Google used Rethink 2026 to say that Merchant Center feed work moves conversions, and its AI surfaces read that feed. Here is what the six conversational attributes are and what Google's 5% figure does and does not show. It includes a one-hour check you can run in Merchant Center this week.
Google Just Told Brands to Fix Their Feeds. Merchant Center Is Now Where Google’s AI Reads Your Product
By Stephen Honight, Founder of Lmo7
Google used its Rethink 2026 retail event to tell brands, in its own numbers, that the Merchant Center feed is now where its AI reads products. The numbers sit in a Google blog post, “Boost your holiday sales with these agentic commerce updates”, published on 16 September 2026 by Ashish Gupta, VP/GM of Merchant Shopping.
On the brands we work with, the feed has usually been plumbing. Whoever runs Shopping ads owns it and it gets opened when products are disapproved. The content team rarely sees it. I think that assignment is now wrong, and there is a check near the end of this post that takes about an hour and will show you where you stand.
A caveat before the mechanism. Google’s figures are Google’s own, about Google’s own product, and one of them comes from a single brand test. The mechanism still holds up, and the work it points to is work you should already be doing on your product pages.
How the feed reaches Google’s AI surfaces
My reading of Google’s material is this. The feed you submit to Merchant Center populates the Shopping Graph, and the Shopping Graph is what Google’s AI shopping surfaces read: AI Mode, AI Overviews, Gemini and the new checkout Google’s AI completes for the shopper (Google blog, 16 September 2026). If that is right, the feed is the source document for how Google’s AI describes and recommends your products.
There is a second effect outside Google. Semrush published a study on 20 January 2026, “ChatGPT Searches Google Shopping to Create its Recommendations”, written by Leigh McKenzie. Semrush ran 100 shopping prompts, each five times. The top ChatGPT product appeared in Google Shopping’s top three results 75% of the time (Semrush, January 2026, 100 prompts, 5 runs each). Semrush also found that ChatGPT runs a second, encoded layer of shopping fan-out queries against Google Shopping.
Semrush is open about the limits. Its own caveat is that the sample is small and the result is a starting point rather than a settled number. I agree with that framing. The 75% comes from one study with a small sample (Semrush, January 2026). It does suggest that a well-kept feed pays on surfaces you do not control directly, including the ChatGPT carousel.
Source for both points: Google blog, 16 September 2026, and the Semrush study at https://www.semrush.com/blog/chatgpt-searches-google-shopping/.
What the conversational attributes are
In May 2026 Merchant Center added conversational attributes, documented in Google’s Merchant Center help (answer 17085370, linked from the 16 September post). You submit them through a supplemental feed or the API, and all six are optional:
- question_and_answer
- document_link
- related_product
- item_group_title
- variant_option
- popularity_rank
I am working from the attribute names here, so check Google’s Merchant Center help for the exact spec and format before anyone builds a supplemental feed.
Three of the six map onto content work that good teams already do on product pages.
The question_and_answer attribute is the same discipline as Q&A seeding for Amazon’s Rufus. You work out what shoppers ask before they buy, then answer in plain sentences that a person and a machine can both lift out whole. If your Amazon listing already carries those answers, the writing is done. It needs a home in Merchant Center.
The related_product attribute is your cross-sell and substitution logic. Which product goes with this one, and which should a shopper consider instead. Many PDPs carry a version of this in a “you may also like” module, but the logic behind it is seldom written down as a deliberate statement.
The variant_option attribute is variant language. How you describe the gap between sizes, flavours, colours or formats, in words a shopper would use. Conversational search is where variant confusion costs you, because a shopper asks for the lighter one or the one for sensitive skin and the system needs something to match against.
This is PDP content moved into the place where Google’s AI reads. The writing is the slow part, and Merchant Center only gives it a field. If the answers already exist on your product pages, reuse them in the supplemental feed.
Google offers one data point on whether submitted content gets used. In testing with lululemon, conversational attributes submitted by the brand were incorporated 50% of the time in relevant product recommendations in AI Mode (Google blog, 16 September 2026). That is Google’s own test with one brand, so I would not forecast your result from it. It does show that submitted content gets used when the field is filled, and an empty field gives the system nothing to quote.
What Google’s 5% figure shows
The headline number in Google’s post is this: “On average, merchants adopting core Merchant Center feed best practices see a 5% increase in conversions the following month.” (Google blog, Ashish Gupta, 16 September 2026.)
It is easy to read too much into that. The figure is Google’s own, about Google’s own product, and it covers core feed best practice broadly rather than enriched feeds or conversational attributes. What it measures is conversions the following month, not AI visibility, and Google does not promise AI visibility off the back of it.
So I would not put the 5% (Google blog, 16 September 2026) into a forecast. My view is that it works better as evidence that basic feed work is underdone. If an average merchant sees a lift after adopting the core practices, then the average feed has gaps in the core practices. Fix those before you write a single conversational attribute, using Google’s own best-practice guidance in Merchant Center. The enrichment sits on top of a clean base, and a gap in the base will undercut it.
