A person opens WhatsApp to message a friend, taps the search bar, and asks a question instead of typing a name. The answer appears inside the chat, drawn from a model trained on the open web and, increasingly, from public posts across Meta's own platforms. No browser opens and no publisher receives a visit. Repeated across more than a billion monthly users, that loop is what Meta AI means for the people who make the content behind the answers. Meta AI is the family of generative assistant features that Meta has embedded across WhatsApp, Instagram, Facebook, Messenger and a standalone meta.ai site, powered by its Llama models and supplemented with live results from Google and Bing. It reads the web to answer questions in place, and the place is an app the user already had open.
What is Meta AI?
Meta AI is Meta's branded AI assistant, not a single app but a layer that runs inside the company's existing products. It first appeared in the search bars of WhatsApp, Instagram, Facebook and Messenger in 2024 and has since expanded to a dedicated meta.ai web experience and a standalone Meta AI app. Under the surface it runs on Meta's open-weight Llama models, with real-time web results supplied by Google and Bing for time-sensitive queries such as news, sports scores and stock prices. The defining characteristic is distribution rather than novelty: the same conversational answer experience offered by ChatGPT, Gemini or Copilot, placed inside apps that billions of people open every day for other reasons.
How big is Meta AI's reach?
By reach, Meta AI is the largest consumer AI assistant in the world. Mark Zuckerberg told shareholders in 2025 that it had passed one billion monthly active users roughly eighteen months after launch, and 2026 estimates put the figure between around 640 million and 1.2 billion across Meta's family of apps depending on how engagement is counted. That headline number carries an important caveat for publishers and advertisers: much of the usage is driven by placement, not deliberate choice. Around 63 percent of Meta AI engagements happen inside WhatsApp, which passed three billion monthly users in 2025, and Instagram crossed three billion in September 2025. Industry analysts have argued that a large share of those interactions are incidental taps on a feature that sits in the way, rather than the high-intent sessions that ChatGPT and Perplexity attract. For a publisher, the practical point holds either way: an enormous audience is asking questions and receiving answers without ever leaving a Meta app.
How does Meta AI source its answers?
Meta AI draws on three pools of content, and only one of them is the open web in the way publishers traditionally understand it. The first is training data: Meta crawls the web with its Meta-ExternalAgent bot to build the datasets behind Llama, so published content can shape the model's answers long after it was read. The second is real-time retrieval, where Meta AI calls Google and Bing for fresh information on current events, which means a publisher's reporting can surface through Meta AI without Meta ever crawling that page directly. The third, and the most strategically significant in 2026, is Meta's own social graph: public Facebook posts, Group discussions, Reels and Marketplace listings are now treated as retrievable inventory. Meta has also confirmed it is building its own search engine to reduce its dependence on Google and Bing, which would close the loop further and route more answers through content Meta already controls.
What is Facebook AI Mode, and why does it matter?
Facebook AI Mode is a conversational search experience that Meta launched globally on 15 June 2026, and it represents the clearest signal yet of where Meta AI is heading. It lives inside the existing Facebook search bar, alongside options such as People and Marketplace, and returns synthesised, conversational answers drawn from public content across Meta's apps instead of a ranked list of links. A user can ask a question, read an answer assembled from Facebook posts, Groups and Reels, and ask follow-ups in the same thread. For publishers this matters because it reframes Facebook search away from link discovery and towards answer delivery, and because the retrieval material is Meta's own social content rather than open-web articles. Where Google AI Mode at least cites external URLs some of the time, Facebook AI Mode is structurally biased towards content created and held inside Meta's walls.
Why is Meta AI a particularly closed loop for publishers?
Meta AI sends almost no referral traffic, which makes it one of the most extractive AI surfaces a publisher will encounter. Other AI assistants at least produce occasional click-throughs to source pages; analyses of AI referral data in 2026 found that Meta AI's billion-plus monthly users generate effectively zero measurable outbound referrals, because messaging and social feeds are environments where leaving the app is rare by design. The content still does work in the system: it may have trained the model, or surfaced through Google and Bing results, or shaped a Facebook AI Mode answer. What it does not do is generate a page view, an ad impression or a subscription prompt for the publisher that produced it. This is the core asymmetry of the AI era expressed in its purest form - the read happens, the value is captured inside the app, and the open-web business model that depends on the visit never gets a chance to run.
