Meta Muse and AI agents: what an 11-point website checklist gets right

Short answer

Which points on the viral Meta Muse checklist have sources behind them?

Most of the 11 points are accessibility and plain HTML advice with published sources behind them. Meta-WebIndexer is a real Meta crawler. Naming it in robots.txt only matters if your site blocks crawlers by default. The "trust score" and the 30 to 90 day refresh rule have no published source. The question heading and the 100-word answer are editorial choices. Meta has not said how Muse chooses what to recommend. So treat any claim that a change made Muse recommend a product with care.

§ 1 The claim

On 19 October 2026 Rob Hoffman, CEO of Contact.so, posted a thread on X.[1] He said Meta Muse was recommending a client's product as the best option on the market. He then listed 11 changes to make a website easier for Muse and other AI agents to read and recommend. The thread spread fast.

I cannot check the client result. The post does not name the product. Muse answers are also private to each user. What I can check is the advice. Each point either has a published source behind it or it does not. This note sorts them.

§ 2 What Meta Muse is

Meta announced Muse on 8 September 2026 as a personal AI agent.[2] Meta says it runs on Muse Secure VM, a dedicated secure computer with its own browser. It can open that browser, fill out forms and complete checkouts for the user. As of October 2026 it is rolling out in the US on iOS, Android and muse.ai.

Neither Meta's launch post nor its post on Muse for small business gives website owners any guidance.[3] They do not say how Muse reads a page, which signals it trusts or how it picks a product to recommend. So any list of "Muse optimisations" is inference. Some of that inference rests on solid ground, because Muse uses a browser and browsers expose pages in documented ways.

§ 3 The 11 points, checked

The table names each point in a few words of my own. The middle column says what the sources show. The last column is my verdict.

So most of the list is accessibility and plain HTML. That is good news for a site owner. The same work helps screen reader users, search engines and AI agents at once.

§ 4 The robots.txt line

Meta's crawler page lists Meta-WebIndexer. It says the crawler improves Meta AI search results. It also says that allowing it in robots.txt helps Meta cite and link to your content in Meta AI's responses.[4] The same page lists Meta-ExternalAgent for uses such as training models. It lists Meta-ExternalFetcher for links a user asks for. Meta says the fetcher may bypass robots.txt.

The suggested group looks like this.

User-agent: meta-webindexer
Allow: /

Under the Robots Exclusion Protocol, a crawler may fetch any URL that no rule disallows for it.[13] So the line changes nothing on a site that already allows every crawler. It matters when a site blocks crawlers by default with a wildcard. There, naming Meta-WebIndexer is how you let it in. Meta's page does not say whether Muse itself relies on this crawler.

§ 5 The two points with no source

The trust score. The thread says broken links lower a site's "trust score". None of the operators I checked documents such a score. Broken links are still worth fixing. Vercel found ChatGPT spent 34.82% of its fetches on 404 pages and Claude 34.16%, against 8.22% for Googlebot.[9] Every wasted fetch is a page the crawler did not read. Google gives the same advice in plainer terms. Links should be real anchor elements with resolvable URLs.[12]

The 30 to 90 day refresh. No operator I checked gives a refresh interval. Google's guidance is about honest dates. Show a visible date and keep it consistent with the structured data.[11] Changing a date without changing the content does not make it fresh. "As of" statements do more, because they date the fact itself. That is the timeliness dimension of data quality, covered in my note on data quality before AI.

§ 6 This site against the list

I ran the points that apply against frankvitetta.net.

• Accessible names. The menu button carries aria-label="Open menu". All five fields on the contact form have a label element.
• Images. The site has no image elements at all. Diagrams are inline SVG with a title and a description, so their content is text.
• Raw HTML. The site is static. Every word of every note is in the HTML before any script runs. The scripts only draw the animations and run the menu and the term search.
• Semantic HTML. Pages use header, nav, main, article, section and footer. The menu is a button element.
• Dates. Each note shows Published and Last updated dates that match its structured data.
• robots.txt. The file allows every crawler. Meta-ExternalAgent and Meta-ExternalFetcher were already named in it. Meta-WebIndexer is now named too.

None of that is special to AI. It is how the site was built from the start.

§ 7 What to take from it

This section is opinion. Do the accessibility and HTML points first. They have sources and they help everyone who visits. Add the robots.txt line if your site blocks crawlers by default. Treat the heading, the 100-word answer and the refresh interval as editorial choices. Test them before you trust them.

Be careful with claims that a change made an agent recommend a product. Agent answers vary between runs and between users. One screenshot is one sample. My note on prompt length and brand mentions shows how much a measured rate can move with the test set alone. The terms used here are defined in the index of terms.