Checkout follows the feed
The UCP section of the announcement is the shortest part of this post because it matters least for UK teams today, but it shows where the feed leads.
Google says UCP “already enables direct checkout for hundreds of thousands of brands and retailers across Google” (Google blog, 16 September 2026). Its named example is the group behind Coach and Kate Spade, whose checkout runs in Search, AI Mode and the Gemini app. UCP integration hub updates, covering cart transfer to the merchant site and enhanced checkout flow testing, are rolling out gradually in the US, with Australia and Canada to follow early next year (Google blog, 16 September 2026). Google also has a Business Agent for YouTube ads in beta in the US.
The Merchant Center help page (support.google.com/merchants/answer/16837055, read 2 October 2026) sets the boundaries. UCP-powered checkout currently applies to products with eligibility in the United States, Canada and Australia, for participating merchants. It is for select merchants only. Only listings using the native_commerce (checkout_eligibility) attribute display the Buy button, and the merchant remains seller of record.
The UK is not on that list. The practical point is that checkout eligibility is an attribute on a listing in your feed, so a clean feed with a named owner is the preparation. If nobody owns yours, assign one now rather than when the surface arrives.
The ChatGPT contrast
We covered the ChatGPT side in an earlier post, “Product Feeds Are Rising. Your PDP Still Decides Whether You Show Up in AI Shopping” (https://www.lmo7.com/blog/product-feeds-rising-pdp-decides-ai-shopping-2026). It covered Profound’s finding that 88.29% of ChatGPT Shopping offers were sourced from the PDP (Profound, Allen Wu, 24 June 2026).
Put the two together and the picture is more useful than either alone: ChatGPT reads the PDP first, while Google’s surfaces read the feed first, and the Semrush finding connects them because Google Shopping rank shows up inside ChatGPT’s recommendations (Semrush, January 2026, small sample). Work on only one of these and a large share of how AI describes your product goes unmanaged.
The feed owner and the content owner therefore need to share a source of truth. If the PDP says one thing about a product and the feed says another, one AI surface will quote each.
A check you can run this week
In our experience this takes about an hour and needs no new tooling.
- Open Merchant Center and run diagnostics. Note every disapproval and warning, because core feed hygiene is the area Google’s 5% figure refers to.
- Take your top 20 SKUs by revenue. Count how many carry any conversational attribute. Write the number down. Whatever the count is, it is the baseline you will measure progress against.
- If you sell in the US, Australia, Canada, India or New Zealand, open the AI performance insights tab. Google says it is now generally available in those five markets, and it compares your brand’s share of voice with other brands across surfaces like AI Mode and AI Overviews (Google blog, 16 September 2026).
If you are in the UK, you will not find that tab yet. Google’s post does not list the UK, and I am not 100% sure when it will arrive. The reporting comes later, but the feed work can start now, because AI Overviews and Gemini are live for UK shoppers even where Google’s Merchant Center reporting on them is not.
Write the feed and the PDP together
The feed and the PDP draw on the same underlying product truths: what the product is, who it suits, how it differs from the next one along and what a shopper is likely to ask. Written separately, the two versions drift apart. Do them together, from one brief, with one owner.
I would also be wary of anyone who tells you the answer is only the feed. A complete feed gets your product into Google’s AI reading. What decides which eligible product gets picked is a different set of signals, led by reviews and brand authority, and those take a sustained programme that should run alongside the feed work rather than after it.
We have seen what sustained AI search work can do. For Trip Drinks, in a 60-day AI search engagement, average position across major LLMs moved from 7th to 3rd and AI referral traffic rose 33% (Lmo7 client work, Trip Drinks, 60-day engagement, from the Lmo7 case study record). That is not a Merchant Center result and I would not claim it as one. It does show that position in AI answers can move inside an engagement window when the content is built for it.
Tracking matters here too. If you change feed content and PDP content together, you need a way to see which AI surfaces moved and by how much. Without it you are guessing at what worked. If the team is short on capacity, that is the point to bring in help.
What to do next
Challenger: Agentic Ops (£500/month). Start here as the default. It covers feed and structured data work, and we run it alongside Agentic Tracking & Content Optimisation (£1,000/month) and Amazon Content Optimisation (£750/month) so the feed, the PDP and the Amazon listing stay aligned off one brief. That is the shared source of truth this post argues for, applied commercially.
DaaS, our data tier: £250 + VAT per month per marketplace. Direct API access with upskilling for teams that want to run the work themselves and see whether feed changes moved anything. With Google’s UK reporting not live yet, your own measurement is the evidence for the next round of budget decisions.
Enterprise: workshops from £5,000. For multi-brand Merchant Center governance and stakeholder alignment, where several teams each hold a piece of the feed and nobody holds the whole. The workshop lands a single owner and a single process before UCP checkout reaches your market.
Run the top 20 SKU check first. The gaps it surfaces will tell you how much of this your team can cover in-house and where help earns its fee.
Stephen Honight is the founder of Lmo7, an agentic commerce agency that helps brands understand what AI-driven commerce means in practice and what to do about it. Current clients include Trip Drinks, Veloforte, Brown-Forman, Haleon, Pelotan and Symprove.