How to Weight ICP Criteria Without Letting One Easy Metric Dominate

Who this worksheet is for

This worksheet is for B2B leaders, revenue operations teams, and researchers converting an ICP hypothesis into a scorecard. It is most useful when the current model overweights employee count, revenue, industry code, or web activity because those fields are easy to acquire. It addresses account-level criteria; it does not choose a universal contact title or replace buying-committee work.

Contents

1 Start with the decision and unit of analysis 2 Remove gates and non-criteria 3 Rate consequence, discrimination, and stability 4 Normalize the raw priorities 5 Cap proxy families and test redundancy 6 Run a sensitivity test 7 Complete the weighting worksheet 8 Illustrative example 9 Common mistakes 10 What this does not mean 11 Common questions 12 Sources and methodology

1. Start with the decision and unit of analysis

Write one sentence before discussing points: This model prioritizes [account or facility] for [defined product, application, relationship type, and commercial motion] in [geography] during [review period]. Weights have no meaning outside that sentence. A distributor recruitment model should emphasize coverage, capability, inventory behavior, and channel conflict. An end-user replacement model may emphasize installed base, operating conditions, serviceability, and replacement exposure. Define whether each field describes a firm, establishment, site, or parent. The Census Bureau's County Business Patterns methodology distinguishes an establishment at one physical location from a company that may operate several establishments. County Business Patterns methodology BLS QCEW publishes establishment, employment, and wage data by geography and industry, subject to its coverage and disclosure rules. BLS QCEW overview Neither dataset tells you automatically which establishment has the target application.

2. Remove gates and non-criteria

Do not use points for a condition that cannot be compensated. If a mandatory qualification is absent, 20 extra points for company size should not make the account eligible. Evaluate hard gates first as pass, fail, or unresolved. Also remove fields that do not explain ICP fit:

• email opened;

• advertisement clicked;

• one content download;

• one reply without fit evidence;

• a contact's seniority alone;

• a commercial database's opaque score. Those may be activity, contact, or research-routing signals. They do not establish application fit, commercial relationship, operating need, or account serviceability. Treat an industry code as a search and classification aid, not a complete ICP. The Census Bureau states that NAICS is the standard used by federal statistical agencies to classify business establishments for collecting and publishing U.S. business statistics. U.S. Census Bureau NAICS A code describes predominant economic activity at the classified unit; it does not prove use of a specific product or buying model.

3. Rate consequence, discrimination, and stability

For each compensating criterion, SellersGPT recommends three ratings: Rating dimension Question Scale Commercial consequence If this criterion is wrong, how much does it damage value, delivery, economics, or fit? 1 = limited, 5 = severe Segment discrimination How strongly does it separate the intended segment from plausible non-fits? 1 = weak, 5 = strong Meaning stability How consistently does this field mean the same thing across accounts and over the review period? 1 = volatile or ambiguous, 3 = stable Calculate: Raw priority = consequence x discrimination x stability This is a SellersGPT editorial formula. The multiplication forces a high raw value to require strength on all three dimensions. Data availability is documented separately and must not increase strategic weight. A critical but difficult criterion should create a research requirement, not lose importance. Official definitions help expose unstable proxies. The SBA's size standards vary by industry and can be expressed in employees or average annual receipts for a specific program; affiliation can also affect the calculation. SBA small business size standards dataset Therefore, "small," "mid-market," and "enterprise" should not inherit one universal employee threshold without a commercial rationale.

4. Normalize the raw priorities

After rating all eligible criteria, convert raw priorities to a 100-point model: Normalized weight = criterion raw priority / sum of all raw priorities x 100 Keep full precision during calculation. If integer display points are required, use a stated largest-remainder method: take the floor of every exact normalized share, then allocate the remaining points one at a time to the largest fractional remainders until the displayed total is 100. Predeclare a stable criterion-ID order to break an exact tie. Record any manual adjustment and why; do not silently increase a preferred criterion after seeing which accounts win. The worksheet should contain: Criterion Definition Consequenc e Discriminati on Stability Raw priority Normalized weight Evidence rule [Name] [Observable pass rule] 1-5 1-5 1-3 Product Share of total [Accepted sources] Weights express the relative contribution of assessable criteria. They do not determine whether evidence is confirmed, contradicted, or unknown. Use the evidence states separately.

5. Cap proxy families and test redundancy

Several fields may represent the same underlying idea. Revenue, employees, site count, floor area, and production lines can all proxy scale. Counting each at full weight can award the same advantage repeatedly. Group related criteria into families:

• application and process;

• commercial relationship;

• operating scale;

• service and delivery;

• market access;

• timing context;

• firmographics. SellersGPT recommends setting a maximum family share before examining account results. For example, a team may cap all generic firmographics together at 15% if they are only weak proxies for application. That threshold is a policy choice, not an official standard. Use three redundancy questions: 1 Would two criteria usually change in the same direction for the same reason? 2 Does one criterion merely restate the other's definition? 3 If one field disappeared, would the second add genuinely new decision information? If the answer suggests duplication, combine the fields, reduce the family cap, or select the more direct measure.

6. Run a sensitivity test

A model is fragile when small, reasonable weight changes reverse many decisions. Test the proposed weights before deployment: 1 Select a labeled set containing clear fits, clear non-fits, and difficult boundary cases. 2 Score it using evidence available as of the same date. 3 Increase each major weight by 20% one at a time and renormalize the others. 4 Decrease it by 20% and repeat. 5 Record changes in eligibility, tier, and order. This is a practical robustness test, not a claim of statistical significance. If one minor proxy repeatedly flips boundary accounts, inspect the criterion definition, threshold, evidence quality, and family cap. GAO's data-reliability guide emphasizes accuracy, completeness, and applicability for the intended use. GAO, Assessing Data Reliability Apply the same purpose test to every scoring input. A complete dataset can still be inapplicable to the decision; a first-party facility document can be highly applicable but limited to one site and date.

