Vinted's static agent-readiness score is 82 / 100. Its robots.txt welcomes every AI agent — search, agent, and training intents are all allowed. An honest agent-identity fetch is not WAF-blocked. By every declared signal, this is a site that's ready for automated visitors.
Then we pointed a real AI agent at it and asked it to shop. Here's the gap.
What the static scan said
On a scan in late August 2026, vinted.nl scored 82.3 overall. Category breakdown: Discoverability 82, Structured 69, Consistency 100, Actionability 84, Trust 72, UCP 86. Citability is not scored for commerce sites — it applies to content-type sites only, and the remaining six weights are renormalized.
The findings an agent cares about (every check is documented in the public rule catalog):
- High — Missing Product schema. No machine-readable
Product/Offerdata — no structured price, availability, or currency. - Medium — No agent commerce surface. No UCP profile and no MCP discovery: no declared, verifiable way for an agent to transact.
- Medium — Machine-invisible trust facts. Shipping and return terms aren't exposed in a machine-readable form.
- Low/info discovery hygiene: no sitemap, no llms.txt, no agent-interface discovery surface.
Read on its own, that's a good-but-improvable report. What it can't tell you is whether an agent can finish a purchase. So we ran one.
You can run the same static scan on your own domain without an account — one free check per visitor per day, up to 10 pages — from the box on scovant.com.
What the agent actually did
We gave Claude Sonnet 4.6 a real browser and one goal: find an item and get to checkout. It behaved like a competent shopper:
- Navigated to vinted.nl, dismissed the consent layer, reached the catalog.
- Searched for "shirt", got results.
- Clicked a product to open its detail page — and landed on the registration page.
- Tried a different product. Same redirect. Tried a price link, tried favorites — every path to an item bounced to registration.
- After thirteen steps the agent stopped and reported, in its own words:
"Vinted requires user registration/login to view product details or proceed to checkout. Every click on a product redirects to the registration page. This is a login wall that affects all unauthenticated users — cannot complete the task without an account."
Two things make this a clean finding. First, it was not a WAF or bot challenge — the anti-bot signal was negative. This is the site's own product design. Second, the wall is honest about who it stops: every logged-out visitor, human or agent. Scovant never creates accounts or enters credentials by design, so it stops exactly where a privacy-minded shopper who won't sign up just to browse would stop.
Outcome: the run failed at an interaction step, and the owner of that failure is the website — not the agent, and not anti-bot.
Why the two layers disagree — and why that's the point
The 82 is not wrong. It measures the site's declared, machine-facing quality, and on those axes Vinted is decent. But an agent doesn't grade markup — it tries to do the task. The static scan flagged missing Product schema; the simulation revealed the agent never even reaches a product page to need it, because browsing itself is gated. Neither layer alone tells the whole story. Together they do:
- Static: if an agent got to a product, it couldn't read the price.
- Simulation: an agent can't get to a product without an account.
That is the entire case for running real simulations instead of trusting a score.
The fix list
- Let unauthenticated visitors see products. A public product page — even a read-only one — is the difference between an agent-invisible catalog and an agent-shoppable one. Gating the whole browse experience behind registration excludes every agent and every logged-out human.
- Publish
Product/Offerstructured data on those pages: price, currency, availability. Moves the highest-severity static finding and gives agents a machine-readable basis to act. - Declare an agent commerce surface — a UCP profile, or at minimum MCP discovery — so agents transact through a contract rather than screen-driving the human UI.
- Machine-legible shipping/returns plus discovery hygiene (sitemap, llms.txt): cheap, and they remove friction for every crawler and agent.
The honest takeaway
An 82 that fails a real checkout isn't a bad score — it's a misleading one, if you read it as "ready for agents." The registration wall is a legitimate product choice for humans; it's also a hard stop for the growing class of shopping agents, and no static score will ever surface it. Declared vs done. That gap is exactly what a simulation is for.
Scovant scanned vinted.nl as part of its public agent-readiness research corpus. The site is not affiliated with or endorsed by Scovant; removal requests are honored. See the research methodology and the checkout benchmark.