How does Meta’s Muse agent use your website?
It fetches the pages people give it and tells them what they say. Muse does this from Meta’s servers, mostly as HatchFeed/1.0. It does not run your JavaScript, it does not read robots.txt, and it rarely opens a second page. We found about 15,600 Muse requests on six websites that Tin Analytics monitors. Here is what they show.
What we foundin short.
- Found: Muse fetches pages under three names, all from one Cloudflare IPv6 range. One of them is a HEAD request that claims to be Chrome 138 on Linux.
- Found: it opens links people give it, not pages it discovers. It never got a 404 and never asked for robots.txt.
- Found: it does not show up in Google Analytics or PostHog. It does show up in server logs, as Chrome.
- Found: the ways to verify it are unreliable. Its signatures came and went, and Cloudflare stopped labelling it a verified AI assistant on 7 November.
- Not seen: Muse’s cloud browser filling in forms or shopping, and its new Chrome extension. Either they did not visit these sites or they look like ordinary people.
Want to know what agents are doing on your site? Book a free 30-minute call with Emre →
What Musedoes for its users.
Muse is Meta’s personal AI agent. It launched on 8 September in the US on iOS, Android and muse.ai. People ask it to research, compare, book and buy. Meta says it runs a Chromium browser on an isolated Linux machine for each user and can fill in forms and check out.
Meta’s crawler documentation lists its crawlers and fetchers. Muse is not among them. Meta publishes no Muse user agent, no robots.txt token and no list of addresses. Security companies have reported that Muse does not identify itself. They measured it on large sites. We looked at what it does on small ones.
Meta is also clear that a Muse visit counts as a person’s visit: When Muse browses the internet, it will appear as your activity
. Muse’s own robots.txt forbids automated collection of its pages without written permission.
What a Muse fetch looks likein your logs.
Every link Muse fetches gets the same three requests, in the same order:
| When | Method | User agent | Signed |
|---|---|---|---|
| 0 s | HEAD | Mozilla/5.0 (X11; Linux x86_64) … Chrome/138.0.0.0 Safari/537.36 | No |
| about 10 s | GET | meta-webindexer/1.1 (+https://developers.facebook.com/docs/sharing/webmasters/crawler) | Yes, for www.meta.com |
| under 2 s later | GET | HatchFeed/1.0 (+https://agent.meta.ai; link preview) | No |
Then HatchFeed asks for the page’s /opengraph-image and /twitter-image. Muse builds a preview card for each link, with the title and image, not just the text.
All three come from Cloudflare’s network (AS13335), from addresses in 2606:54c1:1::/48. The address changes on almost every request. Over 6 and 10 November this range sent 7,162 Muse requests from about 5,000 different addresses, and only 10 requests of any other kind.
agent.meta.ai in the HatchFeed string redirects to muse.ai. The Muse iPhone app opens links in a browser that calls itself FBAN/HatchForiOS. Hatch looks like the app’s internal name.
The signed request explains reports of meta-webindexer from addresses Meta does not claim. Meta’s own indexer comes from Meta’s network. This one comes from Cloudflare, signed with Meta’s key, a few seconds before HatchFeed.
| Muse fetcher | Requests | Sites |
|---|---|---|
| HatchFeed/1.0 | 7,287 | 4 |
| meta-webindexer/1.1, signed for www.meta.com | 2,613 | 4 |
| HEAD with a Chrome/138 Linux user agent | 5,714 | 6 |
These counts cover 9 October to 13 November on the 11 websites that send request records to Tin Analytics.
It’s not curious.
None of Muse’s 15,600 requests got a 404. It does not guess addresses or explore. 85% to 97% of what it fetched were articles, and 5% were home pages. It never fetched a pricing, sign-up or login page. Muse opens a page because a person sent it there, and the pages people send are mostly articles.
Most tasks are one page. The few that went further were short and fast, and the person behind them shows through:
- A daily errand. Muse fetched the home page of a games catalogue at 16:08 UTC on 29, 30 and 31 October. From 2 November it came at 17:08. US clocks went back on 1 November. Someone in the US scheduled Muse to check that page every morning.
- A setup question. On a consumer messaging app, Muse read
/llms.txt, the docs and two integration guides in 9 seconds. Someone had asked it how to connect the product to their tools. - Shopping around. On the games catalogue it went from the home page to the Xbox 360 section and four game pages in 87 seconds.
- A repeat reader. It fetched our aocr.org investigation 12 times over 9 days.
Muse reads it.People do not come.
On a typical day Muse made about 50 requests across all our sites. On 6 and 7 November it made nearly 7,000 a day. Almost all of them were for one page: our post on Claude Code and email addresses, which a popular GitHub issue had just linked to.
| Window (UTC) | Signed Muse fetches | Pageviews by people |
|---|---|---|
| 16 hours before (00:00–16:00, 6 Nov) | 6 | 219 |
| 24 hours after (16:00 6 Nov – 16:00 7 Nov) | 2,428 | 223 |
The Muse fetches began as the people’s visits were fading, rose to 200 an hour by the evening and faded over a day. Each fetch means Muse needed the page for one of its users: to preview it, summarize it or answer a question about it. The people who asked did not come to read it. In two months, Muse’s in-app browser made 4 requests to our sites, all of them for preview images.
This is what an agent audience looks like. Muse fetched the article ten times as often as people opened it, and none of that shows up in the site’s analytics.
robots.txtand llms.txt.
