scan · 40 storefronts · 2026-08-16

Can AI shopping assistants read your store?

Since May 2026 Shopify generates /agents.md, /llms.txt and /.well-known/ucp for every store — the files AI assistants read to understand a catalogue. I checked 40 DTC storefronts on the domain customers actually visit.

GET /agents.md — primary domain 0 / 40
0 57%
Serving agents.md. Shopify handles it; nothing to do.
0 32%
No discovery surface at all — no agents.md, no UCP. Nearly all headless.
0 10%
Shopify-connected — UCP and MCP live — and agents.md still 404s.

Check your own store

Type a domain and this builds the two commands to run. Nothing is sent anywhere — the page has no back end and makes no requests. 


    

    

      
200 · 200Fine. Shopify is handling it.

      
404 · 200Connected, but the file isn’t reaching your domain.

      
404 · 404Usually a headless front end with no route for them.

    

  


  

    
Why redacted

    

      

        The four with a genuine finding aren’t named. They haven’t been told privately, and
        publishing a list of companies as broken is a different act from quietly telling them.
      

      

        The thirteen above are named because being headless isn’t a defect — it’s an architecture
        choice with a side effect almost nobody chose deliberately.
      

    

  


  

    
The surprise

    

      

        A third of these storefronts serve no agent-discovery surface whatsoever on the domain
        everything links to. The custom domain is fronted by something with no route for those
        files, so nothing Shopify generates ever reaches the address a crawler visits.
      

      

        Nobody opted out. They went headless for good reasons and never got agent discovery in the
        first place.
      

    

  


  

    
Method

    

      

        Fetched on the primary domain, never myshopify.com — the host shoppers
        and crawlers actually reach. Redirects followed only within the same site: an apex-to-www
        301 is not a missing file, and treating it as one produced five false positives
        before I caught it. A 302 to /password was recorded as gated and
        unknowable, not as absent.
      

      

        Sample: 40 well-known DTC brands plus the published client lists of six Shopify agencies.
        Convenience sample, not random — indicative, not a census.
      

    

  


  

    
What it isn’t

    

      

        Shopify Catalog syndication is API-based and unaffected. A store with no
        agents.md still reaches ChatGPT, Copilot and Google AI Mode that way. What’s
        lost is the engines that crawl for the files directly.
      

      

        And there is no good published evidence that structured data causes AI citations —
        the strongest study to date found effects indistinguishable from noise. This measures
        whether products are eligible and readable, which is verifiable. It predicts nothing about
        whether an engine will recommend them.
      

    

  


  

    
What I build

    

      

        Feedwright audits a Shopify catalogue for the reasons AI shopping surfaces
        exclude or downrank products — invalid GTINs, unassigned taxonomy categories, unpublished
        products, missing shipping weights, discovery files — then fixes what it safely can, in
        bulk, with a full undo. Every change records the value that was actually in place when it
        ran, so undo restores that value rather than a guess.
      

      

        Not on the App Store yet. Want the audit run on your store before then, or want to argue
        with the numbers above? chamika@feedwright.io