I A/B tested 12 morning check-in questions. Only 3 predicted whether users came back.

We shipped a morning check-in in WoW Brain — a two-minute AI-guided prompt that reads the user's state and picks the cognitive workout for the day. Users like it. It has a 76% completion rate, which is high for anything that asks you to think before coffee.

But then I dug into the retention data and saw something odd. Users who answered the check-in were only slightly more likely to be active in week 4 than users who skipped it. The signal was weak. The prompt was a nice ritual, but it wasn't earning its keep.

So we tore it apart and A/B tested twelve versions over six weeks. This is what we learned.

The setup

We had 3,200 daily active users in the test window. Each got randomly assigned to one of twelve check-in variants for their first fourteen days. Every variant asked five questions. Every variant took under two minutes. The only difference was what questions we asked.

We measured one thing: was the user active in day 21 (a full week after the test window ended)?

The nine that didn't work

Here are the question types that seemed obviously important and turned out to move retention by less than 1%:

• "How well did you sleep last night?" — every wellness app asks this. Users answer honestly. It correlates with nothing.
• "What's your energy level, 1-10?" — same story. Users report it correctly, and it doesn't predict whether they come back.
• "What's your mood today?" — five-emoji scale. Slight negative correlation, actually. Users who reported bad moods came back at the same rate as everyone else.
• "What did you eat for breakfast?" — we tried it. Didn't work.
• "What time did you wake up?" — nothing.
• "How stressed do you feel?" — nothing.
• "How focused do you feel right now?" — surprising to me, but no signal.
• "What's your biggest challenge today?" — free-text. Users wrote thoughtful answers. Zero retention impact.
• "On a scale of 1-10, how motivated are you?" — flat.

Nine questions. All useful for someone doing a research study on their own life. All useless for predicting whether they come back tomorrow.

The three that worked

Three questions had a clear positive lift on day 21 retention. All three had one thing in common: they made the user commit to something small before they closed the app.

Question 1: "What's one thing you want to remember from today?"

This one moved retention by 11 percentage points. The user writes a single sentence — a thought, an observation, a note. That's it.

Why it works: it turns the check-in from a state readout into a future artifact. Tomorrow's evening ritual references that note. The user opens the app tomorrow already knowing there's something waiting for them. It's a self-imposed hook.

Question 2: "If you could only do one workout today, which would matter most?"

Instead of asking about state, we asked about intent. The user picks from three trainer types the AI suggested. It takes ten seconds.

Why it works: users who answered this were 8 percentage points more likely to actually do a workout that day, and 14 points more likely to be active in week 4. Choice creates commitment. Commitment creates return.

Question 3: "What would make today feel like a good day by evening?"

The user writes one to three words. Not a plan, not a goal — a feeling word. "Focused." "Calm." "Not overwhelmed."

This moved retention by 9 points. Why: the answer becomes the compass for the day. When the AI writes the evening ritual, it references what the user said in the morning. The user feels seen. The system feels intelligent. Retention compounds.

The pattern

State questions (how do you feel? how did you sleep?) look useful. They aren't. Users answer them honestly, we get clean data, and nothing changes.

Commitment questions — what do you want to remember, what do you want to do, what would make today good — moved retention every time we tested them. The mechanism is not mysterious: they turn a one-way ritual into a two-way relationship. The user leaves a bookmark. The system picks it up. Tomorrow feels less random.

If you're building anything that needs users to come back tomorrow, stop measuring state. Start collecting commitments. Small ones. Ones the user is already thinking anyway, that you just help them write down.

What we shipped

The final morning check-in has three questions. That's it. The full flow takes 90 seconds on average. Completion rate went from 76% to 91%. Week-4 retention improved by 24 percentage points against the pre-experiment baseline.

The AI coach uses those three answers to shape the entire day: workout selection, evening ritual reference, next morning's greeting. Users describe it as "the app that knows what I need." What they actually mean is "the app that remembers what I said."

If you want to see how the check-in feels as a user, wow-brain.com has a 7-day free trial, no card required. You can see the three questions in action on day one.

Happy to answer questions in the comments — the test setup, the analytics, the AI reference-back mechanism, whatever's useful. And if you're running similar experiments in your own product, I'd love to see what you found.