Muse never asked for robots.txt. It touched llms.txt 13 times, always in the middle of a task. Other agents on the same sites use llms.txt more:
| Agent | robots.txt | llms.txt | All requests |
|---|---|---|---|
| Muse | 0 | 13 (3 reads, 10 HEAD checks) | about 15,600 |
| ChatGPT-User | 13 | 41 | 4,381 |
| Claude Code | 0 | 41 | 1,568 |
If you write an llms.txt, ChatGPT and coding agents will read it. Muse mostly will not.
Your analyticsdoes not see it.
Muse’s fetchers never ran JavaScript: 0 times in 15,600 requests. Google Analytics, PostHog and every other script-based tool never see them. If your Direct traffic went up this autumn, Muse’s fetchers did not cause it.
Tools that count server or CDN logs do see them. The HEAD request claims to be Chrome, so a log-based dashboard counts it as a Chrome visit. On tin.computer it was 7% of the requests that looked like browsers, nearly all of them during one burst.
Two other things from Meta are easy to confuse with Muse:
- Meta’s indexer with a Chrome 145 user agent. It comes from Meta’s own network and adds
compatible; meta-webindexer/1.1to a Chrome string. It ran JavaScript on 0 of 1,901 pages. Its client hints report Chrome 157, so the version in its user agent is frozen. - A link checker that does run JavaScript. It comes from Meta’s network with a Facebook referrer, a 2000×2000 screen, a Los Angeles time zone and old browser versions such as Firefox 59. It opens links that were just shared on Facebook. It shows up in analytics as a one-page visit from facebook.com.
On the consumer messaging app, 33% of Direct sessions came from data centres such as Google Cloud, OVH, Tencent and AWS, and 94% of those never clicked anything. None of them came from Meta. If your Direct traffic is inflated, look for headless browsers in the cloud first.
Verifying Museis unreliable.
Muse signs some requests with Web Bot Auth. In our logs only the meta-webindexer request was ever signed, for www.meta.com. HatchFeed and the Chrome 138 request never were. On 10 and 11 November the signatures were missing entirely. Adobe’s CDN logs show browser requests signed for muse.ai from 7 November. We saw none on our sites.
Cloudflare labels verified bots by category. It labelled Muse’s fetchers “AI Assistant”, then stopped:
| Days (consumer messaging app) | Muse requests | Labelled AI Assistant |
|---|---|---|
| 1–5 November | 132 | 132 (100%) |
| 6 November | 25 | 22 |
| 7–11 November | 434 | 0 |
| 12–13 November | 80 | 3 |
Nothing changed in the requests themselves: same addresses, same user agents, same TLS. ChatGPT-User kept its label on every day. If your rules let verified AI assistants through and challenge other bots, Muse moved from the first group to the second on 7 November, and nothing told you.
How to find Musein your own logs.
- Search your logs for HatchFeed.
grep HatchFeed access.logfinds the preview fetcher. It is the easiest of the three to spot. - Find the HEAD request. Look for
HEADwith exactlyMozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/138.0.0.0 Safari/537.36. It sendsAccept: */*and no Accept-Language, client hints or Sec-Fetch headers, which real Chrome always sends. - Check the address. All three fetchers came from
2606:54c1:1::/48. meta-webindexer from Meta’s own2a03:2880::/29is Meta’s indexer, not Muse. - Cloudflare. To count Muse, make a WAF custom rule with
http.user_agent contains "HatchFeed" or ip.src in {2606:54c1:1::/48}and the action Log. Change the action only if you want Muse users to stop getting your pages. - Log-based analytics. Exclude HEAD requests. Real browsers almost never send them for pages.
Do not rely on robots.txt or on Cloudflare’s verified label. Muse ignores the first and lost the second.
Sourcesand other reports.
Questions.
What is HatchFeed/1.0?
HatchFeed/1.0 (+https://agent.meta.ai; link preview) is the user agent of a fetcher that Meta’s Muse agent uses to load pages and their preview images. agent.meta.ai redirects to muse.ai. In our logs it always came from the same Cloudflare IPv6 range as Muse’s other fetchers.
Does Meta’s Muse respect robots.txt?
In our logs Muse never requested robots.txt: 0 requests in about 15,600. Meta’s crawler documentation does not list Muse and gives it no robots.txt token, so a robots.txt rule cannot name it.
Does Muse show up in Google Analytics or PostHog?
Its server-side fetchers do not. They never ran JavaScript in 15,600 requests, so script-based analytics never saw them. Tools that count server or CDN logs do see them, because one of the fetchers sends HEAD requests with a Chrome user agent.
Is meta-webindexer Muse?
Sometimes. meta-webindexer requests from Meta’s own network are Meta’s indexer. meta-webindexer requests from Cloudflare’s 2606:54c1:1::/48 range, signed for www.meta.com, arrived together with HatchFeed and came from the same addresses. We treat those as Muse.
Does Muse read llms.txt?
Rarely. Muse touched /llms.txt 13 times in about 15,600 requests: 3 full reads and 10 HEAD checks. ChatGPT-User and Claude Code read it more often on the same sites.
How do I block Meta’s Muse?
Match the user agent HatchFeed, HEAD requests with the exact Chrome/138 Linux user agent, and the range 2606:54c1:1::/48. robots.txt will not work. Blocking it also means Muse cannot read your pages for the people who ask it about you.
Agents read your site for people.Make sure they get it right.
Book a free 30-minute call with Emre. We’ll look at which agents visit your site, what they read, and how to turn their visits into customers.
Choose a time for your free 30-minute call.

