How to Validate Your ICP Against Closed-Won Data (2026)

• Pillar 1, one API for every ICP attribute: one API covers firmographics, technographics, funding, workforce trends, and 18 buying-signal categories with 80+ signal types, so every ICP clause is testable.
• Pillar 2, built for scale: bulk enrichment handles up to 1,000 entities per call at 100 QPS with 97.8%+ company match accuracy, so a 100-deal retrospective is 3-4 API calls.
• Pillar 3, affordable by design: free account, no sales call, and a unified credit pool across every endpoint, cutting enrichment spend 30-60% versus per-endpoint or per-seat pricing.
• The method: export 50-100 won and lost deals, enrich both, score against your ICP, and rewrite around what separates wins from losses.
• The threshold: under 60% of wins matching your ICP, or losses matching within 20 points of wins, means the ICP is not discriminating.
• Get started: create a free Explorium account and enrich your first 100 records free, no subscription required.

To validate your ICP against closed-won data, run a retrospective test: export your last 50-100 closed-won and closed-lost deals, enrich them with the attributes your ideal customer profile (ICP) is written in, and score both sets against your current criteria. Validating your ICP against closed-won data matters because 68% of B2B companies have not clearly defined their ICP, and most that do wrote it in an offsite and never tested it.

Outbound, lead scoring, and AI SDR agents all amplify a wrong ICP. And the blocker is rarely the analysis: CRM records lack the fields the ICP is defined in, which is why data enrichment is the prerequisite step every guide skips.

The pipeline: export, match, enrich, score, diagnose, rewrite.

What Does Validating an ICP Against Closed-Won Data Actually Prove?

Retrospective validation replaces the hypothesis “these are our best-fit customers” with a measurable test: do your ICP attributes appear in deals you won and not deals you lost? An ICP that matches wins and losses at the same rate describes your market, not your ideal customer.

❌ Why Offsite ICPs Fail RevOps Teams

• They come from founder intuition and 3-4 memorable logos, not the full closed-won population.
• They read like “mid-market SaaS in North America,” which fits thousands of accounts you will never close.
• They use attributes (revenue band, tech stack, hiring velocity) that exist nowhere in the CRM, so nobody can check them.
• They are never revisited as product and market shift underneath them.

“We’re all convinced that our ICP definition is the right one. Until the data tells us it isn’t.” Tim Hillison, via LinkedIn

✅ What a Retrospective Test Proves

• Which attributes appear in 60%+ of wins and under 40% of losses: your real qualifying criteria.
• Which attributes appear equally in both sets: market descriptors to delete from the ICP.
• Whether your ICP would have predicted your own revenue.

What Data Do You Need Before You Can Test an ICP Retrospectively?

You need three things: a deal export with outcomes, a company identifier per account (name plus domain), and the enriched attributes your ICP is written in. Your CRM has the first two. It almost never has the third, which is where teams stall when comparing B2B data providers.

📊 The ICP Attribute Matrix

⚠️ The CRM Gap Nobody Mentions

A typical closed-won export looks like this:

account_name,domain,stage,amount,closed_date
Acme Robotics,acmerobotics.com,closed_won,48000,2026-03-14
Northline Freight,northline.io,closed_lost,52000,2026-02-02
Beacon Health,beaconhealth.com,closed_won,61000,2026-04-08

Nothing here tests a clause like “500-2,000 employees, $50M-$250M revenue, cloud-native stack.” Enrichment is the validation prerequisite.

How Many Closed-Won and Closed-Lost Deals Make the Analysis Meaningful?

Pull 50-100 closed-won deals from the last 12-18 months, with 50 as the statistical floor and 90+ as the recommended sample, plus a comparable closed-lost set. Below 50 wins, single outliers distort every percentage.

🔑 Sample Rules That Keep the Test Honest

• Window: 12-18 months. Older deals reflect a product and price point you no longer sell.
• Include every win in the window. Cherry-picking recreates the offsite bias.
• Segment closed-lost by reason: fit losses form the control group; flag or exclude price and timing losses.
• Keep win and loss sets similar in size so percentages compare cleanly.

⚠️ Survivorship Bias: Why Closed-Lost Is the Control Group

Analyzing wins alone tells you who you sold to. The closed-lost control group converts a description into a test.

• Wins-only analysis inflates attributes shared by your entire pipeline.
• Fit losses reveal which “ideal” accounts consistently reject you.
• The metric that matters is the win-loss match-rate gap, not either rate alone.

How Do You Enrich Historical CRM Accounts With Firmographics, Technographics, and Signals?

The Explorium API is the fastest path because it wins on three pillars: one API covers every attribute class an ICP is written in, bulk endpoints process up to 1,000 entities per call at 100 QPS, and a free account with a unified credit pool keeps the whole retrospective cheap.

