How to find lookalike companies from your best customers (with a 20-match check)

How to find lookalike companies: the short version

To find lookalike companies, pick about five customers of one type who stayed, give them to a lookalike tool as seeds, compare companies on what they do instead of their industry code, and check the first 20 matches by hand before you build a list.

1. Choose seeds from closed-won customers that are still active after a full renewal.
2. Keep the seed set small and clean. Start with five of one customer type.
3. Compare what companies do: what they sell, to whom, how they charge, their size and their tech.
4. Add negative seeds only when your losses follow a pattern.
5. Run the search by hand, in a sales database or in a dedicated lookalike tool.
6. Check the first 20 matches with yes/no tests you wrote in advance.
7. Research and prioritize the matches before anyone is contacted.

The output of a lookalike run is a list of candidates, not leads. A match says a company resembles your customers. It does not say the company has budget, a problem you solve or a reason to talk this quarter.

Lookalike prospecting or a B2B lookalike audience?

Lookalike prospecting starts from examples instead of filters. You give a tool a few companies that are good customers, and it returns other companies that resemble them. It suits the case where you can point at good customers but cannot say in one sentence what they have in common.

A "B2B lookalike audience" is the advertising version: an ad platform builds an audience that resembles a list you upload, and you reach it with ads. This guide covers prospecting, where you end up with a list of companies to research yourself.

Step 1: Choose seeds to find companies similar to your best customers

Closed-won tells you who you could sell to. Retention tells you who was worth selling to. A list of everyone who ever signed includes customers who left in month three, bought on a heavy discount or needed something you do not build, and a tool will faithfully find more of them.

Apollo's September 2026 guide gives the same advice: pull closed-won customers with strong retention or expansion, not every account that signed. Its academy play adds deals that closed smoothly and accounts that saw quick value. Apollo sells a lookalike feature, so read this as a vendor's view.

Here is a simple way to sort your own customers. The company and numbers are fictional, for illustration only. Ledgerline sells invoice reconciliation software to bookkeeping and accounting firms on annual contracts.

The rule: a seed is a customer still active after at least one full renewal, which is 12 months or more on annual contracts. A negative seed is a customer who left early for a reason tied to what kind of company it is. Everyone else stays out until you know more. Adjust the bar to your own contract cycle. If your tool has no way to weight seeds, this sorting is the weighting.

Two habits help:

• Run each customer type separately. Mixed seeds give mixed results.
• Make each seed a clean example of one business. Ocean.io's help center uses IKEA, which operates across retail, logistics and furniture manufacturing, to show that a filter can steer a search toward the part of a business you care about. Our takeaway: a seed that spans several businesses blurs the signal.

Step 2: How many seed customers do you need?

Vendors disagree, and every number comes from a company that sells a lookalike tool or an ad product.

Even one vendor's two guides, about three weeks apart, give different ranges. We found no test of seed count on any of these pages. Only Lusha gives a reason: a model that reads a pattern across a set needs several examples.

Our reading, a hypothesis and not a finding: a tool that turns one website into a text profile measures closeness to that site, so one example can do. A tool that learns a pattern across a set needs several examples to separate signal from noise. LinkedIn's minimum applies to advertising audiences, so it is not a guide for a prospecting list.

A practical start:

• Use five clean seeds of one customer type. Five is the one number that sits inside Apollo's, Ocean.io's and Lusha's guidance above.
• Try eight to ten as a second run if you have more strong customers of that type and your tool takes them, then compare the two runs.
• Split by type. If you have 30 strong customers across three types, run three groups, not one pooled list.
• Test your own setup. Run the search twice with different subsets of your seeds. If the two top-20 lists barely overlap, seed choice is driving the results, so tighten it.

Step 3: Match on what companies do, not on industry codes

Compare companies on what they sell and to whom, how they charge, their size and their tech. Use industry codes only to exclude sectors you will not sell to, such as government or education.

Why industry codes alone fail

The Census Bureau describes the North American Industry Classification System (NAICS) as production-oriented: it groups establishments by similarity in the processes used to produce goods or services, in order to collect and publish statistics. Its stated purpose is statistics, not finding buyers.

• One code can hold unlike buyers. NAICS 513210, Software Publishers, covers software published through subscriptions or downloads, and its index entries include mobile apps, games and operating systems. A company that sells bookkeeping software by subscription and a studio that publishes mobile games can both fit, though their buyers, sales cycles and budgets have little in common.
• Similar buyers can sit in different codes. In the Census definitions, CPA offices, bookkeeping services and payroll services are 541211, 541219 and 541214. A tool for accounting teams might suit all three.

Researchers have compared text with codes. Gerard Hoberg and Gordon Phillips built competitor sets from the product descriptions in 50,673 annual 10-K statements. In an NBER working paper, later published in the Journal of Political Economy, they report that these sets explain the rivals managers name, and traits such as profitability, better than existing classifications. They studied public companies, so applying it to small private firms is our inference: what a company says it sells is a better similarity signal than the box it was filed in.

Vendors say much the same. Apollo's and Instantly's guides both describe good lookalike search as going beyond industry codes and headcount, and Ocean.io says it builds a meaning-based profile from a website: what a company offers, who it serves and how it positions itself. A community Apify actor takes the other route and matches on the seed's industry, country and a headcount band (half to twice the seed's), and says so plainly. That is the label approach described above, with the benefit that its criteria are explicit.

