ICE Scoring for A/B Test Prioritization (with Template)

An ICE score is a single number that ranks a test idea by rating its Impact, Confidence and Ease from 1 to 10 and multiplying the three together.

Higher scores go to the top of the testing backlog.

The method was created by Sean Ellis to rank growth experiments quickly, and it works best when each score is tied to evidence instead of opinion.

This page covers how to calculate the score, what each letter should mean for an A/B testing team, and a template you can copy.

Every example uses real results from the LiftArchive test database.

If you want the wider comparison of ICE, PIE and PXL, start with our guide to A/B test prioritization.

ICE scoring: how to calculate an ICE score

You rate each idea on three questions, each on a 1-10 scale:

• Impact: if this wins, how much will it move the metric you care about?
• Confidence: how sure are you that it will win, based on evidence?
• Ease: how little time, design, development and sign-off does it take to launch?

The formula in Itamar Gilad's write-up of ICE is:

ICE score = Impact × Confidence × Ease

Scores run from 1 (1 × 1 × 1) to 1,000 (10 × 10 × 10).

Gilad notes that each team can decide what 1-10 means, as long as the rating stays consistent from idea to idea.

Multiply or average?

Some teams average the three scores instead. That changes the ranking, so pick one and stick with it.

Averaging puts the checkout rebuild first. Multiplying puts the quick copy test first, because a 1 on any factor drags the whole score down.

For A/B testing we recommend multiplying: an idea that takes two months to build, or that nobody has evidence for, should not sit near the top just because it scores well elsewhere.

The ICE framework: what each score should mean

The common complaint about ICE is that all three numbers are guesses.

Two people will rate the same idea a 4 and an 8, and the most senior person usually wins.

The fix is a written rubric, agreed before anyone scores anything.

Impact: traffic and value, not excitement

Impact should reflect how many visitors see the change and how close it sits to the purchase or signup.

A cart or checkout change touches fewer people than a homepage change, but those people are much closer to buying.

Size matters less than it feels. Oliver Kenyon removed a single line of copy from a cart page and reported +$40,413 per month (+$1.50 revenue per session).

Kasper Dolk changed one product page headline to the desired outcome and reported +$823,115.54 revenue per year. Neither change needed days of design work.

A useful rule: score Impact on where the change lives and how many buyers see it, not on how big the change is.

Confidence: score the evidence

Confidence is the score most worth getting right.

Gilad's Confidence Meter ties it to evidence types: self-conviction, industry trends and colleagues' opinions together earn only 0.1 out of 10, anecdotal customer requests 0.5, survey or market data around 1.5, qualitative user research around 3, and a tested MVP with real usage 6.5. The point is that opinions barely count.

For A/B testing, the closest thing to strong evidence is a pattern of past tests. Here is a rubric that uses that:

Outcome-led headlines are a good example of a 7-8. Besides Kasper Dolk's result, Vasta tested an outcome-led hero headline on a supplement landing page and saw +35% completed purchases, with only the headline changed.

You can browse the rest in product page headline and copy tests.

One warning when you use any database, including ours: published results lean toward winners.

Of the 1,724 results in LiftArchive, 1,447 are marked as wins, about 84%.

Clean Commit publishes its full test library, and of its 173 results in our database only 16 are wins, 35 lost and 122 were flat.

Your real win rate will look more like the second number, so keep Confidence scores modest unless the evidence is your own.

Ease: count the days

Ease is the easiest score to make objective.

Convert effort into a fixed scale, for example: 10 for a copy change shipped in under a day, 7 for a few days of design and front-end work, 4 for one to two weeks, and 1 for a multi-week rebuild that needs developer time and approvals.

Ease also protects you from big redesigns, which are expensive and hard to learn from.

Clean Commit rebuilt a hockey landing page on the layout of a better-performing page for another sport and saw -38.5% conversion rate and -42.7% revenue per visitor over 29 days.

Dave Diederen compared both routes on one store: adding urgency messaging to checkout produced +6.59% revenue per visitor, while the full checkout redesign lost (-1.18% conversion rate).

A

→

B

+$40,413 per month

Removed a single line of copy from the cart page

CartHeadlines, product titles & copy@oliverkenyon

A

→

B

+11.88% revenue per visitor

Shortened guarantee copy ('or your money back' instead of 'or get 100% of your money back')

Product pageMoney-back guarantee & returns@DaveDiederen

A

→

B

+35% purchases

Outcome-led hero headline on landing page

Landing pageHeadlines, product titles & copyVasta

A

→

B

+6.59% revenue per visitor

Added urgency messaging to checkout (instead of a full checkout redesign)

CheckoutSales, urgency & limited-time offers@DaveDiederen

A

→

B

+$823,115.54 revenue/year

Change the product hero headline to the desired outcome

Product pageHero & above-the-fold section@kasper_dolk

A

→

B

-38.5% conversions

Hockey page rebuilt on another sport's layout

Landing pageLayout, fonts & visual designClean Commit

ICE prioritization in practice: template and worked example

Here is a backlog for an ecommerce store, scored with the rubrics above.

Copy the columns into a spreadsheet and add a row per idea.

The guarantee row uses Dave Diederen's test, where shortening "or get 100% of your money back" to "or your money back" gave +13.08% conversion rate and +11.88% revenue per visitor.

One test is a reason to try it, not proof it will work for you, which is why Confidence is a 6 and not a 9.

The redesign row scores 54 even with the highest Impact.

That is ICE doing its job. If the team still wants the redesign, break it into parts and score each one, so a win tells you what caused it.

Write each row as a proper test hypothesis so the Confidence score has something to point to.

Keep the scores honest

• Score alone, then compare. Have two or three people score each idea separately and average their scores for each factor before multiplying. Big disagreements are worth a short discussion.
• Re-score after each test. A win on a page raises Confidence for follow-up ideas there. A loss lowers it. Log the outcome in a shared test log so the next round starts from evidence.
• Check traffic before you commit. A high ICE score means little if the page can't reach a useful minimum detectable effect in a few weeks. On low-traffic sites, favor bold changes on high-traffic pages.
• Use the top of the list to build a roadmap. Take the top five to ten ideas into your testing roadmap and fill gaps between long tests with quick, high-Ease ones.

When ICE is the wrong tool

ICE ranks individual ideas. It does not tell you which page to work on, so many teams use the PIE framework to pick the page first.

If your team keeps arguing about scores even with a rubric, PXL replaces 1-10 ratings with mostly yes/no questions, which leaves less room for opinion.

ICE is also a poor fit for changes you will ship anyway, such as legal or accessibility fixes, and for ideas you can't test at all.

Those belong in a separate queue. For everything else, ICE is a fast way to sort a long list, as long as Confidence comes from tests and research rather than from whoever argues hardest.

For the rest of the testing process, from hypothesis to analysis, see the A/B testing hub.