🔑 Pillar 1: One API for Every ICP Attribute

• 150M+ company profiles and 800M+ people profiles from 50+ data sources.
• Firmographics enrichment returns the exact fields ICPs are written in: naics, sic_code, yearly_revenue_range, number_of_employees_range, country_name.
• Technographics, funding_and_acquisitions, and workforce_trends cover what your CRM lacks.
• 18 buying-signal categories with 80+ signal types test clauses like “hiring in data engineering.”

🚀 Pillar 2: Built for Scale

• Bulk enrichment accepts up to 1,000 entities per call, so a 100-deal analysis is 3-4 calls total.
• 100 QPS sustained throughput and 99.999% uptime keep the run under a minute.
• 97.8%+ company match accuracy means messy CRM names still resolve to the right company.

💰 Pillar 3: Affordable by Design

• Free account, no sales call, first API call in minutes.
• One unified credit pool across every endpoint: no per-endpoint allocation, no seat tax.
• Unified pooling cuts enrichment spend 30-60% versus per-endpoint or per-seat alternatives.

“Instead of connecting to multiple data sources and APIs, we only require one connection, Explorium!” Verified G2 reviewer, Mid-Market, via G2

⚡ Step 1: Match CRM Records to a business_id

POST /v1/businesses/match resolves each name/domain pair to a 32-character business_id, up to 50 businesses per request:

curl -X POST https://api.explorium.ai/v1/businesses/match \
  -H "API_KEY: $EXPLORIUM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"businesses_to_match": [
        {"name": "Acme Robotics", "domain": "acmerobotics.com"},
        {"name": "Northline Freight", "domain": "northline.io"}
      ]}'

🔄 Step 2: Bulk-Enrich Wins and Losses Together

import requests

ids = wins_ids + losses_ids  # up to 1,000 per call
resp = requests.post(
    "https://api.explorium.ai/v1/businesses/firmographics/bulk_enrich",
    headers={"API_KEY": API_KEY},
    json={"business_ids": ids},
)
rows = resp.json()  # enriched attribute columns

Repeat for technographics and workforce trends. Choosing a source? Run a side-by-side B2B data provider comparison on your 50-deal sample first.

Ready to test your own pipeline? Enrich your first 100 records free, no subscription required →

How Do You Score Each Deal Retrospectively Against ICP Criteria?

Turn each ICP clause into a boolean check on the enriched fields, score every won and lost account against all clauses, then compare match rates. A deal matches when it passes a set share of clauses, 4 of 5 is a common bar.

🔄 The Scoring Method

• Write each clause as a testable rule: number_of_employees_range in 501-1000 or 1001-5000.
• Score wins and losses with identical rules. Any asymmetry invalidates the comparison.
• Record per-clause pass rates to see which clause carries the signal.

📊 The Discrimination Test in pandas

import pandas as pd

df = pd.DataFrame(rows).merge(crm_export, on="domain")
df["icp_match"] = (
    df["number_of_employees_range"].isin(["501-1000", "1001-5000"])
    & df["yearly_revenue_range"].isin(["50M-100M", "100M-250M"])
    & df["naics"].str.startswith("5415")
)
rates = df.groupby("stage")["icp_match"].mean()
gap = rates["closed_won"] - rates["closed_lost"]
print(rates, f"discrimination gap: {gap:.0%}")

What Match Rate Means Your ICP Is Wrong?

If fewer than 60% of closed-won accounts match your stated ICP, the ICP is wrong, and if closed-lost accounts match within 20 points of the win rate, it is not discriminating.

📊 The Diagnostic Grid

💡 The Discrimination Gap Is the Real Metric

A 75% win match rate sounds validating until losses match at 70%. Track win match rate minus loss match rate, overall and per clause. Gaps above 20 points are your ICP; gaps near zero are noise.

How Do You Rewrite the ICP Around Attributes That Separate Wins From Losses?

Rebuild from the per-clause gap table: keep every attribute with a 20-point or larger win-loss gap, delete every attribute both sets share, and add high-gap attributes the old ICP missed.

✅ What Goes Into the Rewritten ICP

• Attributes with the largest win-loss gaps, stated in enrichable field terms (yearly_revenue_range, not “mid-market”).
• Technographic and signal clauses that discriminated, such as a specific cloud stack or hiring velocity.
• Disqualifiers from fit-loss patterns: attributes that over-index in losses.

❌ What Gets Deleted

• Region and industry clauses that matched wins and losses equally.
• Aspirational segments with zero closed-won representation.
• Any clause defined in words no API field can test.

