Most results for this term are about advertising audiences on consumer platforms. The same idea works for building B2B prospect lists, and there it solves a specific problem: filters only capture what you can name, while your best customers share traits you would never think to write down. Below: where the method comes from, how to choose between seeding with customers or with a description, how many seeds to use, and where the output still needs a human check.
What is lookalike modeling?
A method that ranks every company in a database by how similar it is to a seed set. Each company is described by many attributes at once: industry, size, location, technologies, hiring, how it describes itself. The model places all companies in that space of attributes and returns the ones closest to the seeds.
The difference from filtering is the logic. A filter is a yes-or-no rule: software, 50 to 500 employees, Germany. A company either passes or it does not. A lookalike model scores degrees of closeness across all attributes, so a company that misses one filter but matches everything else still ranks high, and a company that passes every filter but resembles none of your customers ranks low.
The term comes from advertising: Facebook's 2013 Lookalike Audiences let advertisers choose between the closest 1% of users in a country for similarity or the closest 5% for reach. The academic basis for large-scale look-alike extension is set out by Ma et al., PMLR 2016 and, for a professional network, by Liu et al., KDD 2016. No public benchmark exists for the accuracy of B2B lookalike lists, so test the output against your own customer data.
Where does the idea come from?
From online advertising. In 2013 Facebook launched Lookalike Audiences: an advertiser uploaded a list of existing customers and the platform found users who resembled them. Advertisers chose between a similarity setting, the closest 1% of users in a country, and a reach setting, the closest 5%.
Research teams then described how to do this at scale. Ma and colleagues at Yahoo (2016) presented a look-alike audience extension system for very large user bases. Liu and colleagues (KDD 2016) described audience expansion from a seed set on a professional network, the closest academic precedent for B2B use. Applied to companies rather than people, the same idea builds prospect lists.
Seed with customers or with a description?
With customers, if you have any. Existing customers encode traits of your ideal profile that you would struggle to put into words: a mix of size, stack, growth stage and business model that no single filter captures. A description captures only what you already know to write.
- Seed with customers: when you have at least one company that bought, stayed and fits the profile you want more of. Use your best customers, not all of them.
- Seed with a description: when you are new to a market, have no customers yet, or are entering a segment unlike your current base. Write what the company does and how it works, not just its industry label.
- Combine them: run both and compare. Companies that appear in both lists are the strongest candidates.
A worked example, with illustrative numbers. Five seed customers, all B2B software companies with an in-house sales team, return a ranked lookalike list of 300 companies. Enrichment adds company size, technologies and open job positions. An AI column then classifies each company against your criteria, and about 60 remain as the list a rep actually works.
How many seeds do you need?
Fewer than most people expect, as long as they are consistent. Leadspicker accepts a sample of one, five or ten companies. One seed returns the neighbours of one company, which is useful when you want more of a single excellent customer. Five to ten seeds let the model find what they have in common and ignore what is unique to each.
Consistency matters more than count. Ten seeds from three different segments pull the model in three directions, and the list lands between them, close to none. If your customers fall into distinct groups, run one lookalike per group. Remove customers who churned or never fit; the model will copy their traits as faithfully as the good ones.
What does a lookalike list still get wrong?
It matches the past, not the present need. A lookalike list is a strong starting point and a weak final answer.
- Similar is not ready. A company can resemble your customers in every attribute and have no reason to buy this quarter. Timing comes from signals, such as hiring or a management change, not from similarity.
- It copies your bias. If your seeds are all one size or one country, so is the output, even when the product would sell elsewhere.
- Surface matches slip through. Two companies can describe themselves in similar words and run different business models.
- It finds companies, not people. The list says which accounts to approach, not who inside them to contact.
The fix is the same in each case: narrow the list with enrichment and a relevance check before any message goes out.
Everything in the Leadspicker database is precomputed into lookalikes. A list can be seeded with one, five or ten existing companies, or with a plain-text description of the ideal customer, without writing a single filter.
Frequently asked questions
How many seed customers do you need for a lookalike list?
One is enough to start, and five to ten consistent customers usually work better because the model can find what they share. Seeds from different segments should go into separate lookalike runs.
Is a lookalike list better than a filter?
It captures traits a filter cannot express, because it scores closeness across all attributes at once. Use filters for hard limits such as region, and lookalikes to rank what is left.
What does a lookalike list still get wrong?
It finds companies that resemble past customers, not companies ready to buy now. It copies any bias in the seeds and does not say whom to contact. Narrow it with enrichment and signals.
