• axidort
    link
    fedilink
    English
    arrow-up
    1
    ·
    10 hours ago

    Building on the point above about 200 vs. adoption: even if the endpoint check is sound, it is a few steps removed from “can an agent buy something here”.

    An answering MCP route tells you the platform ships the route. Whether it is useful depends on things a 40-domain HTTP probe cannot see: is the catalog actually populated and current, can the agent get from a product to a cart to a checkout without a human-only step (captcha, login wall, a payment app that needs a browser redirect), and do shipping, tax and stock come back accurately? A store can pass the endpoint test and still fail at the first cart call.

    Sample caveats worth keeping in mind: 40 domains split across several business types (SaaS, D2C, agencies, local services) leaves only a handful per group, so the percentages are coarse, and a platform default that is on for everyone could make “80%” mostly a statement about one vendor’s rollout. If the domains were picked rather than sampled at random, they may also skew toward sites that already care about this.

    If you want to check your own store, a more telling test than the existence check:

    1. Fetch the endpoint and confirm it returns valid JSON that describes tools, not just a 200.
    2. Run a real task through it: search for a product, add a variant to a cart, get to a checkout URL. Note where it stops.
    3. Compare robots.txt and your CDN/WAF rules with what the bots really get. Bot protection that returns a 403 or a challenge page to known AI user agents is probably the most common silent failure, and it never shows up in your own browser.
    4. Check that product pages carry structured data (price, availability, variants) that matches what is visible on the page, since agents that don’t use MCP still read that.

    The absence of /agents.md is the least surprising result, since it is a proposal rather than something any major agent is documented to look for.