and the US, twice, and 29 refused it, including large chains whose published crawler rules allow the ChatGPT, Perplexity and Claude crawlers. Refusal ran at about 20 percent among large retailers and 2.5 percent among smaller ones.
The gap sits between two settings that nobody reconciles: robots.txt, which states who is welcome, and bot protection, which decides who actually gets through. A store can say yes in one place and no in the other without anyone noticing, and the result looks like weak visibility rather than a blocked door. For a merchant preparing for the peak weeks, that is a cheaper thing to check than any content project.
The audit to run this month
The three surfaces read different fields first, so the audit should be ordered by what each one needs rather than by what is easiest to edit. A title rewrite reaches every downstream answer, while a description edit reaches one surface at a time. The table below is a starting order, not a complete schema, and it assumes the catalog export is already clean enough to compare month to month.
| Surface | What it reads first | Field to fix first |
|---|---|---|
| Copilot through Merchant Center | Catalog eligibility and title structure | Title with brand, material and variant instead of a generic noun |
| Alexa for Shopping | Feed submission and active item status | Availability and shipping window per variant |
| Google AI Mode | Feed attributes plus policy pages | Price condition and return terms inside the description |
The order matters because a fix at the title level changes every downstream answer, while a fix in the description changes one. Start where the leverage is highest, and keep a dated copy of the export so next month's comparison is possible at all.
None of this requires a new team. It requires one owner for the catalog record and one for the public page, plus an agreement that the two carry the same numbers. Most mismatches come from a specification that lives in an image on the listing while the feed sent to the catalog leaves it out.
FAQ
Do product pages still matter if agents read the feed?
Yes, but their job changed. The feed supplies the structured facts and the page confirms them, including the policy details a shopper checks before paying.
Should a small store join all three routes?
No. Check eligibility first. A store that cannot keep delivery windows accurate will produce worse answers by joining more surfaces with the same thin data.
How often should citations be re-checked?
Monthly for a stable catalog, weekly for anything with price or stock movement. Answers shift when models update, so an old snapshot describes a different system.
What is the single cheapest fix?
Reconciling robots.txt with bot protection. It costs an hour and removes a refusal that looks like invisibility from every report.
