Before an AI agent buys from you, it has to read you

On October 21, Walmart announced it is opening its site to approved AI shopping agents, with spending limits that shoppers can set (TechCrunch). Agents that search, compare and check out for a buyer are moving from pitch toward product: Walmart's Agent Profiles are due to launch within the next month, and its gateway is described as a developer beta (Retail Dive).

Walmart is not part of our panel, and we did not measure it. We also can't tell you how those agents will choose. But any tool that compares stores can only compare what it can read. A price, a shipping rule or a return policy that only appears after a page's scripts run may not be there for a simple automated reader.

So we checked a small panel.

What we did

On October 28 and 29, we made plain automated requests to the public pages of 20 online running stores: one request per page, no browser, no scripts run, no login, nothing bought, no store contacted. Our requests identified themselves as Envoyra research, and we respected each site's robots.txt.

Ten were retailers named in the summaries of three fixed web searches. Ten were independent running shops drawn at random from a published 2026 list of the best running stores. We wrote the rules down and had them checked by an AI agent acting as our statistics checker before we read a page. Four small rule changes were logged afterwards; three were applied to all 20 stores, and the fourth was a re-read of three independents on November 2.

This is a convenience panel in one category. It describes these 20 stores, not online stores in general.

What we counted

Return policies and shipping terms were often not found as plain text. On the 10 independents, a return policy was found as plain text on 4 of 10 stores, on the homepage or a page it linked to (5 of 10 counting the standard policy page, which we also checked for the 5 independents we detected as Shopify). Shipping terms were found on 3 of 10 (4 of 10 counting that standard page). On the rest, the policy wasn't on the pages we fetched as plain text, or it only filled in once scripts ran.

Two basics were in place on almost every independent. None of the 10 blocked the AI crawlers we checked for in robots.txt. (robots.txt is a request to automated crawlers, not a lock; some AI providers say pages fetched because a user asked for them may not follow it.) And 9 of 10 had 300 or more words of text on the homepage without scripts running. Other basics were more mixed: a dollar amount on the homepage on 5 of 10 (on one of those it is a free-shipping threshold, not a product price), and contact or location details on 7 of 10 (on one of those the store email appears only in the page's script data, not in text a visitor sees).

Four of the five agent files were identical. 5 of the 10 independents had a file meant for AI agents at the usual address. 4 of those 5 were the same text apart from the store name. Shopify's developer documentation says it manages an agents.md file by default for every store, and those four were on Shopify. The fifth was an llms.txt file on a store we did not detect as Shopify. One of our five Shopify-detected independents returned no agents.md to our request (we did not look into why), so the default did not show on every Shopify store we checked. We did not test whether any AI agent reads these files.

5 of the 10 retailers refused or did not answer our request. 4 refused it and 1 did not answer. Our request said who we were, which may be part of why. It tells you nothing about how those sites treat AI agents, which we didn't test.

In this convenience panel of 20 running stores. 5 of the 10 retailers refused or did not answer our plain request. Public pages only. Source: Envoyra, October 2026.

With groups this small, differences of a few stores can be chance, so we don't compare the two groups. All 11 checks we ran are in the table at the end.

Readable pages are one thing to check. Another is whether AI assistants name your store when buyers ask. Our free scan asks an AI assistant your buyers' questions and shows, in counts, how often you come up and, after you add your email, who comes up instead. Run a free scan

Our check is a plain request. Some agents may run a full browser and see more than we did; we didn't test any. But a policy that a plain request can't read is one more thing that has to go right before a tool that reads pages this way can use it.

Method and limits

We read the raw text each server returned, without running scripts, and followed up to two links from each homepage for policies. For the 5 independents on Shopify, we also checked its standard policy page, and we report both counts. A page that refused or didn't answer our request counts as "can't tell" for every content check; the page-loads check itself records the refusal. The 5 retailers we could read aren't a random 5; they're the ones that didn't refuse us.

The independents come from an award list, so they are probably better run than a typical store. And our search didn't surface a website for 10 of the 20 independents we drew, so the group describes independents a search can find. Our script found a product page for only 6 of the 20 stores, so we don't report product-page checks as findings.

Detection is simple (a currency amount for price, stock words for availability), so a few results at the edges could be wrong. We checked 4 of the 20 stores by hand, all independents, and re-read the homepages of three independents on November 2. Where the script found no policy, the pages had no footer link to one, or filled it in only with scripts; the standard-page check above found one more that the homepage-and-footer rule missed. Four rule changes were made after registration (how we detect the platform, one extra step to find a product page, counting the agent file as "can't tell" when a homepage didn't load, and, on November 2, re-reading the homepages of three independents our script had not read cleanly: one had come back compressed and unread, and for another the script's own code had been counted as words). The first two were made before we looked at any group differences, and the panel was re-scored under them; the last two came after we had seen results. The November 2 re-read changed the independents' counts in three rows (the word "running" or "runner", shipping terms and return policy), and pages may have changed since October 28 and 29. We didn't test any AI agent and make no claim about how any agent behaves. The word-count result moves with the cutoff: at 150 words the independents are 10 of 10, and at 500 words they are 6 of 10. Seven of those ten counts come from our first script, which could count script code as words (it did on one store), and we could not re-check them, so treat the 500-word figure as approximate.

All 11 checks (Yes / No / Can't tell)

For retailers, the 5 that refused or didn't answer count as "can't tell" in every row except "loads". The two product-page rows are shown for completeness, not as findings. Counting the standard Shopify policy page too: shipping terms 4 of 10 and return policy 5 of 10 for the independents; no change for the retailers.

Envoyra is AI visibility intelligence. We show you what AI assistants say about your business and your competitors. You stay in control of what you change. Start with a free scan. Audit and Monitor then show what to fix next, and Monitor checks the same questions every week.

Envoyra re-tallied the table's counts from its store-level records before publishing.