A Claude prompt that takes a seed list of your best customers and returns verified lookalike companies with firmographics. The Lusha connector runs the lookalike model on the company graph. Claude ranks and filters the output by the criteria you actually care about.
Once Lusha is connected in Claude, the connector runs in the background — no special syntax needed. Just paste the prompt and run.
<context>
Expand prospecting from a list of best customers using lookalike modeling.
</context>
<task>
1. Seed list (5-100 domains or LinkedIn URLs):
[PASTE DOMAINS OR LINKEDIN URLS]
2. Find lookalikes similar to the seeds.
3. For each, return:
Company | Domain | Industry | Headcount | Country, city.
4. Filter:
drop under [HEADCOUNT MIN],
drop outside [TARGET GEOGRAPHY] (if specified),
drop seeds.
5. Return top 25 after filtering.
</task>
<constraints>
- Seeds must be domains or LinkedIn URLs (no free-text names).
- Lookalike modeling is broad — filter on headcount and geography to make it usable.
- If fewer than 25 after filtering, flag which filter is too tight.
</constraints>Input: Seed list — 7 mid-market B2B SaaS customers. Filter — headcount 100+, US-based.
Output: 15 lookalikes returned in the first batch, more available via pagination. Two strong matches in the target headcount range surfaced immediately, NYC-based B2B SaaS companies. The rest of the batch was filtered out for being below the headcount floor.
| Company | Headcount | Industry | City |
|---|---|---|---|
| [NYC B2B SaaS, mid-market] | 492 | Software Development | New York, NY |
| [NYC B2B SaaS, mid-market] | 209 | Software Development | New York, NY |
This is a real result from running the prompt against the live Lusha connector. Lookalike modeling returns a broad set of candidates — the filter pass in step 4 of the prompt is what turns the output into a usable list.
ICP discovery usually starts from a hunch. Running it from a seed list of real closed-won customers turns it into a data exercise — pairing the Lusha lookalike model with Claude’s filter pass lets you go from “I think we should target fintechs” to “here are 25 companies that look like our best customers, in our target size and region.”
Two things matter for results to land:
Strong seeds. Five to fifteen of your most-similar closed-won customers produces sharper output than a mixed bag across stages and sizes. The model matches on multiple signals, so seed homogeneity is what tightens the result.
A filter pass. Lookalike modeling is broad by design — it surfaces candidates across the signal space. Claude’s filter step is what narrows the output to your target headcount, region, and industry.
The prompt pulls candidates from 300M+ verified companies under GDPR, CCPA, SOC 2, ISO 27701, ISO 31700, and TRUSTe.
Five minimum, one hundred maximum. Seven to fifteen highly-similar customers produces the sharpest match. Below five, the model has too little signal. Above fifteen, similarity dilutes.
The Lusha lookalike model accepts domains or LinkedIn URLs as seeds. Company names need to be resolved to domains first — Claude can do that step automatically if you paste the names, but accuracy is higher when you provide domains directly.
Lookalike modeling matches across many signals — industry, location, growth pattern, and others. If your seeds vary in size or include any small companies, the candidate pool reflects that. The filter pass in step 4 is built to handle this — set a headcount floor and the noise drops out.
Yes. The Lusha lookalike tool supports an exclusion list. Pass your current customer domains in the prompt and Claude will route them to the exclude parameter.
Company search filters on attributes you specify upfront — industry, headcount, technology. Lookalike modeling finds companies similar to a seed across attributes you didn’t think to specify. Use lookalike when you trust your closed-won data more than your hypothesis.
One data connection. Works in Claude, ChatGPT, your CRM, or any agent you build.
