Product attributes

Product attributes reads how AI shopping answers describe products — the words, specs, and ratings the answer text actually states — and lines your product up against the brands winning the most carousel slots. It answers "what does the AI say about my product, and how does that compare to the shelf leaders?"

What it does

When an AI assistant renders a shopping carousel, its answer text often describes the products — "lightweight aluminium frame", "weighs 289 g", "rated 4.5 out of 5". Product attributes captures exactly those stated descriptions and sorts them into three kinds:

• Characteristics — descriptive text properties like material, style, or fit.
• Facts — checkable specs: numbers with their stated unit (including price with its currency), and stated yes/no properties like "fully waterproof".
• Ratings — ordinal scores the answers state with digits, such as 4.5/5, 9 out of 10, or 88%.

It then shows your product's attributes side by side with the brands that won the most carousel slots in your window, so you can see where the AI describes you differently from the shelf leaders.

Who it's for

• Product and merchandising teams who want to know how AI assistants characterise their products versus competitors.
• Content and PR teams looking for the specs and claims AI answers repeat — the ones worth reinforcing on your own pages and in your knowledge base.

Try it

1. Open Shopping and go to a product detail page (/shopping/products/[id]) for a product that has appeared in at least one carousel. The attributes section only shows for products the AI has actually placed.
2. Find the section headed "How AI describes this product".
3. Switch between the Characteristics, Facts, and Ratings views.
4. Read the matrix: the Attribute rows down the side, your product in the highlighted This product column, and the top competitor brands in the columns beside it. Each column is labelled with how many carousel slots that brand won in the window.
5. Hover any value to see the verbatim quote it came from, and follow the link to the real answer receipt where one is still live.

While extraction is rolling out, the same section instead shows a short status note (see below) — the surface is there, waiting on the pass.

What you see while extraction is rolling out

The attribute-extraction pass is off on production today, so most workspaces see one of these honest states in place of the matrix:

• Not enabled yet — "Attribute extraction isn't enabled yet." The copy explains that shopping answers are being collected, but the flag-gated extraction pass hasn't run for your workspace, and that once enabled every shopping answer is scanned for text-stated attributes — only what the answers literally say, with the quote kept as evidence, never inferred.
• Window not processed yet — once extraction is on, a window whose answers haven't been through the pass shows "This window hasn't been processed yet."
• Nothing to extract — if the carousels listed products without stating any attributes, you see "No stated attributes found." Nothing is invented to fill the table.

How it's computed

• Only what the text states. Every candidate attribute is re-checked against the answer text before it is stored. The evidence must be a verbatim quote; a paraphrase or synonym is rejected ("featherweight" does not ground "lightweight"), numbers must carry their stated unit with no conversion ("0.6 lb" cannot ground "272 g"), and ratings must be written with digits. Under-extraction beats fabrication — an empty result is a valid answer.
• Sample sizes, always. Aggregates carry their denominator: a characteristic reads "8/12 descriptions", never a bare "67%". The denominator is the number of distinct answers describing that brand; the numerator is the answers stating that value.
• No fake math across units. Facts are pooled only within one stated unit; the dominant unit's range is shown and every other unit is listed as excluded — marked "— not pooled" — rather than converted. A 9/10 rating is never rewritten as 4.5/5.
• Missing is an em-dash. A cell with no data shows an em-dash whose tooltip says why.
• The comparison columns are the top competitor brands by carousel-slot count in your window — the brands winning shelf space — with your product always first. This is a side-by-side view of what the answers say, not a statistical claim that any attribute causes a win.
• The rollout flag. Extraction is controlled by the ATTRIBUTE_EXTRACTION flag, off by default because it spends language-model tokens per shopping answer. The history backfill is dry-run unless explicitly told to apply.

Limits

• Extraction is rolling out and off on production today; the comparison view shows the honest states above until it is enabled.
• Sources follow shopping's engine coverage — ChatGPT, Google AI Overviews, Google AI Mode, and Copilot collect today; see Engine coverage.
• Visibility is scoped to your workspace — you only ever see your own brand's data, and competitor columns are drawn only from brands reachable through your own placements. There is no separate plan gate on the feature.

Related

• Product detail — the page this section lives on.
• Shopping analytics — who wins the shelf overall.
• Competitors — how competitor brands are identified and attributed.
• Data honesty — the sample-size and em-dash rules this feature follows.