Meta · Facebook · 2025-2026Facebook Personalization
Tell us why
The tension
You scroll past another post that has nothing to do with you. You hit "Not Interested" — and tomorrow, the same kind of content is back.
Despite years of machine learning advances, 21% of Facebook users still rate their personalization as "very bad." Relevance is the top barrier — especially for young adults, the audience that matters most for long-term retention. Users cite poor interest matching, stale recommendations, and a feeling that their feedback goes into a void.
The problem wasn't that people lacked controls. It was that the controls didn't feel like they worked — and the platform wasn't asking the right questions.
My role
I partnered closely with Content Design, UX Research, Feed Ranking, and Integrity teams to ensure every design decision balanced three tensions: giving users meaningful control, preserving signal quality for ranking models, and keeping the experience simple enough that people actually use it. This work spans H2 2025 through H1 2026, with a long-term vision extending into 2027.
As Product Designer, I owned the end-to-end experience design for Tell Us Why (TUW) — Facebook's contextual feedback framework. I drove the strategy, interaction model, and visual design across five experiment cycles, from initial concept through public launch.
The insight: ask why, not just what
The existing "Interested / Not Interested" buttons captured what users felt — but not why. A "Not Interested" click could mean "I don't like this creator," "this is spam," "this topic isn't for me," or "I've seen this too many times." Without specificity, the ranking model was guessing at intent.
TUW reframes feedback as a conversation. Instead of a single binary signal, we ask a contextual follow-up: Why aren't you interested? The user selects from options grounded in what actually helps ranking — and what actually matters to users. Every option was chosen based on three criteria:
Value to users — validated in prior research
Value to ranking models — either existing models or ones the team could build
Cross-surface consistency — options that work across Feed, Reels, and Video
The framework: a personalization flywheel
The strategic foundation is a flywheel: when controls feel effective, people use them more. More usage generates higher-quality signal. Better signal improves the models that power relevance. Better relevance makes the controls feel more effective. And the loop compounds.
This isn't theoretical. In holdout tests, explicit controls like I/NI showed measurable impact: +0.35% VPVs and +0.17% DAU. Users are increasingly treating feedback tools not as safeguards, but as personalization levers — ways to fine-tune what they see.
TUW is designed to accelerate this flywheel by making every feedback interaction more specific, more legible to ranking, and more immediately satisfying to the user.
User journey
The TUW experience follows a five-step journey:
Discovery— The user encounters a post that does not feel relevant. Frustration or indifference builds.
Engaging— The user taps "Not Interested" or swipes to dismiss. The post is replaced by a tombstone.
Evaluating— The tombstone surfaces contextual follow-up options as selectable pills: "Does not match my interests," "I do not like the creator," "Spam," etc. The user selects one or more reasons.
Re-engaging — A confirmation message acknowledges the feedback and signals that the system heard them. Over subsequent sessions, the user sees fewer posts matching their stated reasons. The feedback loop closes — the control felt effective, reinforcing future usage.
This journey is the backbone of the flywheel: each pass through the loop generates richer signal and builds user trust that the system actually listens.
Rapid experimentation: four rounds, four lessons
The core of the work was a series of rapid experiments, each one a full design cycle from hypothesis to public test, each building on what the last one taught us.
Content strategy: every option earns its place
Designing the TUW option set was as much a content strategy challenge as a design one. Every option had to pass three bars: does research validate it? Can ranking use it? Does it work across surfaces?
Some decisions were uncomfortable. "Political" was removed after V1 — not because users didn't care, but because framing political content as inherently negative created policy tensions. "Misleading" and "Disturbing" were cut for being too ambiguous for ranking models to act on. "Spam" and "Insult" were added to align with Integrity team priorities.
The lesson: in a feedback system, what you choose not to ask is as important as what you ask.
Looking ahead
V4: Modernization
The next evolution moves TUW into a contextual message format — aligned with the Facebook Blueprint team's recommendations and the Reels team's long-term direction. The goal: declutter the interface, simplify the interaction, and create a unified controls format across surfaces.
We explored four design directions — contextual message, clear sections with titles, a minimal TUW-only update, and a bottom sheet — before recommending the contextual message approach for its balance of clarity and consistency.
North star: Meta AI Controls Bot
The long-term vision is a single, AI-powered surface for all user controls. Instead of scattered settings across menus, a conversational interface where users can say what they want — and the system understands, acts, and learns.
The roadmap moves in three phases:
H1 2026: Add low-severity controls (content preferences) to the AI bot
H2 2026: Transition high-severity controls (report, block) while simplifying contextual controls
H1 2027: Full control functionality through the AI bot, with contextual controls as complementary support
This isn't just a UI change. It's a fundamental shift in how people relate to the platform: from managing settings to having a conversation about what they want.
Reflection
This project taught me that feedback design is relationship design. Every time someone tells a platform "I don't want this," they're extending a small act of trust — trusting that the system will listen, learn, and change. When it doesn't, that trust erodes. And once it's gone, people stop giving feedback at all.
The hardest part of this work wasn't designing the UI. It was navigating the space between what users want to express, what ranking models can act on, what policy allows, and what keeps the experience simple enough to actually use. Every experiment was a negotiation between these forces.
What I'm most proud of is the flywheel thinking — designing not just a feature, but a system where better feedback cre