Agents now hit signup flows that bounce off human‑designed browsers, so platforms are turning that traffic into signups via tools like Work OS Agent Registration and Auth Kit.

The core issue lies in the first step of a feed pipeline: retrieval. For months, platforms used engagement as a cheap proxy for relevance, but that signal is easily gamed by bait posts that generate clicks while offering little value. The fix is to shift from behavioral to semantic retrieval, using embeddings to match users with content based on meaning rather than interaction history.

Meta, LinkedIn, and YouTube each adopted semantic models but in different ways. LinkedIn consolidated five legacy systems into a single dual‑encoder built on a fine‑tuned LLa MA-3 model, enabling sub‑50‑millisecond latency. Meta kept a funnel of specialized models, while YouTube generated content identifiers from a generative approach.

These changes aim to reduce the effectiveness of engagement bait, but they introduce trade‑offs in cost, cold‑start handling, and system complexity.