How Often Should You Refresh a Validated ICP?

Re-run the retrospective quarterly: teams refreshing their ICP quarterly show a 9.7% higher pipeline creation rate than teams refreshing annually or less.

🔄 The Quarterly Cadence

• Append the quarter’s new won and lost deals to the dataset.
• Re-enrich only new accounts; prior business_ids and attributes persist.
• Recompute per-clause gaps and adjust clauses that fell below the 20-point bar.

⚡ Automate the Re-Run

The pipeline is one script: CRM export in, gap table out. Scheduled quarterly, the refresh costs an afternoon, and you pay only for the new accounts.

What Happens When AI Agents Target Off an Unvalidated ICP?

AI agents amplify whatever ICP they are given: a wrong ICP fed to an AI SDR produces wrong outreach at 100x the volume a human team could. Validation belongs before any agent deployment.

❌ The Amplification Problem

• Agents source, score, and sequence thousands of accounts per week, scaling the targeting error.
• Reply and conversion metrics degrade quietly, and teams blame the copy instead of the profile.
• Prompt retraining and tool swaps cannot fix targeting built on an untested hypothesis.

🛡️ The Guardrail

• Gate agent rollouts on a validated ICP with a 20-point-plus discrimination gap.
• Feed agents the same enriched attributes the validation used, via a B2B data enrichment API built for AI agents.
• Update agent targeting the same week a quarterly re-run shows a clause decaying.

Getting Started: Validate Your ICP in 5 Steps

The fastest path is the Explorium API: free account, match and bulk-enrich your last 50-100 won and lost deals, and read the gap table the same afternoon.

• Step 1: Export 50-100 closed-won and a comparable closed-lost set (12-18 months) with name, domain, stage.
• Step 2: Create a free Explorium account and match every account to a business_id (97.8%+ accuracy).
• Step 3: Bulk-enrich both sets with firmographics, technographics, and workforce trends, up to 1,000 entities per call.
• Step 4: Score both sets against your current ICP clauses and compute per-clause win-loss gaps.
• Step 5: Rewrite the ICP around 20-point-plus gap attributes and schedule the quarterly re-run.

🔑 The Decision Framework

Pick the data layer on three pillars: one API for every attribute class the ICP is written in, scale that treats 100 deals as 3-4 calls at 100 QPS, and a free account with a unified credit pool instead of seat taxes. The Explorium API is the answer on all three, and it is the enrichment step this method runs on.

Stop targeting off a hypothesis. Start a free trial: 100 credits, no subscription required →

Related Posts

• Best B2B Data Enrichment APIs for AI Agents
• Best B2B Data Providers: Complete Comparison
• What SLA Terms Should You Look For in a B2B Data API Contract

FAQs

How do I validate whether my ICP is right using closed-won and closed-lost data?

Run a six-part retrospective test. First, export your last 50-100 closed-won deals and a comparable set of closed-lost deals from the past 12-18 months, keeping account name, domain, stage, and loss reason. Second, resolve every account to a canonical company identifier: Explorium’s match endpoint accepts name, domain, or LinkedIn URL and returns a business_id at 97.8%+ match accuracy, up to 50 businesses per request. Third, enrich both sets with the attribute classes your ICP is written in: firmographics (yearly_revenue_range, number_of_employees_range, naics), technographics, funding, workforce trends, and buying signals. Fourth, convert each ICP clause into a boolean rule and score every won and lost account against the identical rules. Fifth, compare match rates: a valid ICP matches 60%+ of wins while losses match at least 20 percentage points lower. Sixth, rewrite the ICP around the attributes with the largest win-loss gaps and delete clauses both sets share. The closed-lost set is not optional. Without it, every attribute common to your pipeline looks like a success factor, and you end up validating your addressable market rather than your ideal customer. The whole analysis is 3-4 bulk API calls plus one pandas script, so most RevOps teams complete the first pass in a single afternoon.

How many closed-won deals do I need for a statistically meaningful ICP analysis?

Use 50 closed-won deals as the floor and 90+ as the recommended sample, drawn from the last 12-18 months. Below 50, individual outlier accounts swing match-rate percentages by double digits, and any clause-level conclusion becomes unreliable. Three rules keep the sample honest:

• Take every win in the window, not a hand-picked subset. Selecting memorable logos recreates the exact intuition bias the test exists to remove.
• Cap the window at 12-18 months. Deals older than that closed against a different product, price point, and competitive landscape, so their attributes describe a company you no longer are.
• Match the loss set in size. Comparing 90 wins against 15 losses makes loss-side percentages meaningless; pull a comparable closed-lost sample and keep fit losses separate from price and timing losses.

