Meta documents at least five web crawlers, and they do different jobs. One collects content to train foundation models, one feeds Meta AI search results and citations, one fetches pages when a user asks the assistant to, one serves advertising, and one builds link previews. A publisher's exposure to Meta AI depends on which of those it allows, and on a change to Meta's commercial terms that took effect on 30 October 2026. Meta AI is Meta's branded assistant, a layer that sits inside its existing apps and the standalone Meta AI app, so the audience asking questions is enormous even though the reading happens out of sight.
What is Meta AI?
Meta AI is the conversational assistant that Meta has placed across WhatsApp, Instagram, Facebook, Messenger, Threads and a standalone Meta AI app and website. Users reach it from search bars, group chats and posts, or by opening the app directly. Its defining feature is distribution: the same answer experience offered by ChatGPT or Gemini, placed inside products that billions of people already open every day for other reasons.
What changed in 2026?
The model underneath changed. On 8 April 2026 Meta released Muse Spark, the first model from Meta Superintelligence Labs, which TechCrunch described as a "ground-up overhaul" of Meta's AI efforts. It is presented as a fresh start rather than an update to Llama, and it powers the Meta AI app and web experience. On 12 May 2026 Meta added voice conversations and a Live AI camera mode to the app, and 9to5Mac reported that Muse Spark is also bringing Meta AI to search bars, group chats and posts across Meta's apps, including a Threads chatbot.
On its third-quarter 2026 earnings call, Meta said more than 10 million people in the US and Canada had set up Muse agents in about six weeks, and that Muse is limited to those two countries because of compute demands. Older commentary that describes Meta AI purely as a Llama product is now out of date.
How big is Meta AI?
Meta last gave an absolute figure in October 2025, when it said Meta AI had more than one billion monthly users. Gradually, a statistics site that tracks Meta's disclosures, found no newer total as of 31 October 2026. On the second-quarter 2026 call, Meta said daily interactions with the assistant had risen 60 per cent since the Muse Spark rebuild, and on the third-quarter call Mark Zuckerberg said engagement per person had more than doubled since April. The only audience number from the latest quarter is 3.64 billion daily active people across Meta's family of apps, which counts every app together and says nothing specific about the assistant.
Treat the billion figure with care. Much Meta AI usage is placement-driven, because the assistant sits in search bars people use for other purposes, so a monthly user is not the same as a person deliberately seeking an answer.
How does Meta AI get its content?
Meta AI can draw on several routes, and a publisher can see only some of them.
The first is crawling. Meta-ExternalAgent crawls the web to train foundation AI models or to improve products by indexing content directly. Meta-WebIndexer analyses web content to improve the quality of Meta AI search results and, according to Meta's crawler documentation, to support citing and linking to sources in Meta AI responses.
The second is user-requested fetching. Meta-ExternalFetcher retrieves individual links when a user asks, to support agentic features. Meta's documentation says it may bypass robots.txt because the fetch is made at a user's request.
The third is licensing. Meta used Reuters content to answer news questions in real time from 2024. In December 2025 Digiday reported that Meta had signed seven multi-year AI content deals, naming CNN, Fox News, People Inc. and USA Today Co., with terms undisclosed.
The fourth is Meta's own social content, which is where Facebook AI Mode comes in.
What is Facebook AI Mode?
Facebook AI Mode is a conversational search experience inside the Facebook search bar. The Next Web reported on 15 June 2026 that it was rolling out to US users, returning synthesised answers drawn from public posts in Groups, Reels and Marketplace listings instead of a list of links. The same report said Meta had not said whether group admins or users can opt public posts out, and Meta has made no public statement on publisher attribution or compensation for it. For publishers the significance is directional: Meta is building an answer surface that draws mainly on content it already holds, not on the open web.
Does Meta AI send traffic to publishers?
Nobody outside Meta can say with confidence, because Meta publishes no referral data for the assistant. Meta-WebIndexer's documented purpose includes citing and linking to sources, so links exist in principle. Whether people click them inside messaging and social apps, where leaving the app is uncommon, is unmeasured in any source this article could verify. Earlier claims that referrals are "effectively zero" came from third-party analyses and should be checked against your own data. Meta AI is hard to isolate in analytics because in-app browsers and messaging apps often strip or blur referrer information, which is why AI referral traffic needs deliberate tracking.
