Sourcing Lookalike Accounts

Outcome: a sourced list built from the pattern your best customers share, with the pattern stated explicitly rather than inferred by a model.

Surface: App and MCP server
Level: Intermediate
Uses: enrich_organization · find_company_techstack · find_companies
Credits: ~25 for a 20-customer seed and a 200-company source
Prerequisite: Lesson 06

Extract the pattern, do not ask for it

“Find companies like these” hands the pattern-finding to a model that cannot see your revenue data. Do it yourself — it is three steps and the result is auditable.

Enrich the seed

Take 20–30 of your best customers and run full company enrichment plus tech stack on them. This is the only expensive part, and it is a one-off.

Tabulate, do not eyeball

For each attribute, count how many of the seed share it. Headcount band, country, industry, founded-year band, funding stage, each detected tool.

Keep what concentrates

An attribute shared by 70%+ of the seed is a pattern. An attribute shared by 30% is noise. Write down the ones that concentrate.

The comparison that matters

A pattern in your best customers is only useful if it is not equally present in everyone else.

Run the same enrichment over 20 bad-fit or churned customers. Then compare:

Attribute	Best customers	Bad-fit customers	Useful?
50–500 employees	85%	80%	No — everyone looks like this
Uses a modern data warehouse	75%	20%	Yes — strongly discriminating
Funded in last 24 months	60%	55%	No
Has a dedicated ops role	70%	25%	Yes
Based in Western Europe	90%	88%	No — that is your market, not your ICP

Only the discriminating attributes belong in the lookalike definition. The others describe your market and will produce a large, undifferentiated list.

This comparison is the single most valuable half-hour in the course. Most teams have never run it, and most discover their headline ICP criteria do not discriminate at all — while one unglamorous attribute does almost all the work.

Source against the pattern

Now the sourcing prompt writes itself, using only the discriminating attributes plus the market filters from lesson 06:


Using find_companies, source companies matching:
  - employee count 50-500
  - country in: GB, IE, NL, DE, FR
  - industry: software

Cap at 300. Return name, domain, employee count, country.
Do not enrich yet.

Then for the sourced companies, run find_company_techstack and keep only
those with a modern data warehouse detected.

Report how many survived.

Sourcing on the market filters and then filtering on the discriminating one is deliberate: the stack check is cheap, and running it as a post-filter is far more reliable than hoping a directory filter captures it.

The survivorship trap

Your best customers are the ones who bought and stayed. That is a filtered set, and it carries three biases worth naming:

Channel bias. If most of your seed came from one channel, you are describing that channel’s audience, not your ICP.
Vintage bias. Customers from three years ago reflect an older product. Weight recent wins higher.
Effort bias. Accounts you won because a founder personally worked them for six months are not repeatable at volume.

The correction is to include recent wins, look at accounts across channels, and check whether any pattern is really an artifact of how you sold rather than who they are.

Do this now

Pick 20–30 best customers

By retention and expansion, not just by logo size.

Pick 20 bad-fit customers

Churned, or never expanded, or expensive to serve.

Enrich both sets identically

Company enrichment plus tech stack.

Tabulate both

Percentage per attribute, both sets side by side.

Keep only the discriminating attributes

Those with a large gap between the two sets.

Source with market filters, post-filter on the discriminator

Cap the run.

Sample the output

Read 20 of the sourced companies. Do they feel like your customers? If not, a discriminating attribute is missing.

Check your work

Both seed sets were enriched with the same columns
You have a side-by-side percentage table
Only discriminating attributes are in the definition
The sourced sample reads like your customer base
You have named at least one survivorship bias in your seed

Where this breaks

Building a lookalike from a seed of five customers produces a pattern that is mostly coincidence. Five companies will share several attributes by chance, and the resulting list will look precise while being arbitrary. Twenty is the practical floor. If you do not have twenty good customers yet, use the lesson 06 filter approach and revisit this when you do.

Further automation

Re-run the tabulation quarterly with your newest customers added. The discriminating attribute changes as the product and market change, and catching that early is the difference between a list that keeps working and one that slowly stops. TAM Sourcing builds the full refreshing version.

Next lesson

08 — Research agents that stay on task, where the Research module begins.

Reference for this lesson: find_companies, find_company_techstack, Company techstack, TAM Sourcing.