7. Complete the weighting worksheet

Use this sequence in a facilitated review: A. Decision boundary

• Unit: firm, establishment, facility, parent, or another defined record.

• Product and application.

• Relationship type.

• Geography and service model.

• Decision the model is allowed to make. B. Gate register Gate Pass rule Fail rule Unresolved action Owner [Gate] [Evidence] [Evidence] [Research or hold] [Role] C. Weight derivation Rate consequence, discrimination, and stability independently before revealing the calculated ranking. Keep evidence availability in a separate column. D. Guardrails

• family caps;

• maximum weight for any one criterion;

• minimum evidence coverage for use;

• treatment of unknown and contradiction;

• sensitivity-test acceptance rule;

• review date and owner.

8. Illustrative example

Illustrative example - fictional. A fictional industrial pump exporter weights five criteria for end-user retrofit accounts. The team rates them before seeing account scores: Criterion Consequence Discrimination Stability Raw priority Application fit 5 5 3 75 Relationship fit 4 4 3 48 Operating scale 3 3 2 18 Serviceability 5 4 3 60 Employee count 2 2 2 8 Raw total = 75 + 48 + 18 + 60 + 8 = 209. The exact normalized quotas and integer allocation are: Criterion Exact quota Floor Fractional remainder Largest-remainder points Application fit 75 / 209 x 100 = 35.8852 35 0.8852 36 Relationship fit 48 / 209 x 100 = 22.9665 22 0.9665 23 Operating scale 18 / 209 x 100 = 8.6124 8 0.6124 8 Criterion Exact quota Floor Fractional remainder Largest-remainder points Serviceability 60 / 209 x 100 = 28.7081 28 0.7081 29 Employee count 8 / 209 x 100 = 3.8278 3 0.8278 4 Total 100.0000 96 100 The floors allocate 96 points, leaving four. The four largest fractional remainders are relationship fit (0.9665), application fit (0.8852), employee count (0.8278), and serviceability (0.7081), so each receives one additional point. The reproducible displayed allocation is therefore 36, 23, 8, 29, and 4, totaling

100. Mandatory material compatibility remains a separate hard gate. The result is not a conversion forecast. It only documents relative account-fit priorities for this fictional motion.

9. Common mistakes

• Voting points by opinion: participants negotiate favored accounts rather than criterion logic.

• Rewarding availability: easy firmographics receive weight that belongs to application or service fit.

• Double-counting scale: revenue, employee, site, and line-count proxies all receive full credit.

• Hiding gates in the total: a disqualifying requirement becomes compensable.

• Changing weights after outcomes: the team tunes the model to confirm an existing list.

• Ignoring unknowns: a difficult field becomes assumed fit or silently drops out.

• Mixing record levels: parent revenue and plant conditions are scored as if they describe one unit.

• Skipping sensitivity: a fragile model reaches production without a boundary-case test.

10. What this does not mean

This worksheet does not make a subjective commercial decision mathematically objective. The ratings and caps are explicit governance choices. It does not say employee count, revenue, or NAICS is useless; these fields can be valuable when the ICP explains their causal or operational relevance. It does not establish buying intent, contact authority, or a guaranteed outcome.

11. Common questions

Must weights total 100? No mathematical law requires it, but a 100-point presentation makes shares understandable. Preserve the exact calculation and use the declared largest-remainder rule, including its tie-break rule, when displayed integer points must reconcile. Can a hard gate also carry weight? Avoid doing so for the same decision. Evaluate the gate first. If it passes, score other dimensions; double-counting the gate can exaggerate its influence. What if the most important criterion is hardest to verify? Keep its importance. Add an evidence requirement, an unresolved state, and a research action. Difficulty is not a reason to replace a decisive criterion with a weak proxy. Should historical customer data set the weights? It can inform hypotheses if customer definitions, survivorship, product mix, period, and selection effects are documented. It should not automatically determine a future-market ICP, especially when the strategy or region changes. How often should weights change? Review them when the offer, application, service model, geography, or evidence materially changes. Version the model; do not overwrite the rationale.

12. Sources and methodology

Official sources in this article define NAICS, establishments, QCEW data scope, SBA program size standards, and data-reliability considerations. The rating dimensions, multiplication formula, normalization method, family caps, and sensitivity process are SellersGPT editorial tools, not official government scoring standards. 1 U.S. Census Bureau, North American Industry Classification System. https://www.census.gov/naics/. Accessed August 24, 2026. 2 U.S. Small Business Administration, Small Business Size Standards Dataset. https://data.sba.gov/dataset/small-business-size-standards. Accessed August 24, 2026. 3 U.S. Census Bureau, County Business Patterns Methodology. https://www.census.gov/programs-surveys/cbp/technical-documentation/methodology.html. Accessed August 24, 2026. 4 U.S. Bureau of Labor Statistics, Quarterly Census of Employment and Wages Overview. https://www.bls.gov/cew/overview.htm. Accessed August 24, 2026. 5 U.S. Government Accountability Office, Assessing Data Reliability (GAO-20-283G). https://www.gao.gov/products/gao-20-283g. Accessed August 24, 2026.