Why prioritization is the whole game
A small team runs maybe two or three experiments a week; only high-velocity teams reach ten to twenty — and even that is a stretch when traffic is thin. With so few slots, the real cost of a mediocre test is not the test itself; it is the stronger idea you skipped to run it. A roadmap exists so each slot goes to the highest-expected-value idea, not the loudest voice in standup. Owning that ranked queue is a large part of what a growth PM does.
Build the backlog from your funnel and aha moment
Do not start from a wish list. Start from where the funnel leaks. Walk each step — visit, signup, activation, retention, revenue — and write down the biggest drop-offs; each one is a testable hypothesis. Weight the steps nearest your activation aha moment heavily, since a user who never activates rarely retains. Treat any aha you name as a correlate to validate, not a proven cause: an activation signal predicts retention, it does not prove you caused it. Each leak becomes a backlog row — the hypothesis, the metric it should move, and a rough guess at how much.
Score ideas with ICE
ICE, popularized by Sean Ellis, scores each idea on three axes from 1 to 10: Impact, how much it could move the metric; Confidence, how sure you are it will work; and Ease, how little effort it takes. Multiply the three, sort descending, and work top-down. It is fast because it is three honest guesses. That speed is also the catch: the score is an estimate, not a fact. Read a 512 and a 480 as roughly equal, not as a ruling. The number orders the list; it does not decide for you.
When reach varies, use RICE
ICE breaks down when two ideas differ wildly in how many users they touch. RICE, from Intercom, fixes that by adding Reach and dividing by Effort: (Reach x Impact x Confidence) / Effort. A tooltip seen by every visitor and a tweak buried in a settings page get scored on the same footing. The trade is speed for a reach estimate you have to source. The same warning applies — multiplying four made-up numbers can produce false precision, so keep the inputs coarse and revisit them as you learn.
| Factor | ICE | RICE |
|---|---|---|
| Formula | Impact x Confidence x Ease | (Reach x Impact x Confidence) / Effort |
| Speed | Fast — three 1-10 guesses | Slower — needs a reach number |
| Best for | Triaging a large backlog quickly | Ideas whose reach differs a lot |
| Main trap | Ease hides true effort | False precision from soft inputs |
Review on a cadence, and kill losers
Set a fixed rhythm — a weekly or biweekly review where you read results, promote winners, and retire the rest. Killing on schedule is the discipline that keeps the queue moving; a test left running is a slot left occupied. Be honest that most tests come back flat, and at low traffic many are underpowered before they start — often the right call is not to A/B test at all but to ship and monitor, as testing with low traffic covers. When you do measure, a clean read usually needs a holdout group, which a platform like customer.io or a self-hosted engagement stack can hold aside for you.