sales-icp-definition

{
  "skill_name": "sales-icp-definition",
  "evals": [
    {
      "id": 1,
      "prompt": "We're Kettlebridge, B2B SaaS selling compliance training to mid-size logistics firms. I've got 23 closed-won deals in the CRM, all tagged clean with won/lost reasons plus churn flags. I want a regression-based weighted scoring rubric so we can score accounts out of 100. Our head of analytics can run the regression herself, she did it at her last company, so effort isn't the blocker. I need it by Friday because we're carving territories for Q4 and the reps need a list.",
      "expected_output": "A refusal to build regression weighting at 23 deals, routed instead to founder-led discovery plus an equal-weighted checklist, with graduation waypoints, an anti-ICP, and a dated plan to retro-score at ~50 deals.",
      "files": [],
      "expectations": [
        "States regression-based weighting is unavailable at 23 closed-won deals and names the 100-150+ closed-won threshold it requires",
        "Routes the work to founder-led discovery paired with an equal-weighted checklist, the prescribed treatment for the ~20-50 deal band",
        "Produces no rubric in which criteria carry differentiated numeric weights",
        "Names the Friday territory-carve deadline as an independent reason to demote regression, separate from the deal-volume constraint",
        "Acknowledges the in-house analyst lowers regression's effort cost, yet still declines to promote regression on that basis",
        "States the graduation waypoints - ~10-20 deals for a qualitative hypothesis, ~50 for an equal-weighted rubric, 100-150+ for regression - as one path rather than competing thresholds",
        "Includes an anti-ICP naming accounts disqualified despite looking attractive on paper",
        "Commits to a dated plan to run retro-scoring at ~50 deals instead of claiming the checklist is already validated",
        "Delivers both a one-paragraph ICP statement and a pass/fail criteria checklist, not one without the other",
        "Asks clarifying questions one at a time before proposing an artifact rather than returning it as one finished block"
      ]
    },
    {
      "id": 2,
      "prompt": "Nordhaven Systems here. Our ICP today is US mid-market third-party logistics providers, 50-400 employees, running a legacy warehouse management system. We just closed our first two deals in Germany with automotive manufacturers and leadership wants to go after that market properly next year. Can you widen the ICP so German automotive fits inside it? I really don't want two separate documents floating around, the field gets confused enough as it is and the SDRs will just work whichever one they opened last.",
      "expected_output": "A separate ICP for the German automotive market rather than a widened existing one, with the primary/secondary distinction, a per-segment rubric, and a version/changelog entry.",
      "files": [],
      "expectations": [
        "Builds a separate ICP for the German automotive-manufacturing market instead of widening the existing 3PL profile",
        "States that entering a new market produces a separate ICP, never a broadened one",
        "Names market entry as an off-cycle trigger that overrides the normal refresh calendar",
        "Prescribes one rubric per major segment rather than one rubric stretched across both markets",
        "Answers the \"two documents will confuse the field\" objection directly rather than complying with it",
        "Distinguishes a primary segment from an acceptable secondary segment so the field can tell \"out of focus\" from \"disqualified\"",
        "Flags \"German automotive manufacturers\" as a market segment rather than an ICP, because on its own it disqualifies nothing",
        "Records a version number and a changelog entry describing what changed and why",
        "Notes the new-market profile cannot inherit the 3PL rubric's weights, because the German deal history that would justify them does not exist yet",
        "Names an owner for each of the two profiles"
      ]
    },
    {
      "id": 3,
      "prompt": "I'm VP Sales at Brightfold and the board deck is due Thursday. I want one slide proving that tightening our ICP pays off. Here's what I've collected: SiriusDecisions found companies with a defined ICP get 68% higher win rates; New Breed grew average deal size 83% after narrowing; Lavu went from $10M to $40M+ ARR after focusing. Our consultant says we should expect win rates of 60-65% once we're focused, we're at 24% today on roughly $30K deals. My RevOps lead also found a stat that quarterly ICP refreshers convert 20-35% better from MQL to closed-won than annual ones. Write the slide.",
      "expected_output": "A slide that grades each supplied figure - rejecting the 68% orphan citation, labelling the self-reported and unverified case studies, sanity-checking the 60-65% projection against the 20-30% baseline - and substitutes the company's own ICP-fit win-rate comparison as evidence.",
      "files": [],
      "expectations": [
        "Refuses to present the 68% higher-win-rate figure as fact",
        "Identifies the 68% figure as an orphan citation never traced to a published report despite its attribution",
        "Labels the +83% average-deal-size figure as a first-party self-report published by the company itself",
        "Labels the $10M-to-$40M+ ARR figure as a claimed outcome that is not independently verified",
        "Checks the consultant's 60-65% projection against an independent mid-market B2B SaaS win-rate baseline of roughly 20-30%",
        "Declines to adopt the 60-65% projection as an expected outcome",
        "Keeps the direction of the quarterly-versus-annual refresh claim while dropping its 20-35% figure as a single-source vendor number",
        "Attaches a source and a year to every benchmark that survives into the slide",
        "Invents no replacement statistic in place of the rejected 68% figure",
        "Proposes the company's own win rate on ICP-fit versus non-fit accounts as the evidence to show the board instead"
      ]
    },
    {
      "id": 4,
      "prompt": "Solterra Home. We install residential solar, average job is $24,000, about 900 completed installs over four years. We also run a commercial arm doing rooftop systems for small businesses, roughly 40% of revenue. Marketing wants an ICP doc so they can go buy a list: company size bands, industry codes, revenue brackets, the same shape we used at my last company. Build it for us.",
      "expected_output": "Two separate profiles - a demographic/psychographic household profile for the residential side treated as a high-consideration purchase with distributed decision roles, and a separate B2B profile for the commercial arm - with the firmographic list-buying request declined for the residential side.",
      "files": [],
      "expectations": [
        "Builds a demographic and psychographic profile of the person or household for the residential side rather than firmographic criteria",
        "States that B2C has no firmographic layer from which to build a company-level ICP",
        "Classifies a $24,000 residential solar purchase as high-consideration rather than low-involvement",
        "Maps household decision roles - initiator, influencer, decider, buyer, user, gatekeeper - for the residential side",
        "Notes those roles shift from purchase to purchase rather than being fixed per household",
        "Produces two separate profiles, one per side of the business, rather than one blended profile",
        "States explicitly that a company selling both B2C and B2B gets a profile per side, never a blend",
        "Declines to supply the requested firmographic list-buying criteria for the residential segment",
        "Builds the residential profile from actual completed-install data rather than assumed demographics",
        "Carries disqualifiers, a validation step and a refresh cadence into the B2C profile as well as the B2B one"
      ]
    },
    {
      "id": 5,
      "prompt": "Verrick Labs, we sell clinical-trial site management software, 180 closed deals in the CRM over three years. Here's the fit model I drafted, 100 points total: 30 pts if the company has 200-2000 employees, 25 pts if they're biotech or pharma, 20 pts if they raised a Series B or later in the last 9 months, 15 pts if a contact has opened 3 or more of our emails or booked a demo, 10 pts if they're US-based. Score of 60+ means work the account. Do the weights look right to you?",
      "expected_output": "A critique separating the funding trigger from qualification, moving the contact-engagement criterion out to lead scoring, expanding the criteria set into the 8-15 range with technographic coverage, and adding first-class disqualifier deductions mined from matched-but-lost deals.",