Which Meta crawlers can publishers control?
Meta's crawler page lists these user agents:
- meta-externalagent: foundation model training and direct indexing. Meta says it can be blocked with a robots.txt disallow rule.
- meta-webindexer: Meta AI search quality and citation. Documentation implies robots.txt control but does not state it explicitly.
- meta-externalfetcher: user-requested link fetches for agentic features. May bypass robots.txt.
- meta-externalads: advertising and business products. Compliance not stated.
- facebookexternalhit: link previews when content is shared on Meta apps. Usually honours robots.txt but may bypass it for security checks.
The practical consequence is that blocking meta-externalagent does not remove a site from every Meta AI use. Meta's own crawler documentation lists separate crawlers for AI training and Meta AI search, so a publisher can in principle allow one and block the other. A robots.txt rule is still a request, which is why robots.txt alone cannot stop AI crawlers, and the full token list is in the AI crawler user agent list.
Do Meta's new commercial terms change anything?
They add a contractual layer. Updated Commercial Terms for businesses with Meta accounts took effect on 30 October 2026. Search Engine Journal and Unite.AI report that Meta removed "train" from the crawling clause on 17 October after criticism from the News/Media Alliance and others. The revised clause still lets Meta use automated tools, including test credentials, to access and copy content from sites associated with a business's products or services, but ties the purpose to advertising and business products and to the integrity and security of Meta's products. Meta says its crawling practices have not changed and that robots.txt preferences still apply, although the new wording does not mention robots.txt and legal commentators say it is unclear whether a robots.txt block limits what a business grants by accepting the terms. The detail is covered in do Meta's new commercial terms let it scrape publishers' sites for AI.
What should publishers do about Meta AI?
Start with visibility. Meta AI reads do not appear in client-side analytics, and a declared user agent in a server log is the only reliable trace. Log analysis at the CDN edge shows how often each Meta crawler visits, whether claimed user agents match Meta's published IP ranges, and whether anything arrives undeclared.
Next, set a deliberate policy per crawler instead of one rule for "Meta". Decide separately on training (meta-externalagent), citation and search (meta-webindexer) and user-requested fetches (meta-externalfetcher), and review the Meta business accounts your organisation holds, because the commercial terms attach to them.
Finally, treat licensing as the alternative to blocking. Meta has signed only a handful of deals, so most publishers have none. blankspace works at the CDN edge, detecting AI and Live Search Agent traffic so publishers can see these reads and decide whether to allow, block or monetise them.
Frequently asked questions
Is Meta AI built on Llama?
Not any more. Meta released Muse Spark, from Meta Superintelligence Labs, on 8 April 2026, and it now powers the Meta AI app and web experience. TechCrunch presented it as a ground-up overhaul rather than a Llama update. Meta has said it plans to release new open-weight models, so Llama-era descriptions of Meta AI are out of date.
How do I stop Meta AI using my content?
Add disallow rules to robots.txt for the Meta crawlers you want to exclude, starting with meta-externalagent, which Meta says respects it. This does not cover meta-externalfetcher, which Meta says may bypass robots.txt for user-requested fetches, and it does not undo any contractual permission in Meta's commercial terms. Enforcement of a block needs bot management at the CDN layer.
What is the difference between Meta-ExternalAgent and Meta-WebIndexer?
Meta-ExternalAgent crawls to train foundation models and index content directly. Meta-WebIndexer improves Meta AI search results and supports citing and linking to sources. Because they are separate tokens, a publisher can block training while allowing the crawler that may produce citations, though Meta's documentation does not guarantee citation as a result.
Does Meta pay publishers for AI content?
Some, not most. Meta used Reuters content in Meta AI from 2024, and in December 2025 Digiday reported seven multi-year AI content deals including CNN, Fox News, People Inc. and USA Today Co. Terms were not disclosed, so it is unknown whether payment is lump-sum or usage-based.
What is Facebook AI Mode?
It is a conversational search experience in Facebook's search bar that answers questions from public posts in Groups, Reels and Marketplace instead of listing links. The Next Web reported a US rollout on 15 June 2026. Meta has not said whether public posts can be opted out or how publishers are credited.