Amazon Blocks the AI Shoppers. Your Store Probably Doesn't

On 9 September 2026 the robots.txt file at amazon.com carried a site-wide Disallow for every major AI crawler and shopping agent: GPTBot, OAI-SearchBot and ChatGPT-User, ClaudeBot, Claude-User and Claude-SearchBot, PerplexityBot and Perplexity-User, Google-Extended, GoogleAgent-Mariner and GoogleAgent-Shopping, meta-externalagent, CCBot and Bytespider. Checked the same day, Shopify's own site carried no such rule, and neither did the robots.txt of two large Shopify-hosted stores. When an assistant answers a Black Friday product question this season, it composes that answer from the sites that let it read. That is a fact about access. It is not a number, and the people attaching numbers to it are guessing.

This is a measurement site, so the useful version of this story is not the strategy gossip. It is what you can check, what you cannot, and which of the two decides your Black Friday reporting.

The file, and how to read it yourself

Robots.txt is a plain text file at the root of a domain that names crawlers and tells them which paths they may fetch. It is a stated preference rather than a lock, and the major AI companies publish the agent names they use so that publishers can express one. Fetch any store's file in a browser by adding /robots.txt to the domain.

Three details in Amazon's file are worth more than the headline. The first is that the block is site-wide: Disallow followed by a single slash, not a carve-out for checkout or account pages. The second is that it covers the user-triggered agents, not only the training crawlers. ChatGPT-User, Claude-User and Perplexity-User are the agents that fetch a page because a person asked the assistant to look at it, so the rule reaches the case where a shopper explicitly asks about a listing. The third is that two of the blocked names, GoogleAgent-Mariner and GoogleAgent-Shopping, are agents rather than search crawlers, which tells you the category exists and is being addressed deliberately.

What the file does not tell you

It does not tell you how much traffic assistants send, how many orders they influence, or whether either number is growing. Nobody can derive that from a robots.txt, and the version of this story circulating with large round figures attached is not checkable from the outside. The access question has a public answer. The volume question has a private one, and it lives in your own data.

That distinction is the whole point. A strategy narrative you cannot verify is worth less than a coverage rate you can compute in an hour.

Your store is probably readable already, and that is the easy half

The Shopify sites checked carried no agent rules, which means the default posture for a Shopify-hosted store is open. If you have not deliberately added blocks, an assistant that wants to read your product pages can. Readability is likely not your problem.

Attribution is. Referrals from assistants frequently arrive with no resolvable source, so GA4 files them under Direct or Unassigned along with everything else it could not resolve. The channel does not show up as a channel. It shows up as a hole. When AI sends the shopper, who gets the credit covers the mechanics, measuring the sales ChatGPT and Perplexity drive covers what is and is not observable, and who gets credit when an AI agent buys covers the case where the buyer is software.

Why the peak makes it worse

Coverage is attributed conversions divided by the orders your store actually shipped, and every cost figure you quote is the real figure divided by that number. The Price of Being Found reports its publisher's own census at 48.6% of sessions with no resolvable source over 17 October 2025 to 2 September 2026, one company's data rather than a population estimate. Black Friday pushes the unresolvable share up, because paid traffic, app browsers and consent declines all spike at once, and because promotional discovery routes through more hands than a normal Tuesday does.

So the number moves in the worst possible week, and it moves the denominator under every cost per order in your deck. The number nobody checks has the query.

The timing problem is simple and it is the only genuine deadline in this piece. If you compute your Direct and Unassigned share for the first time on 1 December, you will have a number and nothing to compare it against. Computed this week, the same number becomes a baseline, and the Black Friday reading becomes a change rather than a curiosity. Black Friday 2026 falls on 27 November, which is 79 days from today.

What to do this week

• If you own the budget: open your store's robots.txt and confirm you are not blocking the agents by accident, then compute your coverage rate for the last 40 days and write it down with the date. Ten minutes for the first, an hour for the second.
• If you have to defend the number: put the coverage rate on the same slide as the channel table, before the peak rather than after it, so the December conversation is about a movement you predicted rather than a gap you discovered.

A causal read states its coverage instead of presenting a total that may be half visible, which is the property that matters when a growing share of demand arrives unlabelled. The interactive demo shows that output on a sample store with no signup, and what we publish for machines is the same posture applied to this site.

Robots.txt observations are from amazon.com/robots.txt, shopify.com/robots.txt, allbirds.com/robots.txt and gymshark.com/robots.txt, each fetched on 9 September 2026, and any of them may change without notice. The coverage census (5,858 of 12,054 sessions) is from The Price of Being Found (Edition 2.10), Chapter 8, rated Supported rather than Established by the book's own appendix because the underlying data is not independently reviewable.