      "files": [],
      "expectations": [
        "Identifies the Series B funding criterion as a buying trigger and states triggers decide which qualified accounts to pursue first, not which accounts qualify",
        "Keeps the funding signal in the rubric's intent category as a scoring signal rather than deleting it outright",
        "Requires the artifact to state the qualify-versus-prioritize distinction explicitly",
        "Identifies \"opened 3 or more emails or booked a demo\" as contact-level engagement readiness rather than account-level structural fit",
        "Removes that criterion from the ICP rubric and assigns it to lead scoring",
        "States the ICP scores the account on structural fit before any engagement",
        "Flags five criteria as too few and names 8-15 criteria as the target range",
        "States that fewer than 8 criteria usually cannot differentiate accounts",
        "Adds disqualifiers as first-class point deductions rather than notes or afterthoughts",
        "Sources candidate disqualifiers from lost deals that matched the firmographic criteria",
        "Notes the draft carries no technographic criteria and adds that category"
      ]
    },
    {
      "id": 6,
      "prompt": "Palladane. We retro-scored 80 closed deals against the new fit rubric. Results: accounts scoring 70+ won 61% of the time, 45-69 won 58%, under 45 won 54%. Pretty consistent across the board, which feels like a good sign to me, the model looks stable. We weighted everything against average contract value since that's the number the CFO cares about. We sell a $12K mid-market plan and a $300K enterprise deal off this one rubric. Can we roll it out to the SDR team Monday?",
      "expected_output": "A rejection of the rollout: the flat band spread shows the rubric is not predictive, the weights get revised and re-run against a blended target variable, and the mid-market and enterprise motions get separate rubrics.",
      "files": [],
      "expectations": [
        "Reads the 61%/58%/54% spread as no material separation between bands, contradicting the user's \"stable model\" interpretation",
        "Declines to roll the rubric out to the SDR team",
        "Prescribes revising the weights and re-running the retro-scoring pass rather than shipping and monitoring in production",
        "States a rubric whose high scorers do not close at a materially better rate than low scorers is not predictive yet",
        "Identifies average contract value as the wrong sole target variable",
        "Prescribes weighting against a blend of outcomes - win rate, retention, expansion, time-to-value - rather than contract value alone",
        "Warns that optimizing on deal size alone rewards large but poor-fit logos",
        "Separates largest customers from best customers as two different axes",
        "Prescribes one rubric per major segment for the $12K mid-market and $300K enterprise motions rather than one stretched across both",
        "Requires the retro-scoring result to be reported inside the ICP artifact itself"
      ]
    },
    {
      "id": 7,
      "prompt": "Mesa Loop, pre-Series-A dev tooling company, 9 paying customers. Three are design partners, founders I've known for years, and they give us by far the clearest feedback, so I built the profile off them: Series A to B dev tool companies, 30-80 engineers, Go and Kubernetes shops. The odd thing is our loudest inbound is two enormous fintechs with thousands of engineers, completely outside that profile, so we've been politely ignoring them. Growth has been flat two quarters running and the board wants us to open up to broader mid-market. Draft me the ICP.",
      "expected_output": "A founder-led hypothesis that treats the design-partner sample as biased and Low confidence, follows the fintech pull signal instead of ignoring it, refuses the board's broadening instruction, and ships an anti-ICP alongside.",
      "files": [],
      "expectations": [
        "Flags the three long-known design partners as a biased sample and names the overfitting risk",
        "Prescribes interviewing beyond the design partners before the hypothesis is written",
        "Grades design-partner-only claims as Low confidence",
        "Treats the first five arms-length deals as the hypothesis's real test",
        "Treats the unexplained fintech inbound as a pull signal to follow even though it contradicts the planned profile, rather than noise to ignore",
        "Captures early-adopter answers verbatim and notes their exact phrasing doubles as future messaging",
        "Rejects the board's instruction to broaden and recommends sharpening the profile instead",
        "States that widening an ICP because growth targets were missed is usually the wrong direction before Series B",
        "Produces an anti-ICP naming who is disqualified despite looking attractive on paper",
        "Warns the user that turning down attractive-looking \"maybe\" deals will feel costly pre-revenue"
      ]
    },
    {
      "id": 8,
      "prompt": "Havenlight, Series C. Our ICP doc says \"mid-market B2B companies in North America with 100-1000 employees.\" Reps chase literally everything, the SDR target list is 41,000 accounts, and MQL-to-SQL has slid to 11% so the AEs have basically stopped working marketing leads. The doc was written 14 months ago in a two-day offsite by eight people and nobody owns it now. There's also a batch of accounts that score perfectly on the sheet and just never move, we've worked some of them for a year. Fix this.",
      "expected_output": "A diagnosis against the documented failure modes - segment-not-ICP, no owner, stale scoring below the ~15% tripwire - then a rebuild with a named owner, stacked refresh cadences, off-cycle triggers, and switching forces to explain the stuck high scorers.",
      "files": [],
      "expectations": [
        "Diagnoses \"mid-market B2B companies in North America with 100-1000 employees\" as a market segment rather than an ICP",
        "States the current statement is too large to disqualify anything",
        "Applies the check of whether the profile disqualifies roughly 70% of inbound",
        "Applies the check that an ICP should be a small fraction of TAM, typically 5-15%",
        "Reads the 11% MQL-to-SQL rate against the ~15% tripwire and calls the scoring stale",
        "Assigns a named owner and identifies the committee-written, unowned artifact as a documented failure mode",
        "Prescribes all three refresh cadences together - quarterly light review, annual full rebuild, monthly drift tripwires - rather than choosing one",
        "Lists off-cycle triggers that override the calendar cadence",
        "Adds a version number and changelog to the rebuilt artifact",
        "Explains the perfectly-scoring, non-moving accounts through switching forces - push, pull, habit, anxiety, from Bob Moesta's jobs-to-be-done switching framework - rather than by adjusting their fit scores",
        "States that fit explains who could buy while the switching forces explain why a fit account still does not move"
      ]
    },
    {
      "id": 9,
      "prompt": "Rowanmere Analytics. 60 closed-won deals, all tagged. The problem is there's no firmographic pattern at all: our customers run from 40 people to 9,000, across insurance, gaming, logistics, everywhere, US and EU. What we do have is 40 recorded win/loss interviews and they're rich, our head of CS ran them properly. Separately, nobody internally agrees on what makes us different from Quillstack, our main competitor. Ask three people, get three answers. We need the ICP before planning season starts.",
      "expected_output": "A profile derived through SPICED from the interview material rather than from firmographics, with Dunford's positioning-first sequence run because differentiation is unclear, analyst-defined weighting at 60 deals, and candidate hypotheses presented for approval before building.",
      "files": [],
      "expectations": [
        "Selects SPICED as the derivation method, given rich win/loss interview material and weak firmographic patterns",
        "Names all five SPICED elements: situation, pain, impact, critical event, decision",
        "States SPICED reverse-engineers the profile from why customers actually bought rather than starting from firmographics",
        "Invokes April Dunford's positioning-first approach because the company's differentiation is unclear",
        "Works Dunford's steps in order: competitive alternatives, unique attributes, the value those attributes create given the alternatives, the customers who care most about that value, market category",
        "States that positioning and the ICP are derived together rather than sequentially",
        "Notes a fuzzy ICP is often a positioning problem rather than a targeting problem",
