The case at a glance: 4.2% to 9.2% signup-to-paid
- Pre-change: 98% of signups started the trial, 5% of those ever paid
- Pre-change signup-to-paid: 4.2%
- Post-change: only 22% of signups started the trial
- Post-change trial-to-paid: 42%
- Post-change signup-to-paid: 9.2%
- Traffic: 15-20 SEO signups/day, of which 2-3 are his ICP
- LTV: 5-6+ months, by his own count
He runs a small AI SaaS solo. It reads a user's profile and generates recommended social posts, and almost all of its traffic comes from SEO. The no-card trial made sense in the era when a free user cost a database row. His product calls a model on every trial, so every tourist arrived with real cost of goods attached. One billing switch, every pre-paywall feature left intact, landing page untouched.
What he's actually selling (and why the AI SaaS trial math changed)
The product isn't a payment gate. The product is an identity layer.
In 2026, an email address plus a disposable inbox costs nothing to generate at machine scale. A working card does not. The moment he required one, he became the only operator in his category who knows which of his visitors are people. And because the card attempt happens per signup, every landing page he owns now quietly scores itself on whether it attracts humans who buy. That's a labeled dataset about his own traffic, generated free, every day.
The sequencing is what makes it survivable. The card sits one click after the first generated post — the user has already seen the thing work before being asked to prove they're real. Operators who copy "card-required trial" as a binary setting and bolt it to the signup page post the opposite result in three weeks and conclude the tactic doesn't work.
The second-order effect he undersells: his support inbox changed species. Only buyers write in now. He calls that a time saving. It's actually a roadmap correction — he stopped taking product direction from people who were never going to pay him.
He thought he ran a pricing experiment. He installed a lie detector on his own traffic and found out 78% of it was never addressable.