ICP research with Claude: finding the 500 accounts that actually matter

ICP docs are usually useless

Most ICP documents we inherit from clients are a page of adjectives. "Mid-market SaaS companies in fintech, growing fast, post-Series-A, value-driven leadership." This is not actionable. You can't message "value-driven leadership."

A useful ICP is a list. Five hundred specific accounts. With specific people. With specific reasons each one is worth reaching out to. Claude makes that list approachable.

The three-step loop

Step 1 — Define in concrete terms

Instead of "mid-market fintech," we extract:

• Revenue band (e.g. $10M–$100M ARR)
• Employee count (e.g. 50–500)
• Geography (e.g. US, UK, Singapore)
• Stack signals (e.g. uses Segment, hiring Data Engineers)
• Funding stage (e.g. Series B in last 18 months)
• Pain signal (e.g. recently posted about data quality issues)

Claude helps refine this by reading your existing closed-won deals and finding patterns you didn't articulate.

Here are our 12 best-fit customers and 8 bad-fit ones. What do the good ones share that the bad ones don't?

The output is surprising every time. Usually two criteria nobody on your team had noticed.

Step 2 — Build the list

We pull company data from a mix of sources: Crunchbase, LinkedIn Sales Navigator, Apollo, Clay, Common Crawl, and public job boards. Claude then filters the raw list against the concrete ICP criteria.

Here are 5,000 companies. For each, decide whether it matches the ICP criteria below. Rank the matches 1–5 on how strongly they fit. Output: company, website, employee count, why it matches, why it might not.

This turns 5,000 rows of noise into 500 ranked targets. In one prompt.

Step 3 — Find the people

For each matching company, Claude finds the right role:

For this company, list every plausible decision-maker for [your offering]. Include role, LinkedIn URL if known, tenure, and one signal that they're a good target now (e.g. recent post, recent hire, recent funding).

The output: your list of 500 specific humans, with a reason to reach out.

Why this beats buying a list

Bought lists are dead by the time you buy them. They're filtered for job title, not context. Half the contacts have already churned. The other half have been emailed by ten other agencies this quarter.

The list you build with Claude is yours. It's filtered on context, not just title. It includes a "why now" signal for every contact. Reply rates are 3–5x higher on average.

The unlock: refresh monthly

Here's the bit that most teams miss. Your ICP list isn't a one-time artifact. We refresh ours monthly:

• Remove contacts who changed roles
• Remove contacts who've replied (they're in CRM now)
• Add new contacts whose signals have fired (just got funding, just hired a Head of Data)

Claude runs this monthly refresh in 20 minutes. Without it, your list decays by ~15% a month.

What to feed Claude for this

The quality of the list depends entirely on the quality of the inputs. What we feed:

• Your own closed-won data (for pattern extraction)
• Company data from a reputable source
• LinkedIn profile data (scraped or via Sales Navigator export)
• Signals feeds (Crunchbase for funding, BuiltWith for stack, job boards for hires)

The fewer sources, the shallower the list. Budget for good data.

Where this breaks

Two places.

1. Niche industries with little public data. If your ICP is "family offices managing $50M+" — Claude can't see those. You need human research.

2. Deeply technical fit signals. Claude can tell you a company uses Segment. It can't tell you whether they're using it well. For highly technical ICP work, you still need SMEs.

For 90% of B2B ICPs, though, this loop is ready today.

---

We run this for every client at kickoff. It's part of Revenue.