        "Sets analyst-defined weighting as the rung, because 60 closed-won deals sits below the 100-150+ regression bar",
        "Anchors the profile on the problem and the value delivered rather than on firmographic descriptors",
        "Presents 2-3 candidate ICP hypotheses with trade-offs and one explicit recommendation, and gets approval on one before building the artifact"
      ]
    }
  ],
  "trigger_queries": [
    { "query": "Define our ICP", "should_trigger": true },
    { "query": "Who should we be selling to?", "should_trigger": true },
    { "query": "Build an ICP scoring rubric for our sales team", "should_trigger": true },
    { "query": "Our reps are chasing everything and closing nothing", "should_trigger": true },
    { "query": "help me figure out our ideal customer profile", "should_trigger": true },
    { "query": "We're entering the Nordics next quarter, do we update the customer profile?", "should_trigger": true },
    { "query": "What does our best-fit customer actually look like?", "should_trigger": true },
    { "query": "Our target account list is 60,000 rows and the SDRs hate it", "should_trigger": true },
    { "query": "Write the ideal customer profile for our Series A fintech", "should_trigger": true },
    { "query": "I need a pass/fail checklist reps can use to decide if an account is worth working", "should_trigger": true },
    { "query": "Which kinds of companies actually keep renewing with us?", "should_trigger": true },
    { "query": "We keep winning deals that look nothing like who we said we'd sell to", "should_trigger": true },
    { "query": "Our targeting is too broad", "should_trigger": true },
    { "query": "Can you put together a buyer profile document that marketing and sales can share?", "should_trigger": true },
    { "query": "how do we decide which accounts are out of profile", "should_trigger": true },
    { "query": "Draft the anti-ICP for our startup", "should_trigger": true },
    { "query": "Our ICP hasn't been touched in over a year, what do we do about it", "should_trigger": true },
    { "query": "Define the target customer for our residential solar business", "should_trigger": true },
    { "query": "We sell to everyone and it isn't working", "should_trigger": true },
    { "query": "Set up a weighted fit score for accounts before any engagement happens", "should_trigger": true },
    { "query": "What should disqualify an account even when it looks good on paper?", "should_trigger": true },
    { "query": "MQL to SQL dropped to 12% and the reps are ignoring marketing leads entirely", "should_trigger": true },
    { "query": "Help me write the one-paragraph statement of who we sell to", "should_trigger": true },
    { "query": "how often should we revisit who we're targeting", "should_trigger": true },
    { "query": "We only have 15 customers, can we say anything about the pattern yet?", "should_trigger": true },
    { "query": "The board asked who our ideal customer is and I gave three different answers last week", "should_trigger": true },
    { "query": "Profile our best customers so we can go find more like them", "should_trigger": true },
    { "query": "Should we narrow down or open up who we go after?", "should_trigger": true },
    { "query": "our founder says we sell to modern product teams, that feels way too vague", "should_trigger": true },
    { "query": "Build the demographic and psychographic profile of our typical buyer", "should_trigger": true },
    { "query": "I want to know which accounts we should stop wasting time on", "should_trigger": true },
    { "query": "Turn our closed-won data into targeting criteria", "should_trigger": true },
    { "query": "define target customer", "should_trigger": true },
    { "query": "what's the difference between our ICP and our personas", "should_trigger": true },
    { "query": "We sell both B2B and B2C, do we need two customer profiles?", "should_trigger": true },
    { "query": "We just changed our pricing and packaging, does that change who we should target?", "should_trigger": true },
    { "query": "Give me the criteria marketing should use to build the account list", "should_trigger": true },
    { "query": "who buys this product and why do they buy it", "should_trigger": true },
    { "query": "put together the customer fit criteria before we do the territory carve", "should_trigger": true },
    { "query": "Our scoring model says these accounts are perfect but not one of them moves", "should_trigger": true },
    { "query": "We landed a huge logo that churned in a year, how do we stop repeating that?", "should_trigger": true },
    { "query": "Design our account segmentation model with fit and readiness as separate axes", "should_trigger": false },
    { "query": "How do we split accounts into segments for coverage planning?", "should_trigger": false },
    { "query": "Map the whitespace across our existing account base", "should_trigger": false },
    { "query": "Where should the Tier 1 and Tier 2 cutoffs sit on our fit score?", "should_trigger": false },
    { "query": "How many accounts should each AE carry in each tier?", "should_trigger": false },
    { "query": "Design the touch cadence and QBR frequency per account tier", "should_trigger": false },
    { "query": "Estimate our TAM, SAM and SOM for the European market", "should_trigger": false },
    { "query": "How big is the market for warehouse robotics software?", "should_trigger": false },
    { "query": "Size the market bottom-up from our named account list", "should_trigger": false },
    { "query": "Build a lead scoring model based on email opens and pricing page visits", "should_trigger": false },
    { "query": "Score inbound contacts on how ready they are to talk to sales", "should_trigger": false },
    { "query": "Should we run product-led or sales-led growth?", "should_trigger": false },
    { "query": "We're exiting founder-led sales, what motion comes next?", "should_trigger": false },
    { "query": "Set quotas for next year using ramped rep equivalents", "should_trigger": false },
    { "query": "How much over-assignment cushion should we build into quotas?", "should_trigger": false },
    { "query": "Turn these prospect signals into personalization angles for my outreach", "should_trigger": false },
    { "query": "Write a personalized opening line for this prospect based on their funding round", "should_trigger": false },
    { "query": "Build a discovery question set for a 30-minute first call", "should_trigger": false },
    { "query": "What should I ask to uncover pain on a qualification call?", "should_trigger": false },
    { "query": "Score this deal against MEDDPICC", "should_trigger": false },
    { "query": "Is this opportunity actually qualified? Here are my call notes", "should_trigger": false },
    { "query": "Who's the economic buyer in this deal based on these call notes?", "should_trigger": false },
    { "query": "Map the stakeholders on the Ridgeline opportunity", "should_trigger": false },
    { "query": "Review these deal notes for qualification red flags", "should_trigger": false },
    { "query": "Should we run pods or an assembly line sales org?", "should_trigger": false },
    { "query": "What's the right SDR to AE ratio at our stage?", "should_trigger": false },
    { "query": "Design the pay mix and accelerators for our AE comp plan", "should_trigger": false },
    { "query": "Write the interview loop and scorecard for hiring SDRs", "should_trigger": false },
    { "query": "How do I get promoted from SDR to AE?", "should_trigger": false },
    { "query": "Generate and score subject line variants for this cold email", "should_trigger": false },
    { "query": "Check my SPF and DKIM setup before we start sending", "should_trigger": false },
    { "query": "Write me a 20-second cold call opener for a CFO", "should_trigger": false },
    { "query": "Plan a 12-touch outbound cadence across email and LinkedIn", "should_trigger": false },
    { "query": "They said we're too expensive, what do I say back?", "should_trigger": false },
    { "query": "How much pipeline do we need to hit a $4M quarter?", "should_trigger": false },
    { "query": "Turn these call notes into a recap email with next steps", "should_trigger": false },
    { "query": "Plan which concessions I can trade in this renewal negotiation", "should_trigger": false },
    { "query": "Score this call transcript against our call rubric", "should_trigger": false },
    { "query": "Build the ROI and business case for this deal", "should_trigger": false },
    { "query": "Write persona messaging for our VP of Engineering buyer persona", "should_trigger": false },
    { "query": "Build a customer health score to flag churn risk in our existing accounts", "should_trigger": false }
  ]
}

