When filters run out: targeting by resemblance instead of by criteria

A filter is a theory about your market. Industry, headcount, country, technology: you write down what you believe your customers have in common, and the database returns the companies that match your belief. If the belief is wrong, the list is wrong, and nothing downstream can tell you. Lookalike targeting inverts the order: you hand over the accounts that actually converted, and the model returns companies that resemble them on signals you never thought to filter on.

Why the filter is usually wrong

Ask a founder to describe their ideal customer and you get a segment: mid-market software companies in France, fifty to two hundred people. Look at who actually signed and you find three of them are not software at all, half are under fifty people, and the thing they share is that they had just hired a first sales person. No filter menu has a field for that.

This is not a failure of discipline. It is what happens when the thing that predicts a purchase is not one of the twenty-eight fields the database indexes. Resemblance models catch some of it because they work on the whole shape of a company rather than on the fields you picked.

What it needs to work

A seed list that actually closed. Not a list of people who booked a meeting, not a list of companies you like: accounts that paid. With four or five of them, the model has almost nothing to learn from and will hand you back your own assumptions in a more expensive format. With thirty or forty, it starts to be interesting, and the answer it gives you will usually be uncomfortable, because it will contain companies you had ruled out.

It also needs you to have exhausted the obvious segment, or to be about to. Resemblance is how you widen a market when the list you trust is running dry, and that is a specific moment. Before it, the honest answer is that a hand-written filter is cheaper and good enough.

Where it does not apply

Some markets are defined by a register rather than by resemblance. Every pharmacy in a region, every architect on a professional roll, every company with a specific licence: the list exists, it is exhaustive, and a model can only lose accuracy against it. On those markets you do not want a lookalike engine, you want the register and a way to enrich it.

Local businesses are the other exclusion, for a different reason. A model built on the digital footprint of companies has very little to read on a business whose entire presence is a shop front and a phone number. That rail is served by maps and directories, at a fraction of the cost.

The part that annoys buyers

Ocean.io publishes no price. Their pricing link goes to a product signup, and the plans page returns a 404. Checked on 18 September 2026. It means a sales conversation before you know whether the budget is plausible, which is a real cost in a category where three competitors will tell you their per-credit rate in a click. It does not make the approach wrong. It does mean you should arrive at that call with your seed list already prepared and a number in your head, so the demo is about your data rather than about theirs.

How to test it in an afternoon

Take the twenty accounts you are proudest of. Split them: fifteen as the seed, five held back. Run the model on the fifteen and see whether the five you held back come out in the results. If they do not, the signals it reads do not describe your market, and no amount of credits will fix that. If they do, the companies you had never heard of in the same output are worth a campaign, and that is the entire value of the category in one experiment.