Merchant Center conversational attributes: the six fields for AI Mode

What conversational attributes are

Google describes conversational attributes as a way to help AI systems and conversational agents understand the specific nuances of a product. They sit inside the normal Merchant Center product data specification, so they travel in the same data sources as price and availability, but they carry the kind of detail a shopper asks about in a chat rather than filters on in a grid.

Three facts from Google’s help page set the frame:

• They are completely optional, and can be used in every country.
• Including them does not change the approval status of existing products.
• If a detail is already in description, product_highlight or product_detail, it does not need to be repeated here.

Google names AI Mode in Search as one of the surfaces that use them. It does not publish a measured effect on visibility, and this guide does not claim one.

These fields answer the questions a shopper would otherwise ask a salesperson, in a form an AI surface can quote.

The six attributes and their limits

The six relationship types for related_product are part_of_set, required_part, often_bought_with, substitute, different_brand and accessory. The identifier type is either gtin or id, the latter meaning the product’s ID in your own Merchant Center data.

How to write each value

In a text file, the group attributes pack their parts into one cell. These are the formats Google shows:

• question_and_answer: each pair as "question":"answer", pairs separated by commas, for example "Is it machine washable?":"Yes, on a cold, gentle cycle."
• related_product: type:identifier_type:identifier, entries separated by commas, for example required_part:id:AZ7B,accessory:gtin:811571013579
• variant_option: name:value pairs separated by commas, for example color:green,size:S
• document_link: one or more PDF URLs separated by commas.

Google recommends TSV over CSV for these fields, which also spares quoting every comma inside an answer. In an XML feed, each question and answer pair is its own element:

<g:question_and_answer>
  <g:question>Is it machine washable?</g:question>
  <g:answer>Yes, on a cold, gentle cycle.</g:answer>
</g:question_and_answer>

The product in these examples is a placeholder: the linen dress from the XML feed example, whose variants already share an item_group_id.

item_group_title and variant_option together

Most feeds already group variants with item_group_id. These two attributes describe that group in words. Google’s own example is a phone: the variant titled “Google Pixel 9 Pro 512GB Moonstone” carries the group title “Google Pixel 9” and the variant option display:XL,memory:512GB,color:moonstone.

Google’s rules for the pair:

• Every variant in a group carries the same item_group_title, and it should differ from the variant titles.
• Every variant uses the same set of option names, and no two variants share the same combination of values.
• The values must match what the landing page shows for that variant.

If you already send color and size, the variant option restates them in one place and adds properties that have no attribute of their own, such as a shoe width or a memory size.

How to send them without touching the main feed

Google offers three routes: a supplemental data source, the primary data source, or the Merchant API. It recommends the first. A supplemental data source is a second file keyed on the product id that adds or replaces fields, so a team can maintain questions and answers in a spreadsheet while the main feed keeps coming from the store platform.

The same mechanism is how category overrides work, and the supplemental feed guide walks through linking one to a primary source. A single supplemental file can carry a google_product_category column and a question_and_answer column side by side.

Where good questions come from

The useful questions are the ones shoppers already ask: in support tickets, in pre-sale chat, and in reviews, where a complaint often reveals a question nobody answered before the purchase. “Runs small” is a sizing question; “did not fit my model” is a compatibility question.

The Review Analyzer pulls topics and recurring complaints out of a large set of reviews, which is one way to find them across a catalog. Whatever the source, each answer should be something the product page also supports, for the same reason the feed and the page have to agree on price.