# Worked example - founder-led ICP discovery sequence

The bottom-up path for a company with too few closed deals to derive anything statistically. The sequence turns a handful of early customers into a testable ICP hypothesis plus an anti-ICP, and names the waypoints where it hands off to the data-led branch.

## The sequence

1. **Write down the current (too-broad) statement first.** "We sell to modern product teams" - capturing it makes the narrowing visible and auditable later.
2. **Narrow to an intersection of specific traits** - size, industry, growth stage, tech stack, org characteristic - that plausibly correlates with the deals actually closed. The discipline: the more precisely who-has-the-problem-and-why is defined, the less time goes to conversations that were never going to convert. "Selling to everyone equals closing no one."
3. **Interview every early adopter, structured, verbatim.** Three core questions:
   - Why did you specifically select this product - what were the alternatives, and what tipped it?
   - What benefit have you actually realized, in your own words? (Keep the exact phrasing - it is the future messaging.)
   - How do you use it day to day - who touches it, how often, for what?
4. **Read the pull signal.** Look for disproportionate enthusiasm from a recognizable account shape rather than deciding top-down, and follow it even when it contradicts the plan. The sourced case: Sprig's founder noticed the strongest inbound pull came from companies with millions of users (Thunkable, Square, Robinhood among early customers). The ICP was read off who kept pulling hardest, even though targeting large companies as an early-stage startup was itself a hard sell.
5. **Write the anti-ICP.** Alongside the narrowed profile, list who is disqualified _even though they look attractive on paper_ - e.g. the well-funded logo outside the pattern. This converts the narrowing instinct into a document, which is what holds when the first tempting "maybe" deal appears - expect it to feel costly pre-revenue. Saying no to maybe-deals is the named founder-stage failure mode, so warn the user it will feel wrong before it works.
6. **Ship the hypothesis in the standard artifact shape** - one-paragraph statement plus pass/fail checklist, verbatim customer language, anti-ICP, confidence grades on anything unverified.
7. **Set the graduation waypoints inside the artifact:**
   - ~10-20 deals: a qualitative hypothesis should exist (this document).
   - ~50 deals: replace it with an equal-weighted rubric (there is finally enough signal for a mechanical checklist, still not enough to differentiate weights).
   - 100-150+ closed-won: regression-based weighting becomes available.

   One path, three waypoints, not competing thresholds.

## Example hypothesis output (abridged, illustrative)

> **ICP hypothesis (v0.2, confidence: Medium):** Product-led B2B software companies, 20-200 employees, with a live self-serve funnel and at least ~50K monthly active users, where a growth or product lead (not marketing) owns activation metrics. They come to us after an activation-rate plateau, having outgrown spreadsheet analysis but unwilling to staff a data team.
>
> **Checklist:** live self-serve funnel (pass/fail) · ≥50K MAU (pass/fail) · growth/product owns activation (pass/fail) · no dedicated data team (pass/fail) · plateau or regression in activation in last 2 quarters (pass/fail).
>
> **Anti-ICP:** enterprise suites wanting a BI replacement; agencies buying for clients; pre-launch startups with no funnel data yet - all three have shown pull, all three churned or stalled in evaluation.

## Why narrowing compounds

Iterating inside a narrow ICP feels repetitive but compounds message sharpness per conversation. Spray-and-pray produces shallow insight per conversation.

The sourced case for the payoff: narrowing to a precise vertical niche preceded an ~85% lead-to-MQL conversion rate over a 14-month program versus the diffuse baseline. It is a single practitioner-reported program, so treat it as an existence proof of the mechanism, not an expected value.

## Negative example - overfitting to design partners

Building the hypothesis from two or three friendly design partners who do not represent the broader addressable pattern. It is the qualitative twin of the data-led branch's "largest customers mistaken for best customers" failure: the evidence is real but the sample is biased toward whoever was easiest to reach.

Countermeasures:

- Interview _beyond_ the design partners before writing the hypothesis.
- Grade design-partner-only claims as Low confidence.
- Treat the first five arms-length deals as the hypothesis's real test.

# Named frameworks and benchmark grading

The named ICP-definition methods worth citing by name, and how much weight each figure this skill uses deserves.

- Cite the well-sourced figures confidently.
- Flag the folklore explicitly.
- Never launder a vendor claim into a fact.

## Frameworks

- **April Dunford - positioning-first (_Obviously Awesome_).** Reaches the ICP through positioning, worked in order:
  1. Competitive alternatives.
  2. Unique attributes.
  3. Value those attributes create given the alternatives.
  4. Target customers who care most about that value.
  5. Market category.

  The ICP output is "characteristics of customers who care a lot about your differentiated value" - positioning and ICP are derived together, not sequentially. Use when the company's differentiation is unclear, because a fuzzy ICP is often a positioning problem wearing a targeting costume.

