Meta has released Muse Spark — the first model from Meta Superintelligence Labs, the unit built over the last nine months by Chief AI Officer Alexandr Wang. The model is small, fast, and closed: a deliberate departure from Meta's previous open-source strategy with the Llama family. It will power Meta AI across Facebook, Instagram, WhatsApp, Messenger, and the Ray-Ban Meta AI glasses in the coming weeks.
For most of the technology press, this is a story about Meta's competitive position versus OpenAI and Google. For brands, it is something more immediate: a new AI model, trained on the world's largest social graph, is about to become the primary AI interface for billions of users across the platforms where those brands have spent years building presence.
What Muse Spark actually is
Muse Spark is not positioning itself as a frontier model. Meta is explicit about this — the model is "small and fast by design," built for efficiency rather than raw capability. What makes it significant is deployment context, not benchmark scores. Muse Spark will be embedded inside the products that 3.2 billion people use daily, answering questions about shopping, travel, health, trending topics, and local recommendations — precisely the categories where brand visibility decisions get made.
Meta has also signalled something structurally important about how Muse Spark will source its answers. The Meta AI app will reference content from the company's social media platforms when responding to queries related to shopping, trending topics, and locations. This is not incidental. It means that what Meta's AI says about a brand will be shaped by what exists on Meta's own platforms — Instagram posts, Facebook pages, Reels, user reviews, comments, and the aggregate social signal that Meta has accumulated about how users engage with that brand.
This is a different citation architecture from ChatGPT or Gemini. Those models draw primarily from the open web. Muse Spark will draw from a closed ecosystem — one where Meta controls both the training data and the distribution surface. The implications for brand representation are significant.
A closed model changes the citability equation
The platform coupling research published by NoGood in March documented how ownership and licensing relationships determine which sources each AI model can cite. The most extreme example in their data was X and Grok — 99.7% of X citations across 10 AI surfaces came from a single model, because Grok has native access to X's infrastructure. Muse Spark is likely to produce a structurally similar pattern for Meta's owned surfaces.
A brand's presence on Instagram and Facebook — the quality of its content, the consistency of its description, the engagement signals it has accumulated — will be more directly relevant to its Muse Spark citability than its presence on Reddit or LinkedIn. The universal substrates that matter for ChatGPT and Claude matter less here. What matters is how the brand has shown up on Meta's own platforms, over time, at scale.
And because Muse Spark is a closed, proprietary model, there is no public documentation of exactly how it weights these signals. Meta has said it "hopes to open-source future versions" — but the current model is not open. The training methodology, the citation architecture, the weighting of social signals versus web signals — none of this is publicly known. Brands are operating in a new black box.
The scale of what this means
Meta's capex for AI in 2026 is projected between $115 billion and $135 billion — nearly double last year. This is not infrastructure spending for a minor product update. It is the foundation for a multi-year repositioning of Meta's products around AI-driven interaction. Zuckerberg has been explicit about the endgame: an AI that knows each user's preferences, context, and social environment well enough to act as a genuinely personalised assistant.
When that assistant answers a question about which hotel to book, which brand of running shoes to consider, which financial product to enquire about — the answer will reflect what Muse Spark has learned about that brand from billions of social interactions. The brand that has been consistent, authoritative, and well-represented on Meta's platforms will have a structural advantage in that answer. The brand that has treated Instagram as a content calendar and Facebook as an afterthought will not.
This is the citability problem at a new scale. Until now, AI visibility strategy has focused primarily on the open web — structured data, authoritative backlinks, cross-platform entity consistency, content optimised for machine readability. All of that remains necessary. But Muse Spark adds a new dimension: social platform presence as a direct input to AI brand representation, on a surface used by more than three billion people.
What brands should do now
The practical response to Muse Spark is not to abandon the open-web citability work — that remains the foundation for ChatGPT, Gemini, Claude, and Perplexity. It is to extend the same discipline to Meta's platforms.
This means auditing how the brand is described across its Facebook and Instagram presence — not just what it posts, but how users describe it in comments, reviews, and user-generated content. It means ensuring that the brand's entity definition on Meta's platforms is consistent with how it is described everywhere else. It means treating Instagram captions and Facebook descriptions as structured content, not marketing copy — because Muse Spark will read them as data.
The AI Visibility Index currently tracks citation frequency, position, and sentiment across ChatGPT, Gemini, Claude, Perplexity, and Copilot. As Muse Spark establishes itself as a significant citation surface, monitoring how Meta AI represents a brand — and whether that representation aligns with actual positioning — will become part of the same measurement discipline.
Meta has just added a new judge to the room. It is sitting inside the products your customers use every day. And it is already forming an opinion about your brand.
Sources: CNBC · CNN Business · Meta Superintelligence Labs blog (April 2026).