Discussion
  • eCommerce SEO

Merchant Center is adding conversational attributes and AI share-of-voice insights. Will you write feed copy for AI shopping separately from on-page copy?

Published 1 reply

Discussion starter: an ecommerce SEO question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.

Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.

1 reply

  • admin
    Admin
    Original poster

    Separately formatted, yes. Separately invented, no. I would keep one set of verified product facts and render it differently for the feed and for the page, which is what well-run feeds have done for years: feed titles lead with attributes and carry no promotional text, while the page carries persuasion, depth and proof. What I would not build is a third “AI copy” layer containing claims the product page does not make.

    The bigger surprise is how little of this is writing. Google’s Merchant Center help page lists six conversational attributes, and only one of them, question and answer, is prose. The others are document links, related products, an item group title, variant options and a popularity rank. So the job is mostly product data work, with one narrow writing task.

    What Google has announced, and what it has not

    At Google Marketing Live in May 2026, Google said retailers globally can use conversational attributes so their product descriptions reflect the more conversational way people search. It also announced AI performance insights that compare your brand’s share of voice on AI surfaces with similar brands, for Australia, Canada, India, New Zealand and the US, with rollout over the following months. As of September 2026, Google’s help page says the insights are available for English-language queries in Merchant Center accounts in those five countries, covering AI Mode and AI Overviews.

    The help page adds the operational detail. The attributes are optional and complement your primary product data. Google recommends submitting them through a supplemental data source, though a primary data source or the Merchant API also works. Adding them won’t affect the approval status of existing products. And if the details already sit in your description, product highlight or product detail attributes, you don’t need to duplicate them. Google frames the goal as helping shoppers find product information across AI-driven surfaces such as AI Mode, while also enhancing traditional search experiences.

    What Google has not published: how much weight AI Mode gives these attributes compared with the description or the landing page, whether question and answer text is shown to shoppers word for word, or any uplift figures. Claims that “AI-optimised” feed copy raises visibility by a given amount are ahead of the public evidence.

    Attribute What it holds Writing or data? Who I would have own it
    question_and_answer Question and answer pairs about the product Short factual writing Customer service and merchandising, working from real questions
    document_link Links to product documents in PDF, such as manuals or spec sheets Data Product or technical team
    related_product A relationship type plus an identifier for the related item Data Merchandising, from accessory and compatibility data
    item_group_title A name for the variant family, such as “Organic Cotton Men’s T-Shirt” Naming Feed or PIM owner
    variant_option Name and value pairs for variant dimensions Data PIM owner
    popularity_rank A percentage value, such as 95.5 Computed data Ecommerce analytics

    The feed and the page already had different jobs

    Feed titles do much of the matching work for shopping queries and appear in listings with little of the page’s context around them, so good ones front-load brand, product type and the attributes shoppers filter on, such as size, material, colour or compatibility. A product page H1 can be shorter because the page shows everything else. Merchant Center’s product data specification also asks merchants to keep promotional text, such as free shipping or sale messages, out of titles and descriptions, which rules out much of what persuasive page copy does.

    So many mature feeds already carry different titles and plainer descriptions than the page, and that separation is healthy. It is a different rendering of the same facts. The new temptation is a third version, written in a conversational register “for the AI”, and that is where I would stop.

    Why I would not write a separate AI layer

    • Consistency is checked, at least partly. Merchant Center already compares price and availability against your landing pages. Google has not documented a text-matching check for descriptive claims, but a benefit that appears only in the feed is one a shopper cannot confirm on the page, and unsupported claims can create misrepresentation risk under Merchant Center policies.
    • Drift. Two hand-maintained descriptions per SKU diverge the first time someone updates one and not the other. At tens of thousands of SKUs, reconciling them becomes a job nobody is resourced to do.
    • Reach. Google’s surfaces read your feed, but other assistants take product data through their own crawlers, feeds and protocols. The product page is the one input any crawler-based assistant can reach, so facts that live only in a Google feed reach only the surfaces that ingest that feed.
    • Evidence. No public test that I can point to shows feed-only phrasing outperforming page-consistent phrasing in AI Mode.

    Agentic checkout raises the stakes on accuracy rather than on voice. Google’s May 2026 shopping updates described shoppers checking out with Google Pay across retailers, in Search and in Gemini. In September 2026 Google opened native checkout to eligible US Shopify merchants, with eligible products included automatically. For shoppers who buy this way, feed data may be most of what they see before buying, so a feed claim the page does not back up can turn into a returns or complaints problem.