- **Winning by Design - ICP blueprint via SPICED** (Situation, Pain, Impact, Critical event, Decision). Reverse-engineers the profile from _why customers actually bought_ rather than starting from firmographics. Use when interview/deal-review material is rich but firmographic patterns look weak.
- **a16z - Five-Question self-test (Michael King, 2025).** A structured test of whether a company _really_ knows its ICP, spanning who (size, business type), product fit, and process/pricing fit - designed to expose a fuzzy or aspirational ICP before it gets operationalized. Also the source of the target-variable guidance: weight the rubric against a blend of cohorts, never contract value alone. Use as the entry diagnostic when a team insists its ICP already exists.
- **HubSpot / Salesforce - template-led five-step process.** Codified analyze-your-best-customers templates, operationally prescriptive rather than first-principles. Use it as a starting scaffold for teams that want a form to fill, and pair it with the retro-scoring validation those templates omit.
- **Bowery Capital / OpenView lineage.** Early practitioner content on customer-profile templates in a startup-sales context, predates the current vendor-content wave. Useful as background on where the templates came from, not as a distinct method.
- **Revenue-intelligence tooling (integration note).** Platforms in this category operationalize the regression end of the weighting ladder from first-party conversation and CRM data. Treat them as tooling options if the user already runs one - not as an independent methodology, and never as a prerequisite.

## Figure grades

| Claim                                    | Figure                                                                               | Grade                                                                       | Handling                                                          |
| ---------------------------------------- | ------------------------------------------------------------------------------------ | --------------------------------------------------------------------------- | ----------------------------------------------------------------- |
| ICP share of TAM                         | typically 5-15%                                                                      | Practitioner convergence across sources                                     | Use as an order of magnitude                                      |
| Best customers share structural traits   | ~80% share 3-5 traits                                                                | Commonly cited pattern rather than a published statistic                    | Use as a derivation heuristic, not a statistic                    |
| Rubric size                              | 8-15 criteria                                                                        | Practitioner convergence                                                    | Use as a design bound                                             |
| Weighting ladder thresholds              | equal <~50 closed-won; regression 100-150+                                           | 2026 multi-source synthesis + a16z/vendor methodology                       | Use as decision gates                                             |
| Buying-committee size                    | 6-10 (Gartner, 2017 survey) to ~13 (The State of Business Buying, 2024); others 8-13 | Well-sourced, but spread reflects differing methodologies and years         | Cite the range and attribute each end; never one number           |
| "68% higher win rate with a defined ICP" | attributed to "SiriusDecisions (now Forrester)"                                      | **Orphan citation** - widely attributed, never traced to a published report | Never state as fact; mention only with this caveat                |
| New Breed deal-size lift after narrowing | +83% average deal size                                                               | First-party self-report, specific and published by the company itself       | Strongest of the common case studies; label self-reported         |
| Lavu ARR growth after narrowing          | $10M → $40M+ ARR                                                                     | Narrowing story credible; ARR figure not independently audited              | Present as claimed outcome, never verified                        |
| Mid-market B2B SaaS win-rate baseline    | ~20-30% (roughly $10K-$50K ACV)                                                      | Independent 2026 benchmark compilations                                     | Use to sanity-check any claimed narrowing lift                    |
| Quarterly-vs-annual refresh lift         | 20-35% better MQL-to-closed-won conversion for quarterly refreshers                  | Single-source vendor figure                                                 | Use the direction (fresher beats staler), drop the precise number |
| MQL-to-SQL staleness tripwire            | below ~15% signals stale scoring                                                     | Practitioner convergence                                                    | Use as a tripwire threshold, not a target                         |

## Standing caveat

Much of the practitioner literature behind ICP guidance is vendor content (data providers, scoring tools, CRM platforms) with a direct incentive to make ICP scoring look more precise and more universally beneficial than the evidence supports. When the skill cites a figure from this file, carry the grade with it - a claim's handling instruction is part of the claim.

# Worked example - data-led ICP scoring rubric with retro-scoring pass

An industrial-B2B setting, chosen because it carries the best-documented counterintuitive weighting finding. In the sourced manufacturing case (the FORGED scorecard lineage), firmographics - industry, revenue, headcount - turned out to be the _least_ predictive category, while certifications held, machines on the floor, and supply-chain trigger events predicted far better.

The structure below (categories, criterion counts, 100-point distribution, retro-scoring protocol) is sourced practice. The specific weights and pass-rate numbers are **illustrative arithmetic**, shown so the shape of a passing validation is concrete, never reusable benchmarks.

## Setting

A vendor selling quality-management software to mid-size manufacturers. CRM holds 140 closed deals tagged won/lost, plus churn flags. That volume supports analyst-defined weighting and is approaching the 100-150+ bar for regression.

## Step 1 - segment by value

Rank cohorts by LTV, time-to-value, churn, expansion, and reference potential - then profile the winning cohort, not the whole base. Here the winning cohort was not the largest accounts: the biggest logo by revenue consumed outsized support and churned at month 14 - a warning sign, excluded from the profile deliberately.

## Step 2 - derive criteria from tagged accounts

Tagging the winning cohort's attributes surfaced five shared structural traits (consistent with the commonly cited ~80%-share-3-5-traits pattern), which seeded the rubric. Lost deals that _matched_ firmographics were mined separately for disqualifiers - the two most common loss reasons became deductions, not footnotes.

## The rubric - 11 criteria, 100 points, analyst-defined weights

| #   | Criterion                                                                        | Category       | Weight |
| --- | -------------------------------------------------------------------------------- | -------------- | ------ |
| 1   | ISO 9001 or equivalent certification held                                        | Technographic  | 14     |
| 2   | 10+ CNC/production machines on the floor                                         | Technographic  | 12     |
| 3   | Runs an ERP but no dedicated QMS                                                 | Technographic  | 12     |
| 4   | Supply-chain trigger in last 12 months (new OEM contract, audit failure, recall) | Intent/trigger | 14     |
| 5   | Hiring for quality-engineering roles                                             | Behavioral     | 8      |
| 6   | Multiple visitors from the account on pricing/spec pages                         | Behavioral     | 8      |
| 7   | Discrete manufacturing vertical (not process)                                    | Firmographic   | 8      |
| 8   | 100-800 employees                                                                | Firmographic   | 6      |
| 9   | $20M-$250M revenue band                                                          | Firmographic   | 6      |
| 10  | North America or EU plant locations                                              | Firmographic   | 6      |
| 11  | Named quality leader exists (dir.+ level)                                        | Behavioral     | 6      |

Note the tilt: firmographics carry 26 of 100 points despite being four of eleven criteria - the analyst panel weighted by believed predictive power, informed by the sourced finding that firmographics predicted least in this domain. A generic template would have inverted that.

**Disqualifier deductions** (scored as subtractions, first-class):

- Locked into a competitor QMS contract with 18+ months remaining: **-40**
- Regulated sub-vertical the product lacks compliance modules for: **-40**
- Fewer than 25 employees (support economics never work): **-30**

**Score bands** (the bands feed the tiering skill downstream - this file only defines them): 70+ strong fit · 45-69 moderate · <45 out of profile.

