Quick answer
ICE is a lightweight prioritization method that rates ideas on Impact, Confidence and Ease, commonly on consistent one-to-ten scales, then combines the ratings to create a relative rank. It is associated with Sean Ellis and growth experimentation. Teams should score ideas against the same objective, population and time horizon, cite evidence for confidence and define whether Ease means inverse total effort. ICE is best for rapid triage among roughly comparable experiments. It does not calculate return, resolve strategic dependencies or make subjective estimates objective. Use risk, ethics, prerequisites and portfolio balance as explicit overlays.
What is ICE prioritization?
ICE is a fast method for ranking ideas using three judgments: expected Impact, Confidence in that expectation and Ease of obtaining the change or learning. Sean Ellis developed the approach for rapid growth work, where teams need a lightweight alternative to prolonged debate over a large experiment backlog.
A common implementation rates each factor from one to ten and multiplies the values. Some practitioners average them, which preserves the same ordering when every scale is positive and equally weighted. The formula matters less than using one documented method consistently within a comparable set.
Why growth teams use ICE
Experiment ideas arrive faster than most teams can implement or analyze them. Without a method, the loudest stakeholder, newest idea or easiest task can dominate. ICE forces a brief conversation about potential value, evidence and total work before capacity is committed.
Its speed is both strength and weakness. Ratings are estimates, and multiplying them can make the result look more scientific than the inputs. ICE should create an auditable shortlist and reveal disagreement. Strategy, dependencies, risk and research still determine what the team actually does.
Define Impact, Confidence and Ease
Impact is conditional: how much a defined metric could move for a defined population and horizon if the idea succeeds. Confidence reflects evidence for the mechanism and estimate, not a presenter's enthusiasm. Ease is the inverse of all effort required to obtain a trustworthy decision, not only coding time.
Create behavioral anchors. For example, a mid-level impact rating could represent a bounded, meaningful change rather than the word medium. Confidence levels can map to replicated experiment evidence, direct behavioral evidence, several interviews, a distant analogue or opinion. Ease anchors should use comparable person-time and elapsed-time ranges.
Impact
Estimate movement on one defined outcome if the idea works.
- For whom and by when?
- What magnitude does each anchor mean?
Confidence
Rate the strength of evidence behind the impact and mechanism.
- What evidence supports this?
- What important uncertainty remains?
Ease
Represent inverse total effort and operational complexity consistently.
- Whose work is included?
- What is the path to a valid readout?
Risk
Apply non-score constraints before an idea enters the queue.
- Could this harm customers?
- Is it reversible and compliant?
Decision
Use rank plus judgment to build a coherent learning portfolio.
- What must be learned first?
- Are ideas sufficiently comparable?
Calculate and interpret an ICE score
If each factor uses a one-to-ten scale, a team may calculate Impact multiplied by Confidence multiplied by Ease and sort descending. Keep the component ratings visible. Two ideas with the same total can have different profiles: one may be ambitious but uncertain, another modest and easy.
Do not compare scores created for different objectives or with different rubrics. A retention experiment and an unbounded infrastructure program do not become commensurable because both have numbers. Use separate lanes or a broader portfolio method when reach, cost, strategic option value or dependencies differ materially.
How to prioritize experiments with ICE
State the current constraint, target population, primary outcome and decision window. Remove mandatory compliance work and ideas that fail ethical review. Rewrite remaining items as testable hypotheses with a mechanism and minimum credible design before scoring them.
Have relevant functions score independently, cite evidence and surface large disagreements. Rank the set, then apply prerequisite, capacity and portfolio checks. Select owners and decision dates. Freeze initial ratings so hindsight does not conceal forecast error, and revisit the backlog when evidence or the constraint changes.
- Constraint and objective fixed
- Comparable idea set created
- Hypotheses include mechanism
- Impact anchors documented
- Confidence tied to evidence
- Ease includes full readout effort
- Risk screened separately
- Independent ratings captured
- Disagreements discussed
- Dependencies sequenced
- Initial scores preserved
- Forecasts calibrated after results
ICE prioritization example
Mapwell's hypothetical team first narrows the decision. Scoring every product request together would reward small interface changes and obscure larger obligations. By defining one activation outcome and excluding mandatory safety work, the remaining experiment candidates are more comparable.
The evidence notes are more valuable than a final rank. A disagreement about confidence may reveal that one person has recent usability evidence others have not seen. A low-confidence, high-impact idea may justify a cheap research step before a full experiment rather than immediate rejection.
