Meta’s New Feed Controls Make a Promise. The Ranking System Still Decides.

“Show me more of this” is a clear preference. “I want local science reporting, but not celebrity coverage” is closer to how people actually think about a feed. Meta’s new Facebook controls let users describe such preferences in natural language, while leaving the platform to decide how those words affect ranking.

Meta announced “Your Algorithm” on 13 October 2026. The company says people can view and manage topics affecting Facebook Feed and Reels, and can add a natural-language explanation after using Interested or Not interested. Meta also says these preferences work with existing engagement signals and can affect recommendations in the same session. Those are product claims from Meta; independent measurement of persistence, weighting or outcome quality was not provided in the announcement. ^[1]

This is a meaningful interface change, but it is not Facebook’s first user control. Meta introduced Show More and Show Less in 2022, allowing people to influence ranking from individual posts. The new experience appears to make preferences more explicit and topic-oriented, rather than creating user input from nothing. ^[2]

A preference is an input, not a ranking rule

Recommendation systems combine many signals. Meta’s own system cards describe how AI systems power content recommendations and explain that different surfaces use different ranking processes. The company’s announcement says natural-language feedback is integrated with its existing signals, not that it replaces them. That means a person’s request enters a system whose translation, weighting and tradeoffs remain platform-controlled. ^[3]

Suppose someone asks for more neighborhood reporting and less celebrity news. A ranking system might map “neighborhood” to topics or content categories, then combine that preference with prior watches, follows, comments and available inventory. If the system interprets a local politics story as “news” and shows more of it, the result may be technically consistent with its mapping but not with the person’s intent.

The interface therefore creates a new question for product design: can users tell what their request changed? A preference control is more useful when people can inspect it, revise it, see examples of its effect and reverse the choice. Natural language can lower the cost of expressing nuance, but it can also create a mismatch between a person’s plain-language expectation and the system’s internal categories.

What recommendation research suggests about control

Research on user-controllable recommender profiles explores letting people directly edit the preferences a model uses, rather than relying only on observed behavior. The CHI 2025 paper reports results from its own study and system; its findings should not be generalized automatically to Facebook or every recommender. It does provide a useful design lens: control depends on whether user edits produce predictable recommendation changes. ^[4]

Fairness research raises another issue. A conversational recommendation system may satisfy an individual request while still producing uneven exposure or reinforcing broader imbalances. The 2024 FairCRS work studies fairness in conversational recommendation, a separate system and setting from Meta’s feed. Its relevance here is conceptual: user satisfaction and system-level outcomes are different evaluation questions. ^[5]

For platform owners, an evaluation should test both. Does the system respond to a stated preference? Does it preserve diversity and avoid systematically suppressing particular sources or groups? Does it behave differently for users with sparse histories? Can a person understand why the feed changed? Meta has not published independent answers to those questions for this new feature.

A practical audit for users and builders

Users can try a specific preference rather than a broad instruction, then observe the feed across several sessions. Note whether the interface shows the saved preference, whether a related topic changes, and whether undoing it restores the prior experience. Feature availability and controls can change by account and region, so use what the current app actually displays.

Product teams building recommenders can borrow a small test plan:

1. Capture intent: let users state what they want without forcing every preference into a fixed taxonomy.
2. Show interpretation: translate the request into editable categories or examples so a user can correct a misunderstanding.
3. Measure the effect: compare recommendations before and after a preference, including whether the result lasts beyond one session.
4. Test side effects: track diversity, source exposure and outcomes for users with different histories.
5. Provide reversal: make it easy to clear or revise feedback and explain what the control can and cannot affect.

Our guide to how recommendation pipelines work explains the ranking and feedback stages behind a feed. The core design challenge is shared across platforms: a visible control can make user intent legible, but trust depends on the system’s response being inspectable and bounded.

The promise is legibility

Natural-language feedback is valuable if it gives people a more accessible way to express preferences and the platform makes the resulting change understandable. It is less useful if a request disappears into a ranking model with no way to inspect its interpretation or persistence.

Meta’s announcement is an early product description, not an independent audit. The distinction between direct preference and inferred engagement remains. Better controls can narrow that gap, but the platform still chooses how to combine a person’s words with the rest of the system. The feature’s real test is whether users can see that their choices matter—and identify when they do not.