## Step 3 - the retro-scoring pass (the step most teams skip)

Protocol: score the last 50-100 closed deals - won _and_ lost - against the draft rubric before anyone uses it live. Here, 90 deals (55 won, 35 lost). Illustrative result:

| Draft-rubric band | Deals | Won | Win rate |
| ----------------- | ----- | --- | -------- |
| 70+               | 31    | 26  | 84%      |
| 45-69             | 38    | 22  | 58%      |
| <45               | 21    | 7   | 33%      |

**Reading it**: high scorers close at a materially better rate than low scorers, monotonically across bands - the rubric is predictive, ship it. Had the bands come out flat (say 62% / 58% / 55%), the correct move is to revise weights and re-run, not to ship and hope.

In the first draft of this rubric, criterion 8 (employee band) was weighted at 14. Retro-scoring showed no separation on it, and the points moved to the supply-chain trigger. That reweighting is the entire value of the pass.

**Report the pass inside the artifact** - the validation table travels with the ICP so the next refresh can re-run it against the new quarter's deals.

## Negative example - the rubric that fails

A four-criterion, firmographic-only checklist ("SaaS vertical, 50-500 employees, $10M+ revenue, US-based"), weights guessed in a meeting, built on 12 closed deals, no disqualifiers, never retro-scored.

Every documented failure mode at once:

- Too few criteria to differentiate.
- One category out of four.
- Data volume that supports only equal weighting, yet weights were invented anyway.
- Nothing to subtract for known-bad patterns.
- No evidence the score predicts anything.

It will pass firmographically perfect accounts that just renewed with a competitor, and its owner will not find out until two quarters of outbound have been spent on them.

---
name: sales-icp-definition
description: Defines the company's ideal customer profile (ICP) at sales-leadership altitude - the firmographic, technographic, behavioral and intent criteria, a weighted scoring rubric with explicit disqualifiers, and the refresh cadence that targeting keys off. Works from closed-won/lost deal data, or from founder-led discovery when deal data is thin. Covers B2B account ICPs and the equivalent B2C demographic/psychographic profile. Use whenever the user mentions the ideal or target customer, buyer profile, "who should we sell to", or targeting that is too broad, even without the letters ICP. Do NOT use for TAM/SAM/SOM sizing (mbfinotti/sales-skills@sales-market-sizing) or account fit scoring (mbfinotti/sales-skills@sales-account-segmentation).
license: MIT
metadata:
  author: Maya-Beth Finotti
  version: "1.3.10"
---

# Sales ICP Definition

You are an advisor to sales leadership defining the ideal customer profile - the single upstream artifact that account segmentation, tiering, territory design, and outbound targeting all consume instead of re-deriving. Produce criteria, a scoring rubric (or a founder-stage hypothesis), explicit disqualifiers, and a refresh plan - never the downstream machinery built on top of them.

Stay at the macro altitude:

- Segmenting accounts along the ICP's dimensions belongs to mbfinotti/sales-skills@sales-account-segmentation.
- Sorting accounts into coverage tiers belongs to mbfinotti/sales-skills@sales-account-tiering.
- Sizing the market the ICP filters belongs to mbfinotti/sales-skills@sales-market-sizing.

See References for these and other sibling skills.

## Invocation examples

Each ask enters at a different point. Run the interview first regardless.

- _"Define our ICP"_ - full build, whichever branch the data volume selects.
- _"Our targeting is too broad / sales is chasing everything"_ - diagnostic entry: run the risk check against the failure modes below, then rebuild from the branch step.
- _"Build an ICP scoring rubric"_ - data-led branch; confirm deal volume first, because a rubric built on 15 deals is a guess wearing a spreadsheet.
- _"We're entering a new market - update the ICP"_ - build a **separate** ICP for the new market; never broaden the existing one to cover it.

## Interview

Ask before proposing:

- Ask one question per message.
- Offer the multiple-choice options where given.
- Skip anything already answered by prior context.
- If you can read the company's website, product docs, or CRM exports, offer to auto-draft a v1 for the user to correct instead of interviewing from a blank page, then use the interview to fill only the gaps - most users prefer correcting a draft.

1. How many closed-won deals does the company have with usable data: (a) fewer than ~20, (b) ~20-50, (c) ~50-150, (d) 150+? This selects the branch - ask it first.
2. Is this B2B, B2C, or both? For B2C: is the purchase low-involvement (one person decides in minutes) or high-consideration (household deliberates - home, car, solar, education, financial products)?
3. Is this a first ICP, a refresh of an existing one, or entry into a new market/vertical/geography?
4. Who consumes the ICP - sales targeting, marketing, territory design, product - and who will be its named owner? An ICP defined by committee with no owner is one of the documented failure modes.
5. What deal data exists: CRM records tagged won/lost/churned, win rate by segment, churn and expansion by cohort? Missing data changes how much the rubric can claim.
6. What sales motion runs today (self-serve, sales-led, hybrid), and at what typical deal size? The motion shapes what an ICP-fit account even looks like - see mbfinotti/sales-skills@sales-motion.
7. By what date must the ICP be in use - a planning cycle, a territory carve, a campaign launch?
8. Do you want a one-off win or a compounding asset: (a) a good-enough filter for this quarter's list build, (b) a validated rubric the next three years of targeting rest on?
9. What is your effort ceiling: analyst hours available, whether anyone can run and read a regression, and the political capital to tell the field that some current accounts are out of profile?

Re-rank the weighting ladder against answers 7-9 before proposing anything. Say which answer moved what:

- A hard date inside weeks demotes regression regardless of deal volume (analyst-defined ships in a workshop).
- A compounding mandate (8b) promotes regression plus instrumented drift checks despite the effort.
- A low effort ceiling keeps the work at the current rung and spends the remaining hours on the retro-scoring pass, which is never optional.

## What an ICP is and is not

Draw this line before any scoring work; the documented failures are mostly conflations.

- **TAM, ICP and persona are three layers answering three different decisions:**
  - TAM sizes the opportunity (board reporting, fundraising).
  - The ICP filters which accounts sales and marketing pursue, typically 5-15% of TAM.
  - Personas shape what to say to the people inside those accounts.

  Build the ICP first, layer personas on top. "Mid-market B2B companies in North America" is a market segment, not an ICP - it is too large to disqualify anything.

- **Buying triggers are a fourth, orthogonal input - never fold them into the ICP.** A funding round or a new regulation decides _who to pursue first_ among already-qualified accounts, not which accounts qualify. Capture triggers in the rubric's intent category as scoring signals, but keep the qualify/prioritize distinction explicit in the artifact.
- **ICP scoring is not lead scoring:**
  - The ICP scores the _account_ (company) on structural fit before any engagement.
  - Lead scoring scores _individual contacts_ on engagement readiness.
  - The account-fit half lives here.
  - The contact-readiness half lives in mbfinotti/revops-skills@lead-scoring.

## Pick the branch by data volume - not preference

Two branches, selected by a constraint, so do not rank them:

- **Data-led rubric** - enough closed-won/lost history exists to derive criteria from what actually closed (roughly 50+ deals; below that the "data-led" label flatters a guess).
- **Founder-led discovery** - below ~50 deals: read the ICP off early customers bottom-up, ship a hypothesis plus an anti-ICP, and graduate to the rubric at the waypoints below.

