Earnings calls are usually where companies talk in careful, hedged language about the future. On Meta’s Q2 2026 follow-up call, held July 29, 2026, Susan Li did something rarer: she described, in plain terms, exactly how Meta’s recommendation systems are changing at the architectural level — and it’s about as close as we’ve seen a platform get to confirming the Meta algorithm in 2026 is built on a fundamentally different premise than it was even a year ago.

From Content IDs to the Semantic ID Meta Just Confirmed

For years, recommendation systems have leaned on content IDs — simple reference tags that let a model look up a piece of content without understanding much about it. Susan Li described this old approach directly: content IDs for ranking, she said, carry only a narrow slice of information about what a post actually is.

What’s replacing it is a semantic ID system — one built so the model can go beyond a lookup tag and actually “infer why it’s interesting to someone.”

That’s the whole shift in six words. Not “what is this content” as a label. “Why does this matter to this person” as a reasoned judgement. Meta’s CFO, on an earnings call, describing a semantic ID system built specifically to let the model reason about relevance rather than just retrieve it.

Social Graph vs AI Recommendation: The Shift Meta’s CFO Described

Susan Li didn’t stop at describing the semantic ID system in isolation. She tied it to where Meta is heading architecturally: collapsing what has historically been a multi-stage recommender — separate systems for retrieval, ranking, and delivery — into a single, unified model.

That’s a bigger statement than it first sounds like. A multi-stage recommender built on content IDs is, underneath, still a social-graph-era system: it retrieves candidates based on connections and history, then ranks them with engagement heuristics layered on top. A single model reasoning over semantic IDs is something else entirely — content understood and judged on its own merits, independent of whether the graph already connects you to it.

This is the social graph vs AI recommendation shift, described by the company making it happen. Not gradual tinkering with an existing system. A stated architectural goal of replacing graph-and-heuristic ranking with one model built to understand content directly.

SDO and Clue Labs: Built for This Exact Shift

This is the premise Clue Labs was built on before it had a name for it: that social platforms had already moved from social graphs to AI-driven discovery engines, and that succeeding on them required understanding content the way the algorithm does, not the way a follower count implies.

We call that category Social Discovery Optimisation — SDO. It’s not a rebrand of SEO for social platforms. It’s built on the specific claim that visibility is no longer primarily a function of who follows you, but of whether an AI system can understand what your content is and why it’s worth showing to someone who doesn’t. Susan Li just described that exact claim from the inside: content IDs (the graph-adjacent old model) giving way to semantic IDs (content understood and reasoned about directly), with the end state being one model doing the whole job.

Clue Labs’ Clue Score and content prescriptions exist to help accounts operate inside that reality now, rather than finding out about it once reach has already collapsed.

What This Means for You

A few practical implications follow directly from this shift:

• Your content needs to be legible on its own terms, not just to your existing audience. A semantic ID system is built to judge content by what it is, independent of who already follows the account posting it.
  
• “Why is this interesting” is now a modelled question, not a guess. If you don’t know why a specific post should be interesting to someone who’s never seen your account before, the algorithm’s job just got harder — and so did yours.
  
• The multi-stage-to-single-model shift means fewer places for a weak piece of content to sneak through.Separate retrieval and ranking stages used to give mediocre content multiple chances. A unified model judging content directly has fewer blind spots to exploit.

Why This Matters

Most platforms let creators and marketers guess at the mechanics underneath their feed. On this call, Meta’s own CFO described the mechanics directly — content IDs giving way to semantic IDs, multiple stages collapsing into one model, all in service of a system that can explain why a piece of content matters, not just that it exists.

That’s the shift Clue Labs and the SDO category were built to help people operate inside. Meta didn’t confirm a theory here. They confirmed the architecture.

Now it’s your turn:

Source:
Meta Platforms, Inc., Q2 2026 Follow-Up Call, July 29, 2026 (Susan Li, CFO). Official transcript and call materials: investor.atmeta.com, p.6–7 (content IDs, semantic ID system, unified recommender model).