AI App Retention Benchmarks: What’s a Good 30-Day Retention for an AI Companion?
You’re two weeks post-launch. Your AI companion has decent installs, your D1 numbers look okay, and someone on your board asks: “Is our 30-day retention good?”
And you realize… you have no idea what good looks like.
Not because you havent done your homework. But because the benchmarks you’d normally reference — GameAnalytics for gaming, a16z for SaaS, Apptopia for social — were built for entirely different products. None of them map cleanly to what you’ve built. And every “AI app benchmark” article you find is either three years old or so heavily hedged it’s useless.
So here’s my honest attempt to give you actual reference points. Not perfectly precise, but grounded in what we know about the space, what real products are achieving, and what the data says about why users stay or leave.
^ you, presenting “engagement is up” to your board while D30 retention is undefined
Why the Benchmarks You Already Know Are Lying to You
The classic mobile gaming benchmark — roughly 40% D1 / 20% D7 / 10% D30 — is calibrated for products built entirely around dopamine. Games have streaks, leaderboards, reward loops. The whole architecture is engineered to pull you back. It has nothing to do with whether the product created value.
Social app benchmarks have a different problem. They assume a network effect. When your friends post something, you have an external reason to open the app. Instagram’s retention isnt driven by how good Instagram is, it’s driven by your friend who just got back from Bali. AI companions have none of that. There’s no social graph creating pull.
The closest analog, honestly, is journaling apps and mental health apps. Those products are also built around personal, reflective habits. Solo-use, intrinsic motivation, no external trigger. And journaling apps typically see 30-day retention in the 12-20% range from what’s been reported publicly. That’s a more honest baseline to compare against than mobile gaming.
But even journaling doesnt capture what makes AI companions unique, and uniquely hard: the relationship dynamic. Users aren’t just building a habit. They’re building a relationship with something that (ideally) gets smarter about them over time. That creates a completely different retention mechanic and a completely different failure mode.
The Actual Numbers: What AI Companion Products Are Seeing
Let me be direct about what I know and what I’m inferring. Some of this comes from public data, some from investor commentary, some from patterns we observe across AI products using Agnost.
Day 1 retention: 25-45%
For well-onboarded AI companion products, this is the realistic range. Character.AI reportedly hits 50-60% next-day retention, which is exceptional and reflects a product with strong distribution and a very well-defined use case. If you’re early stage without that brand pull, 30-40% is solid. Below 25% at D1, you have an onboarding problem.
The single biggest driver of D1 retention is whether the first conversation delivered a moment the user didnt expect. Not just “this is useful.” More like, “wait, this actually understood what I meant.” That micro-moment is the difference between a user who comes back tomorrow and one who forgets you exist.
Day 7 retention: 15-30%
This is the habit formation window and it’s brutal. You lose most users here, and its mostly not a product quality problem. It’s a habit formation problem. Users get busy. The novelty cools. And if the product hasnt created an intrinsic reason to return, it just… fades.
The users who survive D7 are disproportionately likely to stick around. Once you get past the early churn window, retention curves flatten in ways that look much better than traditional mobile apps. The problem is getting users there.
Day 30 retention: 8-18%
This is the core benchmark. Based on public data, investor commentary, and what we see at Agnost, 8-18% is the realistic range for most consumer AI companion products.
Chai reportedly maintains around 22% at 30 days. Character.AI sits in the 13-18% range. Replika, running a subscription model with stronger monetization commitment from users, sees healthy 30-day numbers for their paying cohort.
Best-in-class products — the ones with genuine conversation memory and personalization that compounds over time — are pushing toward 20-25% at D30. That’s the number to aim for. If you’re below 8%, you have a fundamental engagement problem that no growth spend will fix.
^ founders who discover their D30 is 4% after six months of “great early traction”
Why AI Companion Retention Is Uniquely Hard
The novelty effect is real and it’s vicious. Users come in curious. They have a few wild conversations. They show their friends. Then the initial magic fades and unless the product has created something genuinely intrinsic, they drift.
This is different from gaming where the core loop is designed specifically to fight novelty decay. And it’s different from SaaS where switching costs and workflow integration create stickiness. AI companions have neither. They have to earn the return visit every single time.
There are a few specific failure modes worth naming:
The memory problem is probably the biggest. Most users expect an AI companion to remember them, to accumulate context, to feel like a relationship that builds. When it doesnt — when every conversation starts from scratch or feels like the AI barely knows you — there’s no relationship forming. Just repeated interactions. And users stop coming back.