Between ~20 and ~50 deals, run founder-led discovery _and_ an equal-weighted checklist - the interviews supply the criteria, the checklist makes them mechanically applicable, and neither pretends to statistical weighting.

## Brainstorm before committing

An ICP hardens fast - list builds, territory carves and campaign spend get keyed to it within weeks. Surface the candidate cuts first.

1. After the interview (and the segment-by-value or pull-signal analysis), present 2-3 candidate ICP hypotheses - e.g. a narrow vertical cut, a broader size-band cut, or a problem-anchored cut. Give each one trade-offs (addressable volume, win-rate concentration, evidence strength) and end with one explicit recommendation. Ask remaining clarifying questions one at a time, multiple-choice where possible.
2. Get explicit approval on a candidate before building the rubric or hypothesis document around it.
3. Build the artifact section by section, validating each with the user before the next:
   1. Definition paragraph.
   2. Criteria and disqualifiers.
   3. Weights.
   4. Validation result.
   5. Refresh plan.

   A wrong definition paragraph invalidates everything downstream, so never present the artifact as one finished block.

4. Gate finalization on user approval of the assembled artifact.

If your harness has persistent memory, store the approved artifact: definition paragraph, criteria, weights, disqualifiers, owner, and refresh dates. This lets later runs and the sibling segmentation/tiering skills start from the recorded decision instead of from scratch.

## Data-led branch: derive, weight, validate

1. **Segment by value before profiling anything.** Rank customer cohorts by LTV, time-to-value, churn, expansion and reference potential, then profile the _winning_ cohort's attributes - never firmographics-first across the whole base. Best customers by fit, not largest by revenue: a big logo that consumed enormous support and churned is a warning sign, not a criterion.
2. **Derive criteria empirically across four categories:**
   - Firmographic - who the company is.
   - Technographic - what it runs.
   - Strategic intent/triggers - what just changed (funding, leadership, expansion).
   - Behavioral - what it is doing (relevant hiring, pricing-page engagement).

   Tag actual accounts and look for shared structural traits. The commonly cited pattern is that ~80% of best customers share 3-5 specific traits. Target 8-15 criteria total - fewer than 8 usually cannot differentiate accounts.

3. **Make disqualifiers a first-class subtraction, not an afterthought.** Mine _lost_ deals that matched firmographics for the real reason they died (price, missing feature, regulatory blocker, location) and score those as deductions. Also ask what disqualifies an account even when it looks attractive on paper.
4. **Weight on the data-maturity ladder** - a ranked menu, ordered here by efficiency:
   - value: `regression-based > analyst-defined > equal weighting`
   - effort: `regression-based (100-150+ tagged deals, statistical skill, weeks) > analyst-defined (a cross-functional workshop) > equal weighting (near-zero)`
   - efficiency: `analyst-defined > equal weighting > regression-based`

   Rung by rung:
   - **Equal weighting** (below ~50 closed-won deals): the data forces this rung - there is not enough signal to differentiate, and pretending otherwise bakes guesses in as precision.
   - **Analyst-defined** (default rung): team judgment about which factors matter, distributed over a 100-point scale by believed predictive power.
   - **Regression-based** (promote once 100-150+ closed-won deals exist and someone can run and read the model): the efficiency order starves this rung, since it is highest value but loses every round on effort. Promote it anyway when the rubric routes real spend (territory carves, outbound budget), because at that point a wrong weight is a budget misallocation.

   This ordering is a default, not a law: an in-house analyst gets regression near-free, and a team already running revenue-intelligence tooling has paid most of its cost. Re-rank against the interview answers and say what moved.

5. **Triangulate the target variable.** Weight against a blend of outcomes - win rate, retention, expansion, time-to-value - never raw deal size alone, which rewards big-but-bad-fit logos (a16z's Michael King's guidance).
6. **Validate by retro-scoring** - the step most teams skip, and the one that separates a rubric from a slide:
   - Score the last 50-100 closed-won _and_ closed-lost deals against the draft.
   - If high scorers did not close at a materially better rate than low scorers, the rubric is not predictive yet: revise weights and re-run, never ship unvalidated.
   - Report the result inside the artifact.
7. **Segment the rubric.** A mid-market product and an enterprise product need different size bands, tech criteria and signals - one rubric per major segment, prioritized by revenue potential, not one rubric stretched across all of them.

Worked rubric with a full retro-scoring pass, and the negative example to avoid: [rubric-worked-example.md](./references/rubric-worked-example.md).

## Founder-led branch: narrow, read pull, write the anti-ICP

1. **Narrow, never broaden.** The most repeated founder mistake is a wide ICP kept out of fear of shrinking the market. The narrower the definition, the faster pattern recognition compounds across conversations - "selling to everyone equals closing no one."
2. **Read the ICP off pull, not push.** Watch for disproportionate inbound enthusiasm from a recognizable account shape and follow it, even when it contradicts the plan - the signal is who keeps pulling hardest toward the product, not a whiteboard guess.
3. **Interview early adopters, capture verbatim.** Ask why they specifically chose the product, what benefit they actually realized, and how they use it day to day. Keep their exact phrases - verbatim language is both the ICP evidence and the future messaging.
4. **Write the anti-ICP as a companion artifact**: who is disqualified even though they look attractive on paper. This is the narrowing discipline made into a document - it is what stops the profile re-broadening the first time a good-looking "maybe" deal appears, which is genuinely hard to resist pre-revenue. Name that tension to the user explicitly.
5. **Set the graduation waypoints in the artifact:**
   - ~10-20 deals: a qualitative hypothesis should exist.
   - ~50 deals: hands off to an equal-weighted rubric.
   - 100-150+ deals: regression becomes available.

   These are waypoints on one bottom-up-to-top-down path, not competing thresholds.

Full discovery sequence, interview script, and the design-partner overfitting trap: [founder-led-discovery-example.md](./references/founder-led-discovery-example.md).

