News: On September 16, 2026, Google published on its Shopping blog a set of pre-holiday commerce updates, and with them the first usage figure for its conversational attributes: in a test run with the brand lululemon, the attributes submitted were incorporated "50% of the time in relevant product recommendations in AI Mode" (Google, September 16, 2026).
Merchant Center's six conversational attributes are not new: Barry Schwartz documented them on Search Engine Roundtable back on May 26, 2026. What is new since yesterday is that we finally know whether they do anything. For four months, filling those fields was a bet: you wrote answers for an AI with no way of knowing whether it read them. Google has just answered, with a number.
What Google announced on September 16
The post covers several fronts. Cart transfer first: a shopper can now move their cart from Google to the merchant's own site to complete the order, a capability "gradually rolling out in the U.S., with Australia and Canada to follow early next year". The AI performance insights report, which measures a brand's share of voice in AI traffic, is now open in Australia, Canada, India, New Zealand and the United States. And finally the product feed enhancements: conversational attributes, a video link attribute, loyalty data integration.
The first number is the one that matters here. It is precise, it is dated, and it addresses a question the industry has been asking for a year: when you structure your data for answer engines, does any of it come back?
The six fields that talk to the AI
The Merchant Center documentation names them unambiguously. question_and_answer lets you submit question-answer pairs about a product. document_link attaches PDFs: manuals, assembly instructions. related_product declares a typed relationship to another product, a compatible accessory or a substitute. item_group_title gives a readable title to a family of variants. variant_option names the axes of variation, size, colour. popularity_rank places a product as a percentile within your own catalogue.
Three of these fields had no equivalent in the standard product specification. They are not just more tags: they are a space where you write in natural language, for a machine that answers in natural language.
On implementation, Google recommends "a supplemental data source", joined to existing products by their id. Adding directly to the primary feed and the Merchant API both remain available. The recommended route has a clear operational benefit: it touches neither the primary feed nor its approval process, so it does not expose your catalogue to a mass rejection over a formatting error.
Are your products cited by AI engines?
We test your commercial queries on AI Mode, ChatGPT and Perplexity, look at who gets cited instead of you, and tell you what is missing from your content and your product data.
Why this is a content story, not a technical one
You could file this announcement under technical integration: a feed, some columns, an import. That would miss the point. The question_and_answer field cannot be populated from an ERP. It requires knowing which questions customers actually ask before buying, and answering each in one accurate sentence. That is editorial work, done by someone who knows the product and the customer, not a database mapping.
The logic matches what we have been seeing on pages for months: answer engines pick up what is phrased in a self-contained, verifiable way. What is new is that Google is opening a dedicated channel where that phrasing is submitted directly, without going through the crawl of a page. For a catalogue, that is an extra surface to speak on, and it is still largely unoccupied. Our six GEO pillars treated structured data as a supporting signal; in e-commerce, it is becoming a channel in its own right.
Then there is measurement. In July, Google opened the beta of the report that gives an official share of voice inside AI Mode; it now moves to general availability across five countries. So the produce, measure, correct loop genuinely exists, just not everywhere yet.
What to do this week
- Pull your twenty real questions. Not invented ones: take support tickets, pre-sales questions, the customer reviews that open with "does this actually...". That is the raw material for
question_and_answer, and you already have it. - Start with your top fifty SKUs.
popularity_rankexists precisely to prioritise. Attacking a whole catalogue at once is the surest way to ship nothing. - Go through a supplemental source. It is Google's own recommendation, and it protects your primary feed while you iterate.
- Date your starting point. If you operate in one of the five covered countries, record your share of voice before enrichment. Without a baseline, in three months you will not know whether the work paid off.
The limits of what we know
They matter, and ignoring them would turn good news into an empty promise. The 50% figure comes from a test with a single brand, reported by Google, with no independent audit and no published methodology; lululemon is a high-awareness apparel catalogue, which is nobody's average case. The wording matters too: it is 50% of relevant product recommendations in AI Mode, not 50% of AI answers in general.
The +5% conversion figure deserves the same caution, for a different reason: in the post it is attached to merchants applying general feed best practices, not specifically to conversational attributes. Presenting it as the return on these six fields would be a slide; it is a correlation observed on a broader practice. Finally, France is not among the five countries where measurement is open, and a missing number is not a favourable one.
We checked every field name and definition in this article against the official Merchant Center documentation, and we confirmed every figure against the September 16, 2026 post. Anything we could not verify is flagged as such.
The Cicero take
There is an irony in this announcement. For twenty years, SEO consisted of guessing what the engine wanted to read. Here, Google opens a form and asks: tell us what your customers ask, and what you answer. The difficulty is no longer technical, it is editorial, and that is exactly what will stall most catalogues: filling six columns is easy, knowing what to put in them is not. The 50% figure comes from a brand that knows its customers. It does not transfer to a feed filled in a hurry.
Sources
- → Google, Shopping blog, September 16, 2026: primary source. The 50% lululemon test, the +5% conversion figure for feed best practices, cart transfer, AI performance insights opening across five countries, video link and loyalty data attributes.
- → Google Merchant Center Help — conversational attributes: official documentation. Exact names of the six attributes, the role of each and the submission methods, quoted verbatim in this article.
- → Search Engine Roundtable, Barry Schwartz, May 26, 2026: independent report, used to date the initial opening of the six attributes and cross-check that they are optional.
- → Search Engine Land, September 16, 2026: independent report on the September 16 agentic commerce announcements.
Frequently asked questions
What is a conversational attribute in Merchant Center?
It is an optional field added to the product feed to help, in Google's documentation wording, AI systems and conversational agents better understand your products' specific nuances. Six attributes are documented: question_and_answer, document_link, related_product, item_group_title, variant_option and popularity_rank. They are optional and complement the existing product data specification rather than replacing it.
Where does the 50% figure come from?
From the post Google published on its Shopping blog on September 16, 2026. It refers to a test run with the brand lululemon, in which the conversational attributes the brand submitted were incorporated 50% of the time in relevant product recommendations in AI Mode. It is a test with one advertiser, reported by Google itself: not an independent audit, not a market average, and not a guaranteed outcome for another catalogue.
How are these attributes submitted?
The Merchant Center documentation lists three routes: a supplemental data source (recommended), adding them directly to the primary data source, or the Merchant API. The supplemental source is recommended because it joins onto existing products by their id, without touching the primary feed or its approval process.
Is this available in France or the EU?
Conversational attributes belong to the Merchant Center feed specification and are not announced as restricted to a country list. The AI performance insights report that lets you measure the effect, however, is described in the September 16, 2026 post as available in Australia, Canada, India, New Zealand and the United States. France is not on that list. In practice: the work is possible before the measurement is.
Growth and SEO & GEO content strategist, I founded Cicero to help businesses build lasting organic visibility: on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.
LinkedIn →