MILA: about 400 recordings read with AI to find where purchases stalled
MILA Stories makes collaborative books and videos: family, friends and colleagues send audio, photos and text over WhatsApp, and it all becomes a gift for a special occasion. The underlying question was whether growth was held back by the product or by distribution. With PostHog's session recordings read by its AI agent, we reviewed the whole funnel in one pass instead of watching sessions one by one.
The challenge
Knowing whether the problem was product or distribution meant seeing what people did inside the funnel, and being able to trust the numbers.
- The printed-book funnel was losing people, but nobody knew at which step or why.
- Reviewing behavior session by session did not scale: it took reading many recordings at once and crossing them with events.
- Sales numbers did not agree with each other: Meta saw 1 in 20 sales and the dashboards counted a single payment processor.
Our approach
We used PostHog to review the product through its session recordings, read with AI and crossed with events, and measured each change against the previous period.
- 01
About 400 recordings at once
We ran PostHog's AI agent over recordings of the printed-book funnel, cross-checked with 30 days of events. It found four bottlenecks in series: half of visitors left on the first screen without touching a field, the date step stalled people (one spent more than 6 minutes), the "why?" step asked for something creative written cold, and at checkout people kept tapping the same spot on the screen, with failed payments.
- 02
A three-screen funnel
MILA's team redesigned the funnel around those steps: date and format together, and the "why?" made optional. Our part was the diagnosis and the before-and-after measurement. Price-to-pay went from 34% to 59%.
- 03
Checkout friction
Recordings and events showed buyers were the ones who struggled most with the payment screen (1 frustration click per session vs. 0.1: tapping the same spot over and over when something doesn't respond), with low errors across the funnel. Stripe's quick-pay button accounted for almost 3 in 4 declines; it was removed with no hit to sales.
- 04
What's worth a double check
As with any analysis, some things are worth a double check. The payment form lives in another company's window embedded in the site (an iframe), which recordings don't see inside, so that stretch is filled in with the processor's own data.
- 05
Data you can trust
As the foundation, we brought events, collected payments, Meta Ads, Google Ads and WhatsApp into PostHog, and counted each sale as money collected. On top of that, we sent sales to Meta through the Conversions API from PostHog: Meta went from crediting 1 in 20 to 3 in 5.
Switching the payment method confirmed the diagnosis
Recordings and events pointed to the payment step as the friction zone. In one market, MILA moved from Stripe to a local processor, and PostHog measured the result. First week vs. the previous one, with the same traffic and 3% less spend on that campaign:
Your brand could be the next case.
Tell us what you're after and we'll build a growth plan on your own accounts.
Let's talk