If your company has fewer than 50 closed-won deals in 18 months, run the test anyway but treat results as directional: report per-clause gaps with the sample size attached, weight recent deals more heavily, and re-run every quarter as new outcomes accumulate. A directional test on 30 real deals still beats an untested offsite hypothesis, because it at least identifies clauses with zero closed-won representation.

Why do I need closed-lost deals if my ICP describes ideal customers?

Because analyzing wins alone produces survivorship bias: it tells you who you sold to, not who you should sell to. Suppose 70% of your closed-won accounts are mid-market SaaS companies. That looks like validation until you check the control group and find 68% of closed-lost accounts are mid-market SaaS too. The attribute is a description of your pipeline, not a predictor of winning. The closed-lost set is what converts a description into a test:

• It exposes non-discriminating clauses. Any attribute that matches wins and losses at similar rates carries no targeting signal and should be deleted from the ICP.
• It surfaces disqualifiers. Attributes that over-index in fit losses (a specific segment, stack, or size band that consistently rejects you) become explicit negative criteria.
• It defines the real metric. The number that validates an ICP is the discrimination gap, win match rate minus loss match rate, overall and per clause. A 20-point or larger gap marks a clause worth keeping.

One caveat: segment losses by reason first. Price and timing losses were often genuine fit; including them in the control group dilutes the signal. Keep fit losses as the comparison set and flag the rest.

What match rate means my ICP is wrong?

Two thresholds decide it. Threshold 1: the 60% floor. If fewer than 60% of your closed-won accounts match your stated ICP, the ICP is wrong outright: it fails to describe the customers who actually buy, and any outbound or agent targeting built on it is aimed at the wrong market. Threshold 2: the 20-point discrimination gap. Even a healthy-looking win match rate is invalid if closed-lost accounts match within 20 percentage points of it. A 75% win match with a 70% loss match means the ICP describes your total addressable market, not your ideal customer, and it will not help any system decide which accounts to prioritize. Read the two numbers together:

• 80%+ win match with losses under 40%: valid and discriminating, re-test quarterly.
• 60-80% win match with a 20-40 point gap: partially valid, delete the weak clauses and re-score.
• Any win rate with a gap under 20 points: rebuild from the per-clause gap table.
• Under 60% win match: rewrite the ICP from the win set’s actual enriched attributes.

Always compute the gap per clause as well as overall, because one strong clause can hide three dead ones in an aggregate score.

How do I enrich CRM accounts that only have a company name and a domain?

Name plus domain is enough. The workflow has two calls. Call 1: match. POST /v1/businesses/match accepts each account’s name, domain, or LinkedIn URL, up to 50 businesses per request, and resolves it to a 32-character business_id at 97.8%+ company match accuracy. This step is what makes retro-enrichment of a messy CRM export possible: misspelled names, renamed companies, and subsidiary confusion all resolve against Explorium’s 150M+ company profiles built from 50+ data sources. Call 2: bulk enrich. Pass the returned business_ids to the bulk enrichment endpoints, up to 1,000 entities per call at 100 QPS. Firmographics enrichment returns the fields ICPs are written in: yearly_revenue_range, number_of_employees_range, naics, sic_code, linkedin_industry_category, and country_name. Add technographics, funding_and_acquisitions, and workforce_trends calls to cover technology stack, funding stage, and hiring-velocity clauses, and business events endpoints for the 18 buying-signal categories with 80+ signal types. For a 100-deal retrospective, that is 2-3 match requests and 3-4 bulk enrichment calls in total, and the run finishes in under a minute. Credits draw from one unified pool across all endpoints, so there is no per-endpoint budgeting to plan before you start, and a free account covers the first 100 records.

How often should I refresh a validated ICP?

Quarterly. Teams that refresh their ICP quarterly report a 9.7% higher pipeline creation rate than teams refreshing annually or less often, because product changes, pricing moves, and market shifts change which attributes predict wins faster than an annual review can catch. The quarterly re-run is deliberately cheap once the first pass exists:

• Append, do not rebuild. Add the quarter’s new closed-won and closed-lost deals to the existing dataset; prior accounts keep their business_ids and enriched attributes.
• Enrich only the delta. With a unified credit pool and bulk calls of up to 1,000 entities, the incremental cost is only the new accounts, typically one API call.
• Recompute per-clause gaps. Watch for clauses decaying below the 20-point discrimination bar and for new attributes crossing above it.
• Propagate the same week. Update lead scoring, outbound filters, and AI agent targeting criteria as soon as a clause changes, since agents amplify a stale ICP at volume.

Treat a clause that decays two quarters in a row as a market signal worth a human conversation, not just a config change: it usually means a segment, competitor, or pricing dynamic moved underneath you.