Research | Meta AI Research

No frame-embedding restriction was observed (neither CSP frame-ancestors nor X-Frame-Options).

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. It is there, but weaker than this check expects.

Send Content-Security-Policy: frame-ancestors 'self' (or 'none', or a list of partner origins), or X-Frame-Options: SAMEORIGIN. It stops other sites from framing your pages to trick visitors into clicking (clickjacking).

X-Content-Type-Options: nosniff is not set.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. It is there, but weaker than this check expects.

Add it — one header, stops MIME-sniffing attacks. Most hosts let you set headers in config (vercel.json, netlify.toml, next.config).

No Content-Security-Policy detected.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. It is there, but weaker than this check expects.

Configure an enforced policy for script and resource origins; use report-only monitoring to prepare changes safely.

No explicit restriction on sensitive browser features was observed.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. It is there, but weaker than this check expects.

Disable unused camera, microphone, or geolocation access using Permissions-Policy.

The X-Powered-By header advertises "Next.js".

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. It is there, but weaker than this check expects.

You can remove it (in Next.js: poweredByHeader: false; in Express: app.disable('x-powered-by')). It adds bytes to every response and tells scanners what to probe; it is not a vulnerability by itself.

No HTML forms were found.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. This check does not apply to the features observed in this sample.

Check JavaScript-generated form transport separately if present.

No enabled, named password inputs in sampled HTML.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. This check does not apply to the features observed in this sample.

Review authentication forms separately when applicable.

Site serves over HTTPS.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. The inspected source provides this signal.

Nothing to do here.

http:// permanently redirects to https://./

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. The inspected source provides this signal.

Nothing to do here.

HSTS is set: max-age=63072000.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. The inspected source provides this signal.

Nothing to do here.

No Referrer-Policy is set, so browsers apply their default, strict-origin-when-cross-origin, which sends only the origin to other sites.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. The inspected source provides this signal.

Nothing to do here.

No insecure subresources detected.

HTTPS protects what visitors send you, and response headers limit what a hostile page can do in their browser. The inspected source provides this signal.

Nothing to do here.

Entity identity is inconsistent: the declared names differ (og:site_name "Meta AI Research", Organization name "Meta AI Research", Organization name "Meta Platforms, Inc."); Organization url https://about.meta.com/ points to a different site than research.meta.ai.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

Use one name for the site everywhere it is declared (Organization/WebSite schema, og:site_name, the title) and point identity URLs at this site.

• https://research.meta.ai/

No /llms.txt file.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

Optional: publish an llms.txt index of useful pages. This proposal is not a requirement for Google AI features, and its absence does not reduce the core score.

No FAQPage or QAPage schema on any crawled page.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

If the page answers recurring questions, use visible question-and-answer content and matching FAQPage markup. QAPage is for a single question with user-submitted answers. Google FAQ rich results are restricted to eligible authoritative government and health sites; markup does not guarantee AI citations.

Only 2% of 136 H2/H3 headings are question-shaped.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

Phrase key subheadings as the questions users actually ask ("How much does X cost?"). This improves the measured question-heading signal; the audit does not guarantee citations.

1/3 question-shaped heading(s) have a 20–500 character paragraph immediately after them.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

Put a direct 20–500 character answer paragraph immediately after each question heading so retrieval systems can quote it cleanly.

• https://research.meta.ai/

Only 3 of 4 provenance signal types were found: author, date, contact.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

Add the missing author, date, breadcrumb, or contact signal so machines can identify provenance and site structure.

• https://research.meta.ai/blog/muse-wonder
• https://research.meta.ai/blog/specialized-models-for-muse-and-metaai
• https://research.meta.ai/blog/supporting-medical-scientific-research
• https://research.meta.ai/blog/muse-spark-open
• https://research.meta.ai/blog/accelerating-scientific-research-in-health-data-science-with-muse-models
• https://research.meta.ai/blog/developing-capable-models-responsibly
• https://research.meta.ai/blog/solving-open-research-problems-together
• https://research.meta.ai/blog/bringing-your-muse-to-life
• https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

No markdown copy answered at /index.md or /blog/muse-wonder.md or /index.html.md.

This is what an AI assistant relies on when deciding whether it can quote your page. It is there, but weaker than this check expects.

Publish a clean markdown copy of each key page at /page.md (or /page.html.md). It strips navigation and scripts, so a retriever spends its budget on your content.

• https://research.meta.ai/
• https://research.meta.ai/blog/muse-wonder

No /llms-full.txt.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Optional companion to llms.txt containing your full content as one markdown file — useful for sites whose pages are JS-heavy.

