Shopify's DotDev talk on AI discovery
What Shopify's product and SEO leads say about AI discovery
This year's DotDev was big on agentic commerce.
I attended a talk on AI discovery by Kate Ragotte, who runs product for Shopify, and Kyle Risley, who leads SEO there. The pitch was blunt - a mythbusting guide on what to ignore, what Shopify already handles, and where real value actually sits.
It was built for Shopify partners, but not all of it is partner-specific. If you run the catalog yourself, the same mechanics apply directly.
One slide summed up the whole thing.
The AI that reads your Shopify catalog can reorganize what you give it. It can't write a fact you never entered. Everything below is what that means mechanically, plus seven checks.
Every Shopify Catalog entry starts the same way. Shopify's ML reads your admin data and builds a structured record out of it, and that record is what gets searched whenever an agent shops across Shopify. Price and stock levels pull from your live store.
Products also get grouped by Universal Product ID, so no matter how many merchants carry the same item, an agent sees one collapsed result. Whether your listing is that result or just feeds into someone else's decides whether queries surface it at all.
Two access points sit on top of that grouped data. The Global Catalog API is built for cross-merchant discovery, comparing options across every seller. The Storefront Catalog MCP is scoped tighter, to shopping inside a single store. Both run on UCP, Universal Commerce Protocol - Shopify's standard for how agents talk to merchants.
I went deeper on how Shopify ranks a listing before an agent ever re-ranks it in
The number Shopify put behind all this: shoppers arriving through Catalog convert at roughly double the rate of shoppers arriving any other way.
Risley's explanation for why tracks with that. AI search collapses discovery and consideration into a single conversation, so by the time a shopper lands on your product page, they're not browsing toward a decision anymore - they've effectively made it.
That's the mechanism. Here are the seven checks that follow from it.
1/ The five fields the model reads first
Title, description, images, options, tags. Everything downstream - filters, comparisons, recommendations gets built from what's already sitting in those five.
2/ Category metafields come before anything else
Taxonomy, Shopify's fixed list - the frame of reference buyers and agents share. Setting the right category opens up the standard attributes for that product, and those attributes are what agents filter on first, ahead of anything written into a description.
3/ Prioritize the attributes buyers actually filter by
Agents filter rather than rank. One empty field on a filtered attribute removes the product from the results entirely, not just from the bottom of them.
Plain-English values, blue, waterproof, linen, need nothing done to them. The model resolves those on its own. Complex ones, something like SKU112, are dead strings it can't parse. Rewrite each as a key:value pair instead, sku: GYS1121.
Options deserve the same scrutiny. Ragotte's guidance on this was that names need to be human-readable, no acronyms or short forms a model wouldn't recognize. Size, color, and material should read the same way on every product in a category.
5/ Descriptions need extractable facts, not adjectives
Skin type, concern, weight, construction, care - whatever your category is specified by.
AI-discoverable shops run descriptions about 37% longer, but the length itself is a side effect. What correlates with visibility is fact density, not word count. Padding a description with adjectives doesn't do anything a shorter, plainer one wouldn't.
6/ Combined listings, even for parents you'll never publish
The same clustering that groups near-identical listings under one Universal Product ID reads your admin structure whether that parent page ever renders on your storefront or not. Choosing not to publish it for SEO reasons doesn't hide the grouping from the model. Skip this step and near-identical listings end up splitting a signal that should have stayed in one place.
7/ None of it counts unless it's public
A field can be filled in the admin and still be invisible on the storefront, hidden behind a theme setting or an unpublished collection.
Agents can't scrape their way to trust. The models enrich and standardize whatever you give them. They don't go looking for what you didn't.
Where to start
Run steps 1 through 7 on your ten best sellers first, not the whole catalog.
Then review your ranking in the Catalog API. Shopify's Agentic Dashboard, the Agentic section inside your Shopify admin, is a good reference point for that. It previews a result set close to the one an agent would actually see for a given query.
Run the five fields above against your own catalog for free.
If you'd rather have this fixed than DIY it, try Atomz on Shopify. Catalog Genius handles the metafield side of this at scale, so it doesn't have to be a manual pass.
- Ankit
If this was useful, the next one will be too.
Weekly. Free.
Unsubscribe with one click.
