The Four Layers of a Useful B2B ICP

Who this guide is for

This guide is for B2B teams that have a target market but still cannot explain which individual companies belong in it. It is designed for industrial products, components, equipment, materials, and business services where company structure, facility conditions, buying route, and product use matter more than a broad demographic label.

Contents

1 Why one-dimensional ICPs fail 2 Layer one: account structure 3 Layer two: application fit 4 Layer three: commercial relationship 5 Layer four: evidence and actionability 6 Connect the layers with hard gates 7 Use the four-layer worksheet 8 Illustrative example 9 Common mistakes 10 What this does not mean 11 Common questions 12 Sources and methodology

1. Why one-dimensional ICPs fail

“Manufacturers in the United States with 50 to 500 employees” is a searchable population, not a complete ICP. It does not identify which entity or site matters, whether the product fits an operating application, how the account would buy, or what evidence is sufficient to act. The problem is not that industry, geography, and size are useless. They are efficient starting filters. The U.S. Census Bureau describes NAICS as the standard used by federal statistical agencies to classify business establishments for collecting and publishing economic statistics. U.S. Census Bureau - NAICS The scope is statistical classification, not sales qualification. A useful ICP must make a decision easier: include, exclude, investigate, or defer a named account record. SellersGPT recommends four layers because each resolves a different source of false fit: 1 Account structure: Are we evaluating the right organizational and operating unit? 2 Application fit: Is there a credible use for the exact offer? 3 Commercial relationship: Through what role and buying path would the offer move? 4 Evidence and actionability: What is known, how reliable is it, and what action is permitted? These layers are an editorial framework, not an official standard.

2. Layer one: account structure

The first layer defines the unit of analysis. A parent company, legal entity, establishment, plant, warehouse, and project location can produce different answers. The Bureau of Labor Statistics defines an establishment as a single economic unit, usually at one physical location and primarily engaged in one economic activity; a firm or company may consist of multiple establishments. BLS - QCEW Concepts That distinction matters because the relevant process may exist at one plant but not another. Record at least:

• parent and legal-entity identity;

• operating facility or site;

• business unit or division where relevant;

• primary and secondary activities;

• geography of use, purchase, and service;

• ownership, channel, or procurement constraints;

• scale measures appropriate to the offer. Public datasets can help bound a universe, but their unit must remain visible. County Business Patterns provides establishment, employment, and payroll statistics by geography and industry. U.S. Census Bureau - County Business Patterns API An establishment count is not a count of independent buying companies, qualified facilities, or reachable accounts. Layer-one output: a named record with an explicit unit, not a floating company name.

3. Layer two: application fit

Application fit explains where, how, and why the offer could create value. It should be specific enough to verify. Compare these statements:

• Weak: “Uses automation.”

• Better: “Operates automated packaging lines.”

• Testable: “Operates washdown packaging lines where the sensor must detect reflective packages before sealing within the stated environmental and response specifications.” Capture five application fields: Field Question Process What operation or workflow contains the use? Function What must the offer do? Conditions What environmental, technical, regulatory, or integration constraints apply? Installed alternative What product, process, workaround, or non-consumption currently performs the function? Consequence What operational outcome changes if the function improves or fails? Application fit can be confirmed, probable, unknown, or contradicted. A product page or industry code may create a plausible hypothesis; direct facility documentation, equipment lists, specifications, permits, tenders, or first-party operational descriptions can strengthen it. Application fit is a hard gate. A large account with no relevant use should not remain qualified because it scores well on revenue or employee count. Layer-two output: an evidence-linked application statement and its constraints.

4. Layer three: commercial relationship

The third layer defines how the account participates in the commercial system. SellersGPT uses four common relationship hypotheses:

• OEM: integrates the offer into a product or system it sells.

• Distributor: resells the offer and may provide coverage, inventory, credit, or service.

• End user: operates or consumes the offer in its own process.

• Contractor or integrator: specifies, combines, installs, commissions, or delivers it in a client project. This layer answers questions that firmographics cannot:

• Who controls or influences the specification?

• Who contracts and pays?

• Who receives the operational value?

• Is demand tied to production, stock, a facility need, or a project?

• Who owns installation, warranty, field service, and lifecycle support? One company may require several relationship records. Do not let a parent-level “manufacturer” label overwrite the fact that its distribution subsidiary resells products while a factory division consumes them. Layer-three output: a role tied to one account unit and application, plus its buying and support path.

5. Layer four: evidence and actionability

The final layer prevents a plausible profile from becoming false certainty. Every critical field receives an evidence state: State Meaning Permitted action Confirmed Direct, current, scope-matched evidence supports the field Use in qualification and prioritization Probable Several facts align, but a material dependency remains Run a bounded verification step Unknown Evidence is absent, stale, or cannot resolve the claim Preserve as unknown; do not score as false Contradicted Reliable evidence conflicts with the claim Exclude or revise the hypothesis For public companies, the SEC provides public access to filings through EDGAR. U.S. Securities and Exchange Commission - Search Filings A filing may support claims about reported segments, material facilities, products, or risks, but its scope and reporting period must be retained. For private companies, first-party websites, facility pages, certification records, procurement documents, product literature, and direct verification may fill different fields. Actionability adds practical conditions:

• Is the account reachable through a lawful, appropriate business path?

• Can the offer be delivered and supported in the relevant geography?

• Does a channel, incumbent, approved-vendor, or contractual restriction apply?

• Is the evidence current enough for the decision?