Mapwell is a hypothetical neighborhood mobility app. New residents can compare walking, cycling and transit options, but many leave before saving a route. The team has a mixed backlog of onboarding, map and notification ideas.
Mapwell limits the scoring set to reversible experiments intended to improve eligible users saving a useful first route within seven days. Larger infrastructure work and mandatory safety fixes are handled separately.
Impact anchors describe plausible movement on that outcome; confidence anchors correspond to named evidence grades; ease includes design, accessibility, engineering, instrumentation, review, experiment duration and analysis.
A route-purpose prompt, a map legend revision and an example commute receive ratings with short evidence notes. Team members score independently, discuss large disagreements and retain ranges instead of manufacturing certainty.
The highest arithmetic score is not automatic. Mapwell checks whether a prerequisite is missing, whether the idea excludes users with limited mobility and whether the portfolio answers the most important uncertainty.
After each test, predicted impact, effort and readout quality are compared with reality. The rubric and confidence anchors change when forecasts repeatedly prove optimistic or one form of hidden work is omitted.
Mapwell, its scores and its experiments are hypothetical. Real prioritization requires organization-specific scales, evidence and risk review.
Make Confidence an evidence grade
Confidence should decline when evidence is indirect, old, poorly measured or drawn from a different audience. Direct randomized evidence about the same mechanism can rate highly, while a competitor screenshot or stakeholder intuition should rate lower. Qualitative research can strongly support a problem mechanism without estimating effect size precisely.
Attach links and note what each source establishes. Separate confidence in desirability, usability, feasibility and expected metric impact when those uncertainties differ. A single blended score can otherwise hide a fatal technical unknown beneath strong customer evidence.
Build a balanced experiment portfolio
Sorting favors easy, well-understood work and can starve foundational bets. Reserve capacity for different learning horizons: optimization of a known path, exploration of an uncertain mechanism, instrumentation repair and strategically necessary enablers. Compare ideas within each lane before allocating capacity across lanes.
Sequence tests when one result changes the value of another. Do not run interacting experiments on the same population without a design that can interpret them. Review opportunity cost and stop maintaining a large scored backlog whose assumptions have expired.
Govern scoring and experimentation
Store scorer, date, rubric version, evidence and assumptions. Limit anonymous score changes and audit whether seniority systematically shifts ratings. Product, data, engineering, design and customer-facing teams should contribute where their knowledge affects effort, risk or the mechanism.
ICE does not override privacy, accessibility, legal, security or customer-trust review. Create rejection and escalation paths outside the arithmetic. An experiment that relies on deception should not become acceptable because someone assigned it a large impact score.
Limitations and common ICE mistakes
ICE omits explicit reach, cost, dependencies, strategic fit and uncertainty ranges. Ratings are ordinal judgments often treated as cardinal numbers. Multiplication implies equal weighting and independence even though impact, confidence and ease can influence one another. These limitations make the score a prompt, not an optimizer.
Common mistakes include scoring vague ideas, changing scales mid-backlog, inflating confidence, treating Ease as development time, and automatically shipping the top rank. Use ICE when speed and comparability matter. Use RICE, cost of delay, expected value or portfolio judgment when the decision needs their additional structure.
ICE makes prioritization assumptions visible. The team still owns the evidence, risk and final decision.
Frequently asked questions
What does ICE stand for in prioritization?
Impact, Confidence and Ease. Together they provide a quick relative rank for experiment or initiative ideas.
What is the ICE score formula?
A common formula is Impact multiplied by Confidence multiplied by Ease, using consistent positive scales. Some teams average the same components; document one approach and keep it stable.
What is the difference between ICE and RICE?
RICE adds Reach and divides by Effort, making population size and effort more explicit. ICE is faster but works best among ideas with roughly comparable reach and scope.
Should the highest ICE score always be done first?
No. Apply strategic fit, dependencies, risk, capacity and portfolio balance. The rank informs a decision rather than issuing one.
How can teams reduce bias in ICE scoring?
Use anchored rubrics, independent initial scores, evidence links, cross-functional review and calibration of forecasts against completed work.
Sources and further reading
- ProductPlan: ICE Scoring Model ↗Practitioner definition, formula and caution about rating variability
- Growth Marketing Summit: Sean Ellis Presentation ↗Creator presentation describing Impact, Confidence and Ease for test prioritization
- Intercom: RICE Prioritization Framework ↗Publisher explanation of an adjacent model that makes reach and effort explicit
- Microsoft Research: Trustworthy Online Controlled Experiments ↗Primary research context for the validity checks required after an experiment is prioritized