ICP Modeling with Data

ICP modeling uses closed-won deal data, churn data, and customer health data to quantitatively identify which company attributes predict success. Instead of guessing "we think mid-market SaaS is our ICP," you analyze 50-100 closed-won deals and find that companies with 80-300 employees, Series A-B funding, and a sales-led motion close 3x faster and churn 60% less. The data tells you the ICP. Your job is to listen.

The principle: the ICP model is only as good as the data behind it. Minimum 50 closed-won deals for a directional model. Minimum 100 for a statistically meaningful one. If you have fewer than 50 wins, use the qualitative icp-definition-framework skill instead.

The Data You Need

Required datasets

Data fields to collect per deal/account

The Analysis Process

Step 1: Build the analysis dataset

Merge all data into one spreadsheet or database with one row per deal/account.

Columns:
company | domain | employees | industry | funding | geography |
gtm_motion | acv | sales_cycle_days | source | champion_title |
tech_stack | outcome (won/lost/churned/active) | churn_date |
expansion_revenue | health_score

Dataset rules:

• Every row must have the outcome field populated (won, lost, churned, active). Without the outcome, the row is unusable
• Enrich missing fields before analysis. A dataset with 40% blank industry fields produces weak industry insights. Enrich from Apollo or Clearbit before analyzing
• Separate "active" customers by tenure. A customer active for 6 months is different from one active for 3 years. Add a "months_active" column

Step 2: Segment and compare

For each attribute, compare win rates, churn rates, and value metrics across segments.

Analysis template (repeat for each attribute):

What this table reveals: 51-200 employees is the sweet spot. Highest win rate (64%), fastest cycle (28 days), low churn (8%), strong expansion (25%). This is the ICP size band.

Attributes to segment by

Step 3: Score and rank

Assign a fit score based on how many ICP attributes a deal matches.

Scoring example:

Then score every deal in the dataset and compare outcomes:

What this proves: Strong-fit deals (70-90) win at 9x the rate of no-fit deals, close 3.6x faster, have 4.6x higher ACV, and churn at 1/9 the rate. The ICP model works.

Step 4: Validate with holdout data

Split your data: use 70% to build the model, 30% to validate.

Validation process:

1. Build the scoring model from 70% of the data (the training set)
2. Score the remaining 30% (the holdout set) using the model
3. Compare: do high-fit deals in the holdout set actually win more, close faster, and churn less?
4. If yes: the model is validated. Deploy it
5. If no: the model is overfit to the training data. Simplify (fewer attributes, fewer segments)

Validation rules:

• If holdout win rate for strong-fit deals is within 10% of training win rate, the model is robust
• If holdout results are dramatically different (win rates don't correlate with fit score), the model is overfit. Remove the weakest attributes and re-test
• Minimum 15 deals in the holdout set for each fit tier. Below that, the sample is too small to validate

Advanced Modeling

Weighted attributes

Not all ICP attributes are equally predictive. Weight them by their correlation with the outcome.

How to weight:

For each attribute, calculate the win rate difference between the best and worst segments.

Weighting rules:

• Attributes with > 40 pp win-rate difference between best and worst segments are strong predictors. Weight them highest
• Attributes with < 10 pp difference are weak predictors. They don't discriminate between fit and non-fit. Consider removing them from the model
• Normalize weights to sum to 100. This makes the final score interpretable (0-100 scale)

Negative indicators (anti-ICP)

Some attributes are strong negative predictors. Include them as score penalties.

Anti-ICP rules:

• A strong negative indicator should be able to disqualify a deal regardless of positive scores. A competitor employee with a 90 fit score is still disqualified
• Anti-ICP attributes should be based on loss AND churn data. An attribute that predicts losses AND churn is a double negative. Weight it heavily

Multi-outcome modeling

Instead of just modeling win/loss, model multiple outcomes.