• What single next action would reduce the most important unknown? Layer-four output: an evidence ledger and an allowed next action, not an intent claim.

6. Connect the layers with hard gates

Do not average the four layers into a score before checking hard gates. Use this sequence: Candidate universe → correct account unit → application pass → relationship pass → service/access pass → prioritized ICP record Each rate retains its denominator:

• Structure-pass rate = records with the correct account unit / candidate records reviewed

• Application-pass rate = records passing application / structure-passed records

• Relationship-pass rate = records with a viable relationship / application-passed records

• Actionable rate = records passing service and access / relationship-passed records Multiplying these rates is valid only because they describe sequential conditional filters over the same documented flow. The product equals final actionable records divided by initial candidates, apart from rounding. Unknowns do not pass or fail automatically. Maintain a separate queue and define the evidence needed to resolve them. Only after gates pass should a score order the viable accounts.

7. Use the four-layer worksheet

Layer Required fields Hard-gate question Evidence state Account structure Parent, entity, site, activity, geography, relevant scale Is this the unit where the application and buying path exist? Confirmed / probable / unknown / contradicted Application Process, function, conditions, alternative, consequence Can the exact offer plausibly perform the required function? Confirmed / probable / unknown / contradicted Relationship OEM, distributor, end user, contractor/integrator; buyer and beneficiary Is there a commercially viable path for this role? Confirmed / probable / unknown / contradicted Evidence/action Source, date, scope, serviceability, access, next check Is there enough current evidence to prioritize? Confirmed / probable / unknown / contradicted Add three governance fields: hypothesis version, decision owner, and review date. An ICP without those fields tends to accumulate exceptions until it no longer means anything.

8. Illustrative example

Illustrative example - explicitly fictional. Alder Process Systems, a fictional supplier of hygienic pump components, starts with 120 employer-establishment records in two states. The team applies the layers sequentially: Stage Records remaining Stage pass rate Candidate records 120 - Correct site and account unit 74 74 / 120 = 61.7% Relevant hygienic fluid application 32 32 / 74 = 43.2% Viable OEM, end-user, or integrator relationship 18 18 / 32 = 56.3% Serviceable and sufficiently evidenced 11 11 / 18 = 61.1% The end-to-end actionable share is 11 / 120 = 9.2%. Using unrounded fractions, the conditional product is (74/120) × (32/74) × (18/32) × (11/18) = 11/120 = 9.2%. The calculation describes only the fictional reviewed universe. It is not a market-size estimate, response forecast, or expected number of qualified leads. Twenty-one application records remain unknown because facility-level process evidence is unavailable; they stay in a research queue rather than being scored as zero. The exercise changes the ICP from “food manufacturers with 50-500 employees” to a testable record: “a serviceable production site operating a specified hygienic fluid process, with a confirmed component application and an evidenced OEM, end-user, or integrator buying path.”

9. Common mistakes

• Treating industry, country, and employee band as a finished ICP.

• Evaluating a parent when the application exists only at one facility.

• Using product relevance as proof of an operating application.

• Mixing OEM, distributor, end-user, and contractor economics in one segment.

• Converting missing evidence into a failed criterion or an average score.

• Counting establishments as independent firms or buyers.

• Treating a click, reply, or page visit as proof that all four layers pass.

• Writing so many exceptions that the ICP can no longer exclude anything.

10. What this does not mean

The four-layer framework does not require perfect information before any action. It requires uncertainty to be visible and the next action to be proportional to the evidence. A probable record can enter a bounded research or validation task without being represented as confirmed. The framework also does not map every person in the purchase. Contact fit and buying-group roles come after the account, application, and relationship are defined. It does not predict response, meetings, opportunities, revenue, or customer success. Finally, the four layers are not a universal legal or statistical definition of an ideal customer profile; they are a SellersGPT decision tool.

11. Common questions

Which layer should come first? Start with account structure because the other evidence must attach to the correct unit. In practice, research may loop between layers when an application reveals the relevant site or division. Can application fit be inferred from NAICS? NAICS can support a hypothesis about predominant activity. It does not prove that the exact process, equipment, condition, or buying path exists at a named site. Should all four layers receive equal weights? No. Use hard gates before weights. If an application is impossible or the account cannot be served, a weighted average should not restore eligibility. How many unknowns are acceptable? There is no universal count. Define critical dependencies. One unknown certification requirement may matter more than five missing descriptive fields. Is account engagement the fourth layer? No. Engagement is an observed activity signal and must retain its definition. It can guide research or timing, but it does not replace account, application, or relationship fit.

12. Sources and methodology

The four-layer model, evidence states, gates, formulas, and worksheet are SellersGPT editorial tools. Source claims preserve their official unit and scope. Public statistics are used to bound and classify populations, not to declare named accounts qualified or interested. 1 ICP02-BLS-QCEW - U.S. Bureau of Labor Statistics, “Quarterly Census of Employment and Wages: Concepts.” https://www.bls.gov/opub/hom/cew/concepts.htm Accessed August 24, 2026. 2 ICP02-CENSUS-NAICS - U.S. Census Bureau, “North American Industry Classification System.” https://www.census.gov/naics/ Accessed August 24, 2026. 3 ICP02-CENSUS-CBP - U.S. Census Bureau, “County Business Patterns APIs.” https://www.census.gov/data/developers/data-sets/cbp-zbp/cbp-api.html Accessed August 24, 2026. 4 ICP02-SEC-EDGAR - U.S. Securities and Exchange Commission, “Search Filings.” https://www.sec.gov/search-filings Accessed August 24, 2026.