Czech commerce in September 2026: no agentic checkout declared, one MCP server, no AI crawler blocked

We probed the public surface of 22 large Czech retail, travel, telecom and ticketing sites for what an AI shopping agent would find. Seventeen answered. None declares a UCP checkout. One, Rohlik, publishes an MCP server. None blocks a single AI crawler. The month-one baseline of a series.

By the Hlido Editor · 2026-09-04

Shopping agents are beginning to buy on a person's behalf, and Shoptet, the platform behind roughly a third of Czech and Slovak e-shops, runs an MCP server and is preparing the UCP and ACP checkout protocols. So we asked a narrow question about the largest Czech commerce sites: what does an AI agent find when it arrives at the front door today?

Hlido is an independent review desk. No site on this panel paid to be here, none was told, and none can pay to change what follows.

What we probed

On 4 September 2026 we fetched the public surface of 22 Czech sites: eight retailers, five travel sellers, three carriers, four ticket sellers, and the marketing sites of the two dominant e-shop platforms. Five things, no account, no purchase, no chat opened:

  • Reachable by agents. Does robots.txt let GPTBot, ClaudeBot, Google-Extended, PerplexityBot, OAI-SearchBot, anthropic-ai and Applebot-Extended read the site?
  • Agentic checkout declared. Is there a /.well-known/ucp manifest, and does it declare cart, checkout or order capabilities? Presence and declared capabilities only, never spec conformance.
  • Discovery files. A real llms.txt (plain text, with the H1 the spec requires), a sitemap, and any MCP server named on the public surface.
  • Catalog on the homepage. Product or Offer structured data.
  • Chat widget disclosure. Where a chat widget is detected, whether the page says it is an AI before the first message, read with the same probe as our Article 50 register.

The full method, every row and every limitation are in the open data file (CC-BY 4.0). The same probe runs for any store, free, at hlido.eu/eshop-check (česky).

The numbers

Seventeen of 22 sites answered a plain HTTP fetch. Five did not: alza.cz, mall.cz, notino.cz, slevomat.cz, ticketmaster.cz. Their homepages refused a non-browser client from a GitHub-hosted runner, which usually means bot protection. We draw no conclusion about them: a site that refuses our fetch may still admit verified agents, and a site that admits us may still refuse a purchase.

Of the 17 that answered:

SignalSitesOf 17
Declare a UCP manifest at /.well-known/ucp00%
Declare an agentic checkout, cart or order capability00%
Block any of the seven AI crawlers in robots.txt00%
Serve a real llms.txt529%
Name an MCP server on the public surface16%
Product or Offer structured data on the homepage00%
Show a chat widget318%
…of which disclose an AI before the first message0—

What that means

Nobody has declared an agentic checkout yet. Not one of the 17 reachable sites publishes a UCP manifest, including the marketing sites of Shoptet and Upgates themselves. Given that Shoptet has said UCP and ACP are in preparation, this is a baseline, not a verdict. It is the number we will re-read on the third of every month.

The doors are open. None of the 17 blocks a single AI crawler. So the constraint on agentic commerce in the Czech market is not access. It is the absence of anything machine-readable to act on once inside.

Rohlik is the one site that speaks to agents on purpose. Its llms.txt is written for AI agents in Czech, names use cases and limits, and points to https://mcp.rohlik.cz/mcp. We spot-checked that file by hand. We did not connect to the server; declaring one and serving one are different claims, and this probe only tests the first. Four others, Bonami, Košík, O2 and T-Mobile, serve a real llms.txt without naming an endpoint.

Homepages carry no product schema. That is expected and weak evidence: Product and Offer markup lives on product pages, which this month's probe did not open. We report it because it was measured, and we will extend the probe to one product page per site before reading anything into it.

Three chat widgets, no visible AI disclosure. Rohlik and RegioJet run Daktela, SMSticket runs Zendesk. On none of the three homepages does visible text say a visitor will be talking to an AI. This is an observation, not a finding: a widget may be human-staffed, in which case Article 50(1) does not apply, and a disclosure that appears only after the chat opens is invisible to a public-surface read. Since 2 August 2026 the duty applies to any AI chatbot with no grace period, so a merchant who runs one should be able to point to where the disclosure sits.

By sector

SectorProbedAnsweredUCPllms.txtMCP namedChat widget
Retail850311
Travel540001
Telecom330200
Ticketing430001
Platform vendor sites220000

What we did not measure

We did not open a product page, add anything to a cart, open a chat, or connect to any MCP endpoint. We did not test whether a shopping agent can complete a purchase, and nothing here is a legal finding or a conformity assessment. Five sites could not be fetched and are excluded from every percentage rather than counted as closed. The panel is the largest public-facing sites per sector, not a market-share ranking; it will grow.

Where a signal is absent from the file, we did not measure it. Where it is present, it was measured on the date shown, by the probe whose source ships in the same repository as this page.

If you run one of these sites

Paste your address into the free check and you will see the same rows. Fixing the two cheapest gaps takes an afternoon: a real llms.txt with an H1, and a first-message line in your chat widget saying it is an AI, if it is one. Declaring a checkout is a larger decision, and it is the one the market has not made yet.

Next reading: 3 October 2026.

How this page was produced. The numbers come from automated probing of public surfaces on the date shown. The written analysis is drafted by an AI system, and posts publish without a person reviewing each one. Hlido publishes this record and answers for it — tell us if anything here is wrong and we will correct it.