A precise partner profile is what turns a vertical of hundreds of thousands of candidate domains into a clean recruitment list. This guide shows how to write an ICP as required, preferred, and disqualifying signals — and how to calibrate it on a reviewed sample until the error rate is low.
"We want quality travel blogs" cannot be screened — it means something different to every reviewer. A machine-screenable ICP breaks the idea into discrete signals, each one a yes/no question an LLM can answer from what the site actually publishes, each shipped with a confidence score.
Every one must hold. These define the floor — a domain that misses any required signal is out, no matter how good it looks otherwise.
Nice to have. They raise the fit score and let you rank the strongest partners first, without excluding a solid site that lacks one.
Any match removes the domain structurally. This is where you kill whole classes of false positive — competitors, own-commerce sites, thin spun content.
Start from how your affiliate team already vets partners by hand, then make each unspoken rule explicit. Five moves get you from intuition to a screenable definition.
Look at the publishers already converting for you. Name the concrete things they share — format, independence, freshness, monetization — not vibes.
Write down the sites you decline on sight: competitors, coupon-only shells, AI-spun filler. Each becomes a disqualifying signal.
Assign every trait to required, preferred, or disqualifying. Keep required tight — over-stuffing the floor shrinks the pool without improving quality.
Rewrite fuzzy wording until a reviewer could answer yes or no from the live site. "Reputable" becomes "hands-on testing with original screenshots."
Decide how many preferred signals lift a domain to "strong fit." This is the dial you tune during calibration, not a fixed rule.
The definition is agreed in writing before any full run. Everyone sees the same signals, so the output is auditable against a fixed spec.
The same three-tier structure adapts to any vertical. Notice how each disqualifier removes a specific, predictable class of false positive rather than a generic "low quality."
A draft ICP is a hypothesis. You confirm it by screening a slice, reviewing the results by hand, and tuning the signal wording until the measured error rate is low — then, and only then, running the full universe.
Illustrative figures. On a large-scale production run on our own classification platform, a reviewed sample of the full run measured ~96% precision against the client's own judgment. Slice size and passes vary by ICP.
A first draft rarely lands clean. On a representative slice you might see the LLM mark a handful of sites as fits that a human would reject — usually because a signal was worded too loosely and caught word-matches instead of intent. Each disagreement points at exactly which definition to sharpen.
Screen a ~1,000-domain slice Review each: fit or not fit Trace errors to a signal Reword, re-screen, remeasure"The disagreements are the valuable part — they tell us precisely which signal definition to tighten."
The calibration principle, applied to every engagementWhen we share worked examples or sample profiles, the underlying domains are withheld and descriptions are generalized, so the specific publishers stay intentionally not traceable — including via web search. This protects the sites without changing the data or the method. Your own pilot and paid deliverables carry the real domains, screened against your real ICP.
The free pilot calibrates your ICP on a reviewed sample, then delivers your first 20 qualified publishers plus a full-run projection. No cost, no obligation.
Request a Free Pilot