Can publishers control how Meta AI uses their content?
Publishers have one practical control, and it is blunt. Meta's training crawler, Meta-ExternalAgent, is documented as respecting robots.txt, so a disallow directive for that user agent will stop Meta collecting a site's content for Llama training and for its own search index. The user-agent string is meta-externalagent, and unlike some crawlers it is generally considered compliant. The limits matter as much as the control. Blocking Meta-ExternalAgent does not affect facebookexternalhit, the separate bot that generates link previews when content is shared on Facebook, Instagram, Messenger and WhatsApp, so social sharing keeps working. It also does nothing about content that reaches Meta AI through Google and Bing results, because that retrieval depends on those engines' crawlers, not Meta's. And Meta currently offers no granular opt-out: there is no way to permit training but block real-time answers, or to allow general web answers while excluding a site from Facebook AI Mode. The decision a publisher faces is therefore all-or-nothing at the training layer, and largely out of reach at the answer layer.
What should publishers do about Meta AI?
The first step is measurement, because the read is invisible in the tools most publishers rely on. Meta AI consumption shows up neither in Google Analytics, which only records human page views, nor in referral reports, because there is rarely a referral. Server-side log analysis at the CDN edge is the only place the Meta-ExternalAgent crawl and other AI reads are observable, which means the question of how much content Meta is consuming can only be answered where the requests actually arrive. The second step is a deliberate access decision: whether to block Meta-ExternalAgent in robots.txt, accept the training use, or treat the AI read as inventory to be monitored and monetised rather than simply given away. This is the layer blankspace operates at. By detecting AI and Live Search Agent traffic at the CDN edge, blankspace lets publishers see the reads that analytics cannot and decide what to do with them, rather than discovering after the fact that a billion-user assistant has been answering from their work for free.
Frequently asked questions
Is Meta AI the same as ChatGPT?
No. Meta AI is Meta's own assistant, running on its Llama models and embedded inside WhatsApp, Instagram, Facebook and Messenger, whereas ChatGPT is OpenAI's standalone product. The most important difference for publishers is distribution and behaviour: Meta AI reaches its audience through apps people already use for messaging and social media, and it sends almost no referral traffic, while ChatGPT is a destination people open deliberately and at least occasionally cites and links to sources.
Does Meta AI use my website content?
Very likely, through more than one route. Meta crawls the open web with its Meta-ExternalAgent bot to train Llama and build its search index, so published content can shape Meta AI's answers. Content can also reach Meta AI indirectly when the assistant pulls real-time results from Google and Bing. A robots.txt block on meta-externalagent stops the direct training crawl, but it does not prevent your content surfacing through those third-party search results.
How do I block Meta AI from training on my site?
Add a disallow rule for the user agent meta-externalagent in your robots.txt file. Meta states that this crawler respects robots.txt, so the block should stop collection for Llama training and Meta's own search index. Note that this is separate from facebookexternalhit, the bot that creates link previews when your content is shared on Meta's platforms, so blocking the training crawler will not break social sharing previews.
Does Meta AI send any traffic to publishers?
In practice, almost none. Because Meta AI answers inside messaging and social apps, where leaving the app is uncommon, 2026 analyses of AI referral data found that its billion-plus monthly users produce effectively zero measurable outbound referrals. Publishers should assume Meta AI is a content-consumption surface rather than a traffic source, and plan their AI strategy on that basis.
What is the difference between Meta AI and Facebook AI Mode?
Meta AI is the overall assistant that runs across all of Meta's apps. Facebook AI Mode, launched globally on 15 June 2026, is a specific conversational search experience inside the Facebook search bar that answers questions using public content from Facebook posts, Groups, Reels and Marketplace rather than returning a list of links. Facebook AI Mode is significant because it leans on Meta's own social graph as retrieval material, which biases its answers towards content held inside Meta's platforms rather than open-web articles.