ICE Scoring

Change one person's Confidence score from 8 to 5 and the top-ranked idea on your list drops to third. Nothing about the idea changed. That is the property of ICE scoring nobody mentions in the one-paragraph explanations.

ICE is still worth running. The value is in the argument it forces, and that value survives knowing exactly how fragile the number is.

What is ICE scoring?

ICE scores an idea as Impact x Confidence x Ease, with each of the three rated 1 to 10. The highest product goes first. It comes out of Sean Ellis's growth work and the GrowthHackers community, where it was built for ranking growth experiments rather than roadmap features.

• Impact. How much this moves the metric you care about, if it works.
• Confidence. How sure you are that it will work. Evidence only. Enthusiasm does not count.
• Ease. How little effort it takes to ship. High means cheap.

The design intent matters. ICE was made for a weekly experiment meeting where a growth team has thirty ideas and forty hours. It was never meant to arbitrate a quarterly roadmap, and most of the complaints about it come from using it for the second job.

A worked ICE example, with the arithmetic

Five ideas for a B2B signup flow, scored. The clustering is the first thing to notice.

Now the fragility. In the scoring session one person argues that the form-field evidence comes from a consumer case study and does not transfer to a B2B buyer who expects to fill in a form. Confidence on idea one goes from 8 to 5. The new score is 7 x 5 x 9 = 315, which puts it third, behind SSO at 432 and the empty state at 378.

One subjective input, moved three points by one person's opinion, reordered the top of the list. And the ranking is presented as a number, which makes it look like a measurement.

Second thing the table shows: every score sits between 144 and 504, and four of the five sit between 300 and 504. Almost nobody writes a 2 or a 3 for Impact, because scoring your own idea a 2 is an odd thing to do out loud. The real range of an ICE score in practice is narrow, so small differences in a subjective input decide the order.

Third, look at the bottom row. Migrating auth scores 9 on Impact and 8 on Confidence, which is the best evidence-weighted case on the list. It ranks last because Ease is a 2. It will rank last in every session for as long as it stays big, which is the structural work problem below.

Where does ICE scoring break down?

1. Multiplying three subjective scores creates false precision. 8 x 7 x 6 and 7 x 8 x 6 both give 336. The output has three digits and the inputs have about one digit of real resolution. Treat the score as a band rather than a rank.
2. Scores cluster high. People rate their own ideas 6 to 9. The distribution is compressed, so the differences that decide the order are inside the noise of the estimate.
3. Ease starves structural work. Anything that takes a quarter scores 2 on Ease and never reaches the top of the list. Platform migrations, data model changes and test infrastructure are exactly the work that makes everything after it cheaper, and ICE ranks them last by construction. Run them from a separate allocation rather than trying to win the score.
4. ICE has no reach term. Two ideas with the same per-user impact score the same whether they touch 40,000 users or 400. This is the single largest gap and it is the reason RICE exists.
5. Nobody recalibrates. Confidence is the one input you can check afterwards, by going back and asking whether the 9s actually worked. Teams almost never do, so Confidence stays a mood rather than becoming a track record.
6. The score gets used as a decision. The number's real job is to make three people say a value out loud and then discover they disagree about why. When the spreadsheet's sort order becomes the roadmap, you have thrown away the only part that was working.
7. It cannot express a dependency. If idea five has to ship before idea two is possible, no product of three scores will tell you that. Sequence constraints sit outside the model entirely.

ICE or RICE: which should you use?

RICE was published by Intercom, where Sean McBride wrote it up as the scoring model their product team built after finding existing models did not fit. It adds Reach and swaps Ease for Effort, as a divisor: Reach x Impact x Confidence / Effort.

The practical rule: if your ideas all touch roughly the same number of users, ICE is enough and it takes a fraction of the time. As soon as reach differs by an order of magnitude between ideas, ICE will rank them wrongly and you want RICE.

How do interviewers ask about ICE scoring?

Prioritisation questions are the most common execution question in a PM loop, and naming a model is the cheap part of the answer.

• "How would you prioritise these five things?" Score them out loud, then say which score you are least sure about. Naming your weakest input is what distinguishes an answer from a recital.
• "Engineering says the migration is the most important thing and it never gets prioritised. Why?" The Ease question. Answer with the structural-work problem and propose a fixed allocation outside the score.
• "Why would you use ICE over RICE?" Speed, and comparable reach across the candidates. If you cannot say when you would switch, the interviewer learns that you memorised two acronyms.
• "Two ideas score the same. What breaks the tie?" Good answers reach for something outside the model: reversibility, what you learn from shipping it, or a dependency that unblocks later work.

One more thing that lands well: say that you keep the scores after shipping and check whether the high-Confidence bets worked. Interviewers hear a lot of candidates describe a scoring model. Very few describe closing the loop on one.