Your feeds are turning into search engines

For a long time, software had two ways to decide what you saw. Search matched the words you typed to content that answered them. Feeds matched you to people who behaved like you, then served what they engaged with. Those two methods are merging, and the merge is happening inside the products most people would never call a search engine.

Look at what LinkedIn has published about its own feed. As of late 2025, retrieval, the stage that decides which posts are even eligible to show you, runs on a fine-tuned version of Meta's LLaMA 3 used as a dual encoder. It turns members and posts into semantic embeddings and matches on those, consolidating several older retrieval systems into one. The ranking layer is heading the same way. LinkedIn's pre-production model, 360Brew, is a 150-billion-parameter system that starts from a Mixtral mixture-of-experts base, trained on LinkedIn's own data to fold a stack of hand-built, task-specific models into one you can prompt in plain language. On the company's own offline tests it matches or beats those production models, so for now it reads as a clear direction of travel rather than a shipped result. Strip away the scale and the shape is familiar. The feed now reads the meaning of a post and the meaning of your history, and pairs them. That's closer to how a search engine reads a query than to how a feed used to guess.

Audio is moving the same way, independently. Spotify has deployed a system it describes as semantic-ID generative retrieval, which frames recommendation as following an instruction over a catalogue of content turned into discrete semantic tokens. In a March 2026 production test, Spotify reports it lifted non-habitual podcast streaming by up to 5.4% and new-show discovery by up to 14.3%, though those are the company's own figures and the upper end of a range. The direction shows up in the buttons, too. Prompted Playlists now work for podcasts, so you can type the topic you want and have Spotify retrieve against it, and the company used its May 2026 Investor Day to describe an interface where you tell it what you want in words. Spotify's research team has been explicit that it sees one generative approach unifying search and recommendation, with search supplying the topic and recommendation supplying the behaviour.

So the trend is real, and it sits in primary sources rather than the usual stack of marketing predictions. Two large platforms have rebuilt discovery around matching meaning to intent.

Now the part the advice industry tends to skip.

These are hybrid systems. The semantic layer is added to collaborative filtering, not swapped in for it. Spotify's own description has its new system keep a collaborative-filtering signal as part of the input and run as one more source of candidates alongside the existing ones. LinkedIn's embeddings blend what a post is about with how people have behaved around it. "Your feed is becoming a search engine" is a fair description of the change in the mix. "Behaviour no longer matters" isn't, and the evidence doesn't back it.

The same caution applies to the creator advice that has attached itself to it. The popular version says the algorithm now punishes "topic-hopping", so you should keep most of your output, often quoted as 70 to 80%, on a single pillar. None of the engineering papers say that. They describe systems that encode topical preference to retrieve relevant content, not a penalty for posting across subjects, and no platform has published a content-mix ratio tied to reach. The pillar advice is a reasonable working habit borrowed from older content-strategy rules, not a documented platform rule. Encoding what you tend to read, so the system can find more of it, is a different thing from suppressing the person who writes about more than one subject.

What does follow from the evidence is quieter and more durable. If discovery is moving from "people like you also listened to this" towards "this matches what you're looking for", then the input that compounds is understanding what your audience is actually looking for. Not the format of the week, not the posting cadence, not the guess about this quarter's ranking signal. The words they use, the questions they're really asking, the stage they're at when they ask. That's the discipline search has always rewarded, and it's now reaching surfaces that were never thought of as search.

Which reframes the question most people ask. The common one is how to beat a given platform's algorithm this quarter. The more useful one is whether you actually understand what your audience is looking for, because that's the single input the next model rewrite won't reset.