## The artifact

Express the ICP as **both** a one-paragraph plain-language statement **and** a pass/fail criteria checklist - never one without the other. The paragraph carries the nuance a checklist flattens. The checklist is what gets applied mechanically to an account. Add:

- **Disqualifiers / anti-persona**, paired with the top objections heard in sales - disqualification and objection-handling are one section, not afterthoughts.
- **Switching forces** for accounts that fit but do not move: Push (frustrations with the current solution), Pull (what attracts them), Habit (what keeps them stuck), Anxiety (what worries them about switching) - Bob Moesta's Jobs-to-be-Done switching framework. Fit explains who could buy. The forces explain why a fit account still does not.
- **Confidence grades** on any claim about fit that is not independently verifiable:
  - High: two independent sources or an official page.
  - Medium: one credible source plus consistent circumstantial evidence.
  - Low: flagged as uncertain, never asserted.
- **A version number and changelog** - each refresh records what changed and why, so drift is auditable.
- **Primary and secondary segments.** The ICP drives focus. It does not exclude all others. Name the acceptable secondary profile so the field knows the difference between "out of focus" and "disqualified".

```
CONTEXT: branch used · deal-data volume · consuming teams · named owner
DEFINITION: one-paragraph ICP statement · pass/fail criteria checklist
RUBRIC (data-led): criteria by category with weights · disqualifier deductions · score bands
HYPOTHESIS (founder-led): narrowed statement · pull evidence · anti-ICP · graduation waypoints
VALIDATION: retro-scoring result, or the dated plan to run it at ~50 deals
LANGUAGE: verbatim customer phrases · switching forces (push/pull/habit/anxiety)
REFRESH: quarterly light review · annual rebuild · drift tripwires · off-cycle triggers · version/changelog
RISKS: failure modes checked · confidence grades on unverified claims
```

## Refresh plan

An ICP untouched for six months is not just outdated - it points the team at a market that has already moved. Stack three cadences rather than picking one:

- **Quarterly light review**: a cross-functional session (sales, marketing, RevOps) re-scoring the quarter's closed-won/lost/churned deals against the current rubric.
- **Annual full rebuild** of the criteria and weights themselves.
- **Monthly drift tripwires** between reviews: scoring distributions, reply rates, routing accuracy.

Off-cycle triggers that override the calendar:

- A pricing/packaging change.
- A launch that changes who gets value.
- Entering a new market, which gets a **separate** ICP, not a broadened one.
- Repeated closed-won deals landing outside the profile.
- An unexplained win-rate or cycle-length shift among ICP-matched accounts.
- MQL-to-SQL conversion falling below roughly 15%, or reps routing around MQLs entirely - both read as stale scoring.

Technographic data decays faster than firmographic data, so a technographics-heavy rubric needs more frequent data re-verification regardless of cadence.

## B2B vs B2C

The unit of decision changes, so the layering model itself changes shape - not just the criteria inside it.

**B2B**: the unit is the account plus its buying committee - surveys put the committee at 6-10 decision-makers (Gartner, 2017) up to ~13 stakeholders (The State of Business Buying, 2024). Cite the range, not one figure. Criteria stay firmographic/technographic at the company level, owned by RevOps, marketing leadership or the CRO.

**B2C**: there is no firmographic layer to build a company-level ICP from. The structural equivalent is a **demographic/psychographic profile of a person or household** (age, income, life stage, location, values), playing exactly the role the ICP plays in B2B.

- High-consideration purchases: the household distributes decision roles (initiator, influencer, decider, buyer, user, gatekeeper) the way a committee does, and roles shift per purchase.
- Low-involvement purchases: collapse the roles into one person, and the committee-mapping machinery drops out.

**Shared, explicitly:**

- Both profiles are built from actual closed-deal/purchase data rather than assumptions.
- Both layer behavioral/psychographic signals on top of the base.
- Both need disqualifiers, validation and a refresh cadence.
- The narrowing discipline transfers unchanged.

## Failure modes

Run the finished artifact against each of these before it ships - as concrete questions, not abstract advice:

- **The FOMO trap** - too broad. Check: does the ICP disqualify roughly 70% of inbound? If not, it is not doing its job.
- **ICP conflated with TAM** - a target list too large to be actionable. Check: is the ICP a small fraction of TAM (typically 5-15%)?
- **ICP conflated with a market segment** - "companies in industry X" is a slice, not where the company wins, retains and expands.
- **Firmographic-only scoring** - a firmographically perfect account can still fail: mid-reorg, just renewed with a competitor, no champion. People buy from people, not businesses. Keep the intent and behavioral categories, and hand persona work downstream.
- **Built from a conference-room brainstorm** - check: was every criterion derived from tagged won/lost/churned accounts (or, founder-stage, from real customer interviews)?
- **Largest customers mistaken for best customers** - revenue size and fit are different axes.
- **Descriptors instead of the problem** - "Series B, 200-person SaaS" says who, not why they buy. Anchor on the problem and value delivered. Firmographics are the filter for who has that problem.
- **Static artifact, no owner** - check: named owner, refresh dates, version number.
- **Expansion confused with refinement** - before Series B, sharpening almost always beats broadening. Widening the ICP because growth targets were missed is usually the wrong direction at that stage.

## Measurement

The artifact is not done until all of these pass. Iterate until 100%:

- The definition includes both the paragraph and the pass/fail checklist, with disqualifiers as first-class entries.
- The retro-scoring result is reported (or, founder-stage, the dated plan to run it at ~50 deals) - and if high scorers did not outperform low scorers, the weights were revised before shipping.
- A named owner, the three-cadence refresh plan, and the off-cycle trigger list appear in the artifact.
- Every cited benchmark carries its source and year. The widely repeated "68% higher win rate with a defined ICP" figure is an orphan citation that was never traced to a published report - it never appears as fact (see the figure-grading reference below).
- The B2B or B2C scope is stated explicitly, and a both-sided company gets a profile per side, not a blend.

Outcome KPIs to track after the ICP ships:

- Win rate on ICP-fit vs. non-fit accounts (the rubric's live validation).
- Share of new pipeline inside the ICP.
- MQL-to-SQL conversion against the ~15% tripwire.
- The drift signals from the refresh plan.

For calibration when judging claimed narrowing lifts: independent compilations put mid-market B2B SaaS win rates around 20-30% - treat any case study promising far more with the skepticism the figure-grading reference sets out.

## References

- See mbfinotti/sales-skills@sales-account-segmentation for segmenting accounts along the dimensions this ICP defines - segmentation consumes the ICP, it never re-derives it.
- See mbfinotti/sales-skills@sales-account-tiering for the tier cutoffs and coverage levels built on ICP-fit scores.
- See mbfinotti/sales-skills@sales-market-sizing for TAM/SAM/SOM estimation - the sizing layer above the ICP; this skill filters, that one counts.
- See mbfinotti/sales-skills@sales-motion for the motion context that shapes the ICP - a self-serve motion and an enterprise motion produce different profiles from the same market.
- See mbfinotti/sales-skills@sales-quota-setting for the territory and quota math that keys off ICP-based account pools.
- See mbfinotti/revops-skills@lead-scoring for the contact-readiness half of scoring - this skill owns account fit, that one owns lead behavior.
- See [./references/rubric-worked-example.md](./references/rubric-worked-example.md) for the worked scoring rubric with its retro-scoring pass and the negative example.
- See [./references/founder-led-discovery-example.md](./references/founder-led-discovery-example.md) for the founder-led discovery sequence, interview script and anti-ICP example.
- See [./references/named-frameworks-and-figure-grading.md](./references/named-frameworks-and-figure-grading.md) for the named frameworks (Dunford, SPICED, a16z Five-Question) and how much weight every benchmark this skill cites deserves.