XO Whitepaper
Xo appChatroom

ROO - In-Chat AI

Overview

ROO is XO's in-chat AI companion. Rather than a single chatbot, ROO is a suite of AI features users can summon inside any chatroom to help open, sustain, and rescue conversations. ROO is built on a multi-model stack (GPT-5, Gemini Flash) tuned per use case — fast and cheap for high-frequency suggestions, high-quality for moments that matter.

  • Non-intrusive: ROO only acts when a user summons it
  • Mutual by design: games require both users to participate before reveal
  • Evidence-driven: features ship behind A/B tests; ones that don't move retention or matches get sunset
ROO features are summoned 100k+ times per day across the XO chatroom.

Feature Lineup

FeatureWhat it doesWhen it helps
Clapback (Decode / Rizz Up)Reads the last few turns and suggests a sharp, on-tone replyStuck on what to say next
NewsFeed IcebreakerDrafts a personalised opener from another user's postFirst message — converts a like into a real conversation
Close SaverDetects a fading conversation and offers a save moveA chat is one silence away from dying
Vibe Check / Chemistry GameMutual compatibility quiz with shared revealMid-conversation, when you want to go deeper without the awkwardness
Cheat CodePremium top-tier reply suggestions for high-stakes momentsWhen the conversation is going somewhere and you don't want to fumble it
Truth or Dare / Red Flags / RPS / DiceLightweight mini-gamesFiller that keeps energy up between deeper turns

Why ROO Works

XO's conversation success research identified four structural signals that separate matches that lead to friendship from those that fizzle:

  1. Symmetry — both sides putting in comparable effort (strongest signal, d=+1.69)
  2. Self-disclosure depth — Level 3+ vulnerability gates 5.6× higher friendship rate
  3. Reciprocity rhythm — response latency and turn-taking balance
  4. Attachment-style compatibility — secure × secure has ~2× success vs anxious × avoidant

Each ROO feature is mapped to one of these levers:

  • Clapback / Rizz Up / Cheat Code address Symmetry (help the under-responding side keep pace) and Disclosure (suggest replies that surface real opinions, not small talk)
  • Vibe Check explicitly elevates Disclosure with mutual reveal mechanics
  • NewsFeed Icebreaker rewrites a flat "hi" into a content-anchored opener — directly targeting first-turn Disclosure
  • Close Saver intervenes at the Reciprocity break-point, before the conversation goes silent

Built-in Discipline

ROO is governed by two operational rules:

  1. A/B before scale — every new ROO feature ships behind a flag with a control cohort. We measure 7-day match outcome, not just acceptance rate.
  2. Sunset what doesn't lift — features that show null causal lift after sufficient sample are killed (Wingman, AI Twin, AI Memory Loop have all been sunset for this reason). The lineup above is what survived.

This is why the public lineup is short. Most "AI dating assistants" ship a dozen features and never measure if any of them work. ROO ships fewer and proves each one.

FAQ

What is ROO on XO?
ROO is XO's in-chat AI companion — a suite of AI features users can summon inside any chatroom to help open, sustain, and rescue conversations. ROO is non-intrusive (only acts when summoned) and is built on a multi-model stack (GPT-5, Gemini Flash) tuned per use case.
What AI features does ROO offer?
ROO's lineup: Clapback (Decode / Rizz Up) for reply suggestions; NewsFeed Icebreaker for opening messages from another user's post; Close Saver for rescuing fading conversations; Vibe Check / Chemistry Game for mutual compatibility quizzes; Cheat Code for premium top-tier reply suggestions; and lightweight mini-games (Truth or Dare, Red Flags, Rock-Paper-Scissors, Dice).
Does ROO read my private messages?
ROO only reads recent chatroom context when a user explicitly summons a feature (e.g. tapping Clapback for a reply suggestion). It does not background-scan conversations. Suggestions are generated server-side and not stored beyond standard logging for safety review.
What happened to the AI bartender, Wingman, AI Twin, and AI Memory Loop?
All have been sunset. After A/B testing showed null causal lift on retention or match outcomes despite high acceptance rates, they were removed. Wingman was killed via system config flip in mid-2026; AI Twin and AI Memory Loop never shipped beyond beta. The current ROO lineup is what survived rigorous A/B measurement.
How does ROO improve conversation success?
Each ROO feature is mapped to one of four structural signals XO's research identified as predictive of friendship outcomes: symmetry (both sides putting in comparable effort), self-disclosure depth (Level 3+ vulnerability gates 5.6× higher friendship rate), reciprocity rhythm (response latency balance), and attachment-style compatibility. For example, Vibe Check explicitly elevates Disclosure through mutual reveal mechanics, while Close Saver intervenes at the Reciprocity break-point.

On this page