When an AI agent shops for someone, it does not show them the internet. It shows them about six products. The interesting question is how it gets from everything on sale to those six.
For Meta's Muse agent, there is now a partial answer, and it is uncomfortable for any brand that is not already famous. A file that reportedly sits inside the agent tells it, in three separate places, to put well-known merchants first.
Where this comes from
Profound, a company that tracks how AI tools describe brands, says its research team pulled an instruction file from Muse's system files on September 25. It shared the text with State of Brand, which published an analysis on October 8. PPC Land reported on both on October 11.
Two caveats before the details. Meta has not commented, nobody outside Profound has verified the file, and Profound itself warns it may have changed since late September. Profound also sells software in this space, so it has a commercial interest in the story. Treat what follows as a credible report, not a confirmed spec.
What the file reportedly says
According to the reporting, the rules include:
- Favor well-known merchants. The file treats this as a priority in the ranking step, the display instructions and the formatting rules. The top five picks should come from well-known sellers. It does not define "well-known".
- Drop price outliers. Listings priced well above or below a category's normal range are discarded. A newer brand that prices low to get noticed can be filtered out by the same rule meant to catch junk.
- Brand-named searches go to the brand. If the shopper names a brand or retailer, that brand's own store goes first, and direct sellers are preferred over resellers by default.
- Three stores, ten products. On the open web, the browser visits up to three merchant sites and collects up to ten products per search. It can stop after two sites if it already has enough.
- Pages must prove they are real. A link has to land on a product page, show stock, show an add-to-cart control and carry a real product image.
What it looks like in practice
Profound also ran a small test, published October 2: 50 prompts in one retail category, through one Muse account, on one day. Profound calls it exploratory, and that is the right word.
In the median prompt, Meta's own product catalog returned about 60 products from 30 merchants. The browser visited about three sites and came back with four products from two of them. The final answer was six product cards: four from the browser, two from the catalog.
Put differently, the catalog supplied about 94% of the candidates and about a third of the slots. Two websites supplied the rest. Being listed in the catalog got a merchant into the pool. Being one of the two or three sites the agent chose to open got it onto the screen.
You can't see it happening
There is a second finding that matters as much as the ranking. Profound reports that Muse does not cite the sites it reads, and browses without naming itself, routing through residential connections so the visit looks like a person's.
So a merchant has no referral line in analytics that says "an agent considered you and passed", and no citation in the answer to check. The file also does not say how the three stores are picked in the first place. The decision that matters most is the one with the least visibility.
The bigger picture
The rule in this file is about sneakers and sofas, but the logic is not specific to retail. Every engine that answers "which one should I buy" has to cut a long list to a short one, and the cheapest signal available is whether it has seen the name before, in enough places, described the same way.
That is a hard rule for a young company, and a useful one to know about. You cannot edit an agent's instructions. You can find out whether you are in the handful of sources it opens when a buyer asks about your category, and whether the pages it lands on say plainly what you sell, what it costs and who it is for. Asking once tells you little, since answers vary from run to run. Asking the same questions repeatedly, and reading which pages get cited, tells you where the gap is.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.