Multi-outcome rules:

• The strongest ICP attributes are those that predict positive outcomes across ALL models. "51-200 employees" predicting both higher win rate AND lower churn AND higher expansion is a triple signal. That attribute belongs in the ICP with maximum weight
• An attribute that predicts wins but also predicts churn is a trap. "Healthcare companies buy often but churn at 40%" means healthcare is not ICP despite a decent win rate. Look at the full customer lifecycle, not just the sale

Implementing the ICP Score in CRM

CRM fields

Automation

Trigger: New account created OR enrichment data updated
  ↓
Action: Calculate ICP score from enrichment fields
  Employee count → size score
  Industry → industry score
  Funding stage → stage score
  Geography → geo score
  Tech stack → stack score
  Anti-ICP checks → penalties
  ↓
Set icp_score = sum of all dimension scores
Set icp_tier based on score thresholds
Set icp_last_scored = today
  ↓
If icp_tier = "Tier 1" or "Tier 2": flag for prioritization
If icp_tier = "Not ICP": flag for review or disqualification

Implementation rules

• Score automatically on account creation. The ICP score should populate within seconds of an account entering the CRM, not after a human reviews it
• Re-score when enrichment data changes. If a company raises a new round (funding stage changes) or grows (employee count changes), the ICP score should update
• Make the score visible to reps. The ICP tier should appear on the account record, in list views, and in the lead routing logic. A score buried in a custom field nobody sees is useless
• Don't hide the scoring logic. The icp_score_details field shows how the score was calculated. When a rep asks "why is this Tier 2 and not Tier 1?" they can see the breakdown

Model Maintenance

Quarterly review process

Maintenance rules

• Quarterly is the minimum review cadence. Every quarter, re-run the analysis with fresh data. Annual reviews are too infrequent for a fast-moving SaaS company
• Track model accuracy over time. "What % of Tier 1 deals closed this quarter?" and "What % of churned customers were Tier 1?" If accuracy is declining, the model needs recalibration
• The model should get better over time, not stay static. Each quarter adds more data points. More data = better segmentation = more accurate predictions. The Q4 model should outperform the Q1 model

Tools for ICP Modeling

Tool selection rules

• Start with a spreadsheet. Export CRM data. Do the segmentation manually. You'll learn more from hands-on analysis than from any tool's automated output
• Graduate to BI tools at 200+ deals. Spreadsheets get unwieldy above 200 rows. BI tools make segmentation visual and shareable
• Statistical modeling at 500+ deals. Below 500 deals, regression models overfit. Above 500, logistic regression can identify non-obvious attribute interactions that manual analysis misses
• Predictive scoring vendors at $10M+ ARR. Below $10M, the data volume doesn't justify the tool cost. Above $10M, automated ICP scoring saves RevOps 5-10 hours per week

Measurement

Anti-Pattern Check

• Building the model from 15 deals. The sample size is too small. Patterns from 15 deals are likely noise, not signal. Wait until you have 50+ wins and 50+ losses before modeling
• Using only win data (no losses). The model identifies who buys. Without loss data, it can't identify who doesn't buy. The contrast between wins and losses reveals the discriminating attributes. Include both
• Ignoring churn data. A company type that buys easily but churns at 40% is not ICP. It's a trap. Include churn data in the model. The best ICP attributes predict wins AND retention
• Over-weighting one attribute. "100% of our wins are in the US" when 95% of prospects are also in the US. Geography isn't a discriminating factor in this case. Compare the win rate for the attribute vs the base rate
• Building the model once and never updating. The model from 6 months ago was built on different data, a different product, and a different market. Refresh quarterly with new win/loss/churn data
• Using the model to exclude leads instead of prioritize. The ICP model should determine routing priority and sales effort allocation. It should not be a hard gate that prevents non-ICP leads from ever being contacted. Tier 3 accounts can still be worked at lower priority
• No anti-ICP in the model. The model only has positive scores. A government agency with 3 positive attributes scores 45/90 and enters the pipeline. Include negative indicators that disqualify regardless of positive fit
• Scoring without enrichment. The model requires employee count, industry, and funding stage. If 40% of accounts are missing these fields, 40% of scores are wrong. Enrich before scoring