What to compare instead

Instantly's guide lists much the same signals. The weights depend on your product: for a vertical product, what a company sells and to whom dominates, while for one that depends on an integration, tech stack matters more. Before you search, write down the two or three signals that were true of your best wins. The ideal customer profile guide covers how to find them.

Step 4: Use negative seeds to tell the search what to avoid

A negative seed is a company you do not want more of. The idea is standard in similarity search: Qdrant's recommendation API, for example, accepts lists of positive and negative examples.

A good negative seed left, or was a poor fit, for a reason tied to what kind of company it is: too small, in a regulated sector, or in need of a feature you do not build. A bad one lost on timing, lost to a champion who left, or lost for a reason you never recorded. Those do not describe a kind of company, so they only add noise.

Keep it to a handful of clear cases. Put hard rules in filters, not seeds: current customers, competitors, anything you will not sell to. Check that your tool excludes your customer list (Lusha, for one, says it does), and whether a negative seed removes resembling companies or only ranks them lower.

Step 5: Free and paid ways to run a lookalike search

Details are as read on 2026-10-17. Prices and limits change.

Each tool decides what "similar" means from its own data, and none can tell you why a customer bought. That is why the check in Step 6 matters more than the tool.

To try the idea before you pick a tool, our free lookalike finder takes up to 3 customer websites and returns up to 10 similar companies, each with a note on what it shares with your customers and how it differs. It needs no sign-up and allows 3 searches a day.

Step 6: Validate the first 20 lookalike matches by hand

Before you build a list from a lookalike run, review a sample of it.

1. Take the top 20 exactly as returned. Do not choose the 20 that look best.
2. Before opening the first one, write three or four yes/no tests drawn from your best wins. For example: same kind of product for the same kind of buyer, same way of charging, inside your size band, a team that would use your product. Add one disqualifier: a competitor or an existing customer.
3. Spend three to five minutes on each, reading the homepage and pricing page. That is one to two hours in total. Mark a company good only if it passes every test. Unsure counts as not good.
4. Count the good ones.

Read the count with care, because twenty is a small sample. A 95% Wilson score interval, which suits small samples, puts the plausible range for the true share of good fits at about 48% to 85% when 14 of 20 are good. For 10 of 20 the range is about 30% to 70%, and for 16 of 20 it is about 58% to 92%.

So 20 can catch a list that is clearly bad or clearly good: with 5 or fewer good the range stays under 50%, and with 15 or more it stays above 50%. It cannot tell 70% from 80%. That is enough to decide what to do next. These rules of thumb are ours, not statistical cutoffs:

• 14 or more good: go on to a bigger batch, and keep reviewing a small sample of each new batch.
• 8 to 13 good: read the misses. What do they share? That trait is your next filter or negative seed. Change the seeds and re-run, not only the filters.
• Fewer than 8 good: stop. The seeds, or the tool, are not capturing what makes your customers alike.

This check measures resemblance, not reply rates, so do not read a good score as a forecast.

Step 7: Turn lookalike matches into a researched list

A reviewed batch is still a list of company names. To make it usable:

• Clean it. Remove duplicates, current customers, competitors and past rejects. Keep each company's match reason beside it.
• Research each account before contacting anyone: what it does, what changed recently, who owns the problem. The guide to researching a prospect keeps claims tied to sources.
• Order by timing. Similarity decides who is on the list. Signals such as hiring or funding decide who goes first. The guide to B2B buying signals covers how to use them.
• Find the person and verify the address before you write anything.
• Start small. Twenty-five to fifty reviewed accounts teach you more than 500 names you cannot personalize.

Finding a company is not permission to email its staff. The rules depend on the country and on who the recipient is, so check them before you send. The cold email and GDPR guide is a starting point. This is general information, not legal advice.

How FoxKnock handles lookalike prospecting

FoxKnock sells a lookalike mode, so read this section as a vendor's description.

FoxKnock's Lookalikes mode takes 1 to 10 seed domains you paste in (its form says 3 to 5 of your best customers work best), plus optional negative seeds for customers you lost. It reads the seeds' websites and ranks up to 150 candidates from a similar-company search, a web search and your CRM. Candidates that resemble a negative seed rank lower, and strong matches carry a "resembles a lost customer" flag.

The top 12 matches get a short note on what each company shares with your seeds and how it differs, so you can run the 20-match check with reasons in front of you. FoxKnock sends the text it gathered to Claude, by Anthropic, to write those notes (if AI is unavailable, they fall back to keyword overlap). Searching is free. Accepting a result costs 15 credits at default prices, which can change, and companies already in your CRM are free to accept.

It has limits. It judges similarity from public web text, has no company database of its own beyond your CRM, and cannot know why a customer bought. Without a meaning-based embedding model set up, it falls back to a built-in word-based one, and the run details say so. See lookalike companies for details.

If you want to try this on your own market, the Free plan includes 500 one-time credits and needs no card (they are added once you verify your email). That covers about 33 accepted companies at default prices. Sign up free, paste five of your best customers and accept only the matches worth keeping.