The Post-Attribution Playbook for Growth — Eric Seufert, Mobile Dev Memo

Modeled on an internal Zapier document: a matrix with each organizational function as a row and levels of AI implementation as columns, from unacceptable/no implementation to full transformation. Not a top-down mandate: each team fills in what full transformation and unacceptable non-adoption look like for its function, then plans a roadmap between them and the resources needed. Opting out is not allowed. Started by agreeing company principles for AI with the founders.

Collect what competitors are running (e.g. Facebook Ads Library ads and other visible material) into an S3 bucket, then use an LLM agent to interpret the concepts and suggest why they work. Replaces the old weekly manual routine of copying competitor ads into a Google Doc and circulating it. Use the output as concept ideas to test; hallucination risk is contained because the ad platform picks winners.

Built an internal tool (Draper, ~2018) that generated all permutations of ad components and deployed them to Facebook in a constant cycle, with no person touching the output, while Facebook's value optimization (VO) found high-value audiences. Each weekly company stand-up showed the best ad of the week. Lesson: judge the process by win rate and change the inputs; do not try to explain why a given ad won.

Signal engineering: create a light hurdle for users to clear (here a CAPTCHA added only to test intent) that is a proxy for LTV; users who clear it are treated as high-value, and that event is sent back to the ad platform so it optimizes toward them in real time.

When optimizing for maximum spend under a ROAS constraint: max out the biggest channel until it hits the ROAS threshold, then move to channel two until it hits the threshold, then channel three. Only add a channel when existing ones are saturated or it creates an interaction effect (e.g. brand channels). A new channel must exceed, not match, the ROAS existing channels would get on the same budget, because it adds integration, creative-format and measurement overhead. High ROAS at tiny spend is not a reason to add it.

Instead of guessing a terminal LTV from ~10 days of data: start at modest spend (e.g. $5K/day) with a high early bar (~150% ROAS by Day 7). Let cohorts age and track Day 20/30/60 ROAS. If cohorts reach e.g. 150% by Day 30, lower the early target, change bids and grow the budget; re-check at Day 60 (e.g. 120%), raise budget again, aiming for a landing point around 110%. If a campaign can't hit ~150% Day 7 ROAS at very low spend, go back to the drawing board.

In creative testing, aim to identify losers as fast as possible: an ad getting no delivery the same or next day is not a winner and is killed, even if it might perform at average, because it won't grow spend. Winners take longer to prove out, so let them run.