44 concrete figures across 271 words.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Good — keep facts specific and current, and date-stamp them.

Description is short or non-declarative: "Research, releases, and ideas from Meta AI.…"

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Prefer a complete declarative sentence over slogans or questions — it should stand alone when quoted out of context.

No readable llms.txt was available to inspect for Markdown links.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Publish a readable llms.txt, then link its important pages with same-origin Markdown URLs.

No readable /llms.txt, so its structure cannot be checked.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Publish llms.txt first — the llms.txt check covers that — then this one verifies the structure.

The homepage answered Accept: text/markdown with text/html; charset=utf-8.

This is what an AI assistant relies on when deciding whether it can quote your page. This check does not apply to the features observed in this sample.

Optional: return a stripped markdown body when the request asks for text/markdown. Some hosts and CDNs do this for you; a plain HTML answer is not a defect.

No OpenAPI interface was advertised by sampled HTML.

This is what an AI assistant relies on when deciding whether it can quote your page. This check does not apply to the features observed in this sample.

An API is optional. Add a service-desc link if you already publish a machine-readable API specification.

The audited URL is allowed for evaluated answer-search crawlers.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

9 article(s) with author + date metadata.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

Semantic structure present: main×1, article×18, section×2, nav×5, header×2, footer×1.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

Paragraphs get to the point without preamble.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

Paragraphs average 62 words, lists present — chunk-friendly.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

Opening paragraph reads as a definition.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

dateModified present in schema.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

226 outbound reference(s) across crawled pages.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

Brand declared: og:site_name "Meta AI Research" + identity schema.

This is what an AI assistant relies on when deciding whether it can quote your page. The inspected source provides this signal.

Nothing to do here.

No privacy policy page is linked or crawlable.

A publisher review looks for this before it accepts a site. Right now the sampled pages don't provide it.

Publish an accessible privacy policy that discloses cookies and data use by Google and other third-party ad vendors.

1 sampled page(s) still read as unfinished — found "coming soon".

A publisher review looks for this before it accepts a site. Right now the sampled pages don't provide it.

Finish those pages, or redirect, noindex or remove them. Google policy prohibits ads on screens under construction.

• https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

1/10 editorial pages expose fewer than 250 publisher words; those 10 pages hold 14543 publisher words in total.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Review whether each page fully serves its purpose. Expand incomplete editorial content with useful information; keep concise navigation and tool pages when they help visitors.

• https://research.meta.ai/

The newest article has 509 words of body text.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Check that the page covers the question end to end. Short is fine when the topic is small; it reads as thin when the title promises a guide.

• https://research.meta.ai/blog/specialized-models-for-muse-and-metaai

No terms of use page found.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Consider adding terms that identify the service rules and publisher. This is a transparency heuristic, not an AdSense requirement.

No about page found.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Add an About page that identifies the publisher, site purpose, and relevant experience. This supports content transparency but is not a standalone AdSense requirement.

No contact page or mailto link found.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Add a working contact route or visible email so users can reach the publisher. This is a trust heuristic, not a published approval rule.

1/10 sampled editorial pages lack detectable author attribution.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Identify who created editorial content with a visible byline, author link, or matching Article author markup. This is a transparency heuristic, not an AdSense approval rule.

• https://research.meta.ai/

4/9 attributed editorial pages provide no discoverable author or publisher background.

A publisher review looks for this before it accepts a site. It is there, but weaker than this check expects.

Link the byline to an author profile or About page, or add truthful structured author details. This is a transparency heuristic, not an AdSense approval rule.

• https://research.meta.ai/blog/specialized-models-for-muse-and-metaai
• https://research.meta.ai/blog/muse-spark-open
• https://research.meta.ai/blog/developing-capable-models-responsibly
• https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

A privacy page was not present in the crawl sample, so its disclosures could not be verified.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Confirm the policy explains cookies or similar storage, Google or advertising vendors, and third-party data use.

No public contact email was found in the sampled home, About, or Contact pages.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

A public email is optional. The separate contact-method check covers whether readers have a way to reach the publisher.

No ads.txt yet.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

When AdSense provides your publisher entry, publish it exactly at /ads.txt and monitor authorization status in your account.

No cookie-consent banner or CMP script detected.

This observation could not be established from the available evidence. Resolve the collection limit or inspect the relevant source before deciding whether a change is needed.

Before serving Google ads in the EEA, UK, or Switzerland, implement a Google-certified consent management platform that integrates with the IAB TCF.