The “what do I actually use this for” problem is the second killer. Users who have a clear, recurring reason to open the app retain at dramatically higher rates than users who downloaded it and aren’t quite sure what their use case is.
And then there’s the re-engagement vacuum. Social apps have your friend’s activity. Email apps have the inbox. What brings someone back to an AI companion? For most products, the honest answer is: not much, unless the product proactively creates the moment.
The Conversation Patterns That Actually Predict Retention
This is where it gets interesting, and where most analytics tools completely miss the signal.
DAU/MAU ratios and session counts dont tell you why users are staying or leaving. The signal is in the conversations themselves.
Across products we track at Agnost, a few patterns show up consistently:
Users who have a conversation with 5 or more turns in their first session retain at roughly 2-3x the rate of users who dont. A five-turn conversation means the user got past the awkward introduction phase and actually engaged with something meaningful. It’s not a perfect proxy for quality, but it’s a strong leading indicator.
Users who return within 48 hours of their first session have dramatically higher 30-day retention. That second visit is everything. It’s the signal that a habit is forming. If you’re not actively trying to trigger that second visit within 48 hours, you’re leaving your best retention lever on the table.
Users who have at least one emotionally resonant conversation in their first week — something that felt personal, relevant, maybe a little vulnerable — almost always come back. The product has to create the conditions for that conversation to happen. It wont happen organically for most users without some nudge.
Every AI companion product has what I call an “activation conversation” — the specific type of conversation that separates users who stick around from users who churn. Finding yours is one of the most valuable things you can do with your data. It’s product-specific and you wont find it in industry benchmarks.
What the Best Teams Are Actually Doing About This
The products winning on retention are not winning because they have better AI. They’re winning because they’ve built retention mechanics that are native to the conversational medium.
A few things that separate the top performers from the pack:
Conversation memory that compounds. Not just remembering your name. Remembering what you said last Tuesday about your job situation and bringing it up naturally two weeks later. That’s the feature that creates the “it actually knows me” moment. It’s technically hard and most teams deprioritize it. Big mistake.
Re-engagement triggers based on conversation context, not timestamps. The teams doing this well are sending push notifications that reference something the user actually said. “You mentioned you had a big presentation this week. How did it go?” That converts at multiples of generic “come back to the app” notifications.
Tracking conversation health as a leading retention indicator. Are a user’s conversations getting longer over time? More personal? More varied? Or are they getting shorter and more repetitive? A flattening or declining conversation quality score is a churn signal that shows up weeks before the user actually leaves. You need to be watching it.
^ your product team the moment they find the activation conversation pattern in their data
The Metric You’re Probably Not Tracking
Most teams obsess over DAU/MAU ratio as their stickiness metric. For AI companions, it’s a necessary but insufficient signal.
The more predictive metric is what I’d call conversation quality retention: are the user’s conversations improving over time, or plateauing?
A user who has the same type of surface-level conversation every week is at churn risk even if their session frequency looks fine. A user whose conversations are getting longer, more specific, and more personal is building a genuine relationship with the product and they’re not going anywhere.
This metric doesn’t exist in standard analytics tools. It lives entirely in the conversation data. And most teams never look at it because their analytics stack isnt built to surface it.
We built Agnost specifically to surface patterns like this — conversation-level signals that predict retention weeks before it shows up in your standard metrics. If you’re serious about understanding why your AI companion’s D30 looks the way it does, standard product analytics won’t get you there.
Where This Is All Going
The category is still figuring out its native retention mechanics. We’re in the phase where most teams are applying social app and gaming playbooks to a product that needs its own framework.
The products that win at retention over the next two years will be the ones that build analytics infrastructure native to conversation, that understand the difference between a healthy conversation arc and a dying one, and that intervene based on conversation signals rather than time-since-last-open.
For now: if you’re hitting 15%+ at D30, you’re in the top tier of this category. If you’re between 8-15%, you have a real business but a real retention problem worth solving. Below 8%, the product hasn’t created the intrinsic value that brings people back, and that’s a product conversation before it’s a growth conversation.
TL;DR: D30 retention for AI companions runs 8-18% for most products, with top performers hitting 20-25%. Standard mobile benchmarks are the wrong comparison. The metrics that actually predict whether users stay live in the conversations, not in your standard analytics dashboard.
Reading Time: ~8 min