What Meta's Muse recommends: first data from 3 small sites

On November 1 2026 we started asking Muse, Meta’s AI assistant, the same buyer questions we ask ChatGPT, Gemini, Claude and Perplexity. Three small sites, five questions each. One of them, a home dumpling business, was named in 3 of 5 answers on day one, and ranked first in two. The other two were named in none.

The pattern was simple. Muse answered from the open web. When it found a small business’s own website, it read it and quoted it. When the answer came from pages that never mention the business, the business was not in it.

What we ran

• The model: muse-spark-1.3 through Meta’s developer API, with its web_search tool on. The model decides when to search. It returns its sources as citations.
• The questions: 5 buyer questions per site, the same daily set the other four assistants get. 15 answers in all.
• Follow-up probes: 9 extra questions on the same day, to check how Muse handled the businesses by name.
• The cost: about $1.54 for the day.

The three sites:

1. Daily Dumplings (dailydumplings.com), a one-person frozen dumpling business in Hollywood, Florida. It is also our first case study.
2. A one-person sticker shop on Etsy, with its own small website. It is mine, which is why it is in the test.
3. Heard of You (heardofyou.io), this site.

Results by site

Daily Dumplings: named when the question matched its site

When Muse won, it was because of the website. It opened dailydumplings.com and repeated the delivery zones, prices and pickup times correctly. Asked about the business by name, it got every fact right.

The two misses are telling. The Fort Lauderdale and khinkali answers were built from delivery apps and restaurant guides, where a home kitchen does not appear. The site sells khinkali, but it had no page about them, so Muse treated the question as a restaurant question.

The sticker shop: quoted, but credited to Etsy

This was the strangest result. Muse quoted three of the shop’s own listings word for word, prices included. It found them on Etsy’s search pages and gave the credit to “Etsy”.

A probe confirmed it. Asked who sells one of those exact sticker sheets, Muse named the shop correctly. So Muse knows the shop exists. It just never sees the shop’s name while answering a category question, because the listing titles do not contain it.

Muse never opened the shop’s own website for a category question.

Heard of You: not in the pages Muse read

Asked about Heard of You by name, Muse read heardofyou.io and summarised it correctly. Asked buyer questions, it named other vendors, found through their own pricing pages and through “best AI visibility tools” lists. It quoted prices word for word.

Three of our five questions were “how do I get recommended” questions. Muse answered those with a list of steps and named no vendor at all, ours or anyone’s.

What Muse searched

Across the day, Muse made 104 web searches, opened 56 pages and saw about 1,000 search results.

• Daily Dumplings questions: 390 results. The most common sources were dailydumplings.com (26), guide.in.ua, a Ukrainian diaspora directory (24), DoorDash (23) and Postmates (15). Then ordering pages, MapQuest, Wanderlog, restaurant directories and food guides.
• Sticker questions: 134 results, almost all marketplaces: Etsy (38), Redbubble (21), TeePublic (11).
• Heard of You questions: 335 results from 238 different sites, spread thin across vendor blogs, Medium and dev.to.

What was missing matters too. In about 1,000 results we saw no Reddit, almost no Yelp, no Google Maps and no Facebook.

When we asked about Daily Dumplings’ Instagram, Muse opened the public Instagram page like any other web page and said it could not see much without logging in. It did not pull anything from inside Meta’s apps.

What Meta says about its crawlers

Meta’s crawler documentation lists several user agents. The ones that matter here:

All three of our sites allow every crawler (User-agent: * with Allow), so meta-webindexer was already let in.

Where the search results come from is less clear. Meta has not named a search provider for 2026. In October 2024, Meta AI used Google and Bing, and Meta was reported to be building its own crawler and index. Since December 2025, paid deals with publishers supply news.

Our results looked more like a separate index than Google’s: lots of low-quality pages that Google usually filters out, and only one of about 1,000 result links carried a Google shopping tag. That is an inference from one day, not proof.

The app can do more than what we measured

We measured the API. The Muse app, launched in September 2026, can also draw on accounts you follow on Instagram, Facebook and Threads. It can shop through partner retailers and book tables. None of that shows in our numbers.

So a Meta presence still matters for real app users. It just did not decide anything in the answers we could see.

What we are changing

Only things that are true and that we can do on our own sites:

• Daily Dumplings now has a khinkali page and a Fort Lauderdale and Broward delivery page, in English, Russian and Spanish, built only from facts already on the site.
• Heard of You rewrote its pricing page so the prices and the words buyers use (“one-time”, “done-for-you”, “for solo founders”) sit in plain sentences. We also published an honest Peec AI alternatives comparison.
• The sticker shop is considering putting its name in a few Etsy listing titles. If it does, we will check after two weeks whether Etsy views drop.

No fake reviews, no paid placements dressed up as editorial, no new accounts pretending to be customers.

Limits

• One day: 15 answers and 9 probes. Muse’s answers change from run to run, so these are early signals, not rates.
• Three sites, all run by me or my family. That is why we can share every number, and also why it is a small sample.
• The API model only. The consumer app may answer differently.
• Nothing here says why Muse chose a source, only which sources it chose.

We will re-run the same questions daily and publish the 30-day picture in December. The data on this page is free to reuse under CC BY 4.0 with a link to this page.

Questions