    One fact base, three renderings

    The structure I would use is a single approved source of product claims, with each output channel drawing from it in its own format.

    Verified product facts: specs, materials, dimensions, compatibility, certifications, care
    ↓
    Approved claim library: only what you can substantiate, with the source recorded for each claim
    ↓
    Rendered three ways: feed attributes (attribute-led title, plain description, highlights, details), conversational attributes (question and answer pairs, relationships, variants), and the product page (H1, copy, specs table, visible Q&A, reviews, Product markup)
    ↓
    Check: every claim in every rendering traces back to the library, and nothing appears in the feed that the page cannot show

    This is closer to how a newsroom handles one story across print, web and a push alert than to writing three stories. The headline, length and emphasis change with the channel. The facts do not, and a correction made once flows to every version.

    Where feed-first work does pay

    • Fill the basics before the new attributes. Google says conversational attributes need not repeat what is already in the description, product highlight or product detail attributes. If those are thin, fixing them comes first, and every surface that reads the feed benefits.
    • Write question and answer pairs from real questions. Mine support tickets, pre-sales chat, review questions, on-site search and return reasons. Answer in one or two factual sentences a staff member would stand behind.
    • Publish the same pairs on the product page. That keeps the claims verifiable and gives crawlers outside Google the same answers. FAQ rich results are gone from Google, so the on-page version earns its place by answering shoppers, not by winning a SERP feature.
    • Load relationships from data you already hold. Compatibility tables, accessory rules and cross-sell logic may already live in the PIM or platform. Feeding them into related_product answers “what works with X” questions without any new copy.
    • Name variant families properly. Use item_group_title and variant_option where your platform’s parent names are generic or the variant dimensions are unclear.

    How the answer shifts by catalogue

    Situation What I would do Why
    A few hundred SKUs, platform-generated feed Write strong page copy, let the feed inherit it with attribute-led titles, and add question and answer pairs to top sellers by hand Manual work is feasible and consistency comes almost free
    Tens of thousands of SKUs with a PIM Rules-based feed rendering from attributes, with question and answer pairs only for top-revenue and high-question SKUs Hand-written copy at that scale drifts
    Reseller of branded goods using manufacturer copy Put the effort into question and answer pairs, relationships, and delivery or service facts that competitors lack Every reseller has the same manufacturer description
    Regulated or technical products No feed-only claims, compliance review of every answer, and document_link for manuals and spec sheets Claims need substantiation, and the documents carry it
    Strong page content, thin feed Port page facts into product highlight and product detail before anything new Google says those need not be repeated in conversational attributes

    How to tell whether any of it moved visibility

    Google’s help page describes share of voice against competitors, filterable by product category, and does not mention a view by individual product. So the report can show a trend without showing which products moved it. For a readout you need a matched-product test.

    1. Split one category into two matched groups by price band, sales, stock depth and seasonality. The treatment group gets conversational attributes plus filled highlight and detail gaps. The control group stays as it is.
    2. Primary metric: item-level impressions and clicks, treatment against control, from Merchant Center for free listings and from Google Ads for Shopping ads.
    3. Secondary metrics: conversion rate, return rate and repeat pre-sales questions for treatment products. If the question and answer pairs are doing their job, avoidable returns and repeated questions should fall even where visibility does not move.
    4. Duration: long enough for the feed to be reprocessed plus four full weekly cycles, because shopping demand varies by day of the week. Avoid sale events, and extend the test if items have low impression counts.
    5. Hold paid media steady. The same feed drives Shopping ads, so keep bids and budgets stable for both groups and tell the paid team before the test starts. A bidding change can swamp the signal.

    If treatment products gain impressions and clicks relative to control without hurting conversion, roll out by revenue tier. If the result is flat, keep question and answer pairs where they reduce support load or returns, and do not build a separate copy function for AI shopping.

    The first week’s work is small. Export your top 200 products by revenue, list the questions customers ask about each from support and reviews, check which answers already exist on the page and in the feed’s highlight and detail fields, and fill the gaps in both places at once. That shows quickly whether you have a writing problem or a data problem, and in many catalogues it turns out to be the second.

    Need help with this on your own site?

    Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.

    See SEO services

Add a reply

Replies are open to approved members. Log in or apply to join.