The shopping query your product page never sees.

A shopper used to type "running shoes" into Google and scan a grid. Now they type "a good neutral running shoe with extra cushion for a heavy runner under $150" into AI Mode or the Gemini app, and an assistant reads the whole sentence before it decides what to show. Google says queries in AI Mode run about three times longer than traditional Shopping searches. Your feed was built for the short version. It carries a title, a price, a handful of attributes, and a description written to fit a results grid, and none of it was written to answer a sentence.

In May 2026, at Google Marketing Live, Google shipped the first lever that changes this inside the feed itself. Conversational Attributes are a set of optional product-data fields in Merchant Center that let you add plain-language descriptions and question-and-answer content, which Google's AI then uses to match your products against conversational queries across AI Mode, AI Overviews, and Gemini (Search Engine Land, Google blog). This is the part most teams will miss. The matching now happens against feed copy your product page never gets a chance to influence.

What Conversational Attributes actually are

The standard product feed answers a database. Title, GTIN, price, availability, color, size. Each field maps to a filter or a column, and the description has historically been a short block of marketing copy that Shopping mostly ignored for matching. The feed's job was to be machine-readable, not conversational.

Conversational Attributes sit on top of that. They are additive metadata, six optional fields that Google's documentation frames as a way to "reflect the more conversational way people search" (Google Merchant Center Help). Three of them carry most of the weight:

• question_and_answer lets you submit FAQ-style pairs as structured data. A question, then a plain-language answer. "Does it have a headphone jack?" paired with "This version doesn't have a headphone jack." You can submit many per product.
• item_group_title is the family name for a set of variants. Where title is the full SKU name like "Google Pixel 9 Pro 512GB Moonstone", the group title is just "Google Pixel 9", so an assistant can talk about the product family first and drill into a variant later.
• Document links point to PDFs tied to the product. Manuals, assembly guides, datasheets, ingredient lists, care instructions. The kind of detail a buyer asks about but rarely finds in a description.

One detail matters more than it looks. These fields are additive, so submitting them does not change the approval status of your existing products. Misformat one and you do not lose Shopping eligibility for the SKU. Google ignores the attribute and keeps serving the listing. The downside risk of starting is close to zero.

Why feed copy now competes for matches your product page never sees

Here is the shift worth sitting with. For years, the question "does this jacket pack down small enough for carry-on" was answered, if at all, on your product page, in a review, or in a support thread. A buyer typing that into Google landed on a page and read until they found out. The page did the work.

In an AI shopping answer, the assistant assembles a recommendation before the buyer visits anyone's page. It reads structured data across many products and picks a shortlist. If the packable detail lives only in your page copy and not in your feed, your product can be absent from the very comparison where it would have won. The page never gets loaded, because the product never got shortlisted.

That is the real change. Conversational Attributes are the first place you can write for these queries inside the feed, separate from the page, in language the matching layer reads directly. Your beautifully written product page still matters for the buyer who arrives. It is no longer the only surface deciding whether they arrive at all.

How to turn the questions buyers ask into attribute copy

The instinct most teams bring to a new field is to stuff it. Resist that. Google's guidance is blunt about keeping the copy clean, no ALL CAPS, no promotional text, no keyword cramming, and writing conversationally works against stuffing anyway, because forced keywords break the natural rhythm the field is built to reward.

Better source material already exists. You do not have to invent the questions buyers ask, because they have been asking them for years. Mine four places:

• Pre-sale chat logs and support tickets. The questions people ask before buying are the questions they now ask an assistant.
• Product page FAQs and the "questions" section under reviews.
• Returns data. A common reason for return is usually a question the listing failed to answer up front.
• Your own search bar. The long queries people type on your site are conversational queries in miniature.

Turn each recurring question into a question_and_answer pair, answered the way you would answer a customer, not the way you would write ad copy. Compare the two registers:

The right-hand column answers the constraint inside the question. It names the trade-off instead of hiding it. An assistant matching "a waterproof jacket I can wear commuting in light rain" can use that second sentence. It cannot use "10,000mm" without knowing what the number means for the buyer, and most buyers do not.

Write for the constraint, not the keyword. "Fits a 15-inch laptop", "dishwasher safe", "works with the old charger", "true to size". These are the phrases that turn into matches, and they read like a person because a person asked them first.

Reading AI Performance Insights without chasing noise

Alongside the attributes, Google launched AI Performance Insights, a reporting view that shows your share of voice across AI shopping surfaces, benchmarked against brands similar to yours, for journeys that start in AI Mode, AI Overviews, or the Gemini app (Search Engine Land, MarTech). It also breaks performance across the funnel and flags which product attributes you are missing with a completeness score. It is rolling out to the US, Canada, Australia, India, and New Zealand first.

This is the first time Google has handed retailers a share-of-voice number for AI shopping. Treat it carefully. Share of voice on a new surface is volatile, and the temptation will be to react to every weekly swing. Do not. A single week's drop usually tells you nothing. Three things are worth watching, and most are not.

• The attribute completeness score is the cleanest signal in the report. It is a checklist, not a popularity contest, and closing the gaps is fully in your control.
• The trend over four to six weeks, not the week-to-week line. Look at direction, not jitter.
• The product term and attribute insights, which show the language buyers actually use. That is free research for your question_and_answer copy, pointing you at the questions worth answering next.

The funnel view and the headline share-of-voice figure are useful context, but they move for reasons you cannot always attribute, including competitors editing their own feeds. Let them inform, not steer.

A first pass for a team that cannot rewrite the whole catalogue

Most SME teams reading this run a catalogue too large to rewrite by hand and too small to throw a feed engineer at. You do not need to do the whole thing. Not at once. You need a defensible first pass.

Start with your top 20 products by revenue. These earn the effort, and they give you a clean read on whether the work moves anything before you scale it. For each one, do three things. Add an item_group_title so variants collapse into a family. Write three to five question_and_answer pairs from your real support and returns data, not invented ones. Link any manual or spec sheet you already have as a PDF.

Submit it through a supplemental data source, which is Google's recommended path and keeps the conversational copy separate from your primary feed so you can change it without touching the listings shoppers already see (Google Merchant Center Help). Larger teams can push the same fields through the Merchant API.

Then leave it for a month. Let the attribute completeness score climb on those 20, watch the four-week trend rather than the daily one, and read the product term insights for the next batch of questions to answer. If the top 20 move, you have a repeatable process and a reason to fund it across the catalogue. If they do not, you spent a few hours, not a quarter.

The feed used to be plumbing. It is becoming the place where the first version of your product gets described to a buyer who may never reach your site, and the retailers who write it that way will be the ones the assistants can actually recommend.