Lookalike audiences: how to pick a seed that actually works

A lookalike audience is a model built from a list you supply, used to find people who resemble it. The modelling is the platform's job and it is generally good at it. Your job is choosing what "resemble" should mean, and that is entirely decided by the seed.

The default behaviour — upload everybody who ever bought — produces a model of your average customer. Your average customer is dominated by one-time buyers of your cheapest product, because there are more of them than anybody else.

Seeds worth trying, in order

1. Repeat buyers. The single best general-purpose seed. Two or more completed orders, refunds excluded. Models the behaviour you actually want more of.
2. Top decile by 12-month revenue. Sharper, smaller, and best where your margin varies a lot by customer.
3. Buyers of one product line. For a specific campaign. Narrow, and the most likely to beat broad targeting on a specific push.
4. Subscribers who became customers. Useful when the goal is list growth rather than immediate sales.

What is rarely worth it: all customers, all subscribers, or anybody who visited the site. Those describe your traffic, not your good outcomes.

Size: bigger is not better

• A few thousand good records beats fifty thousand mixed ones. Signal quality dominates once you are over the minimum.
• Watch the match rate, not the upload count. An audience of 20,000 that matches at 30 % is a seed of 6,000, and the platform is modelling on that.
• If a good seed is below the minimum, widen the definition one step — repeat buyers over 24 months instead of 12 — rather than abandoning quality entirely.

Cleaning the seed before you upload it

Everything that is not a good customer in the seed makes the model slightly worse, and these are cheap to remove:

1. Refunded and returned orders excluded from the value and frequency calculation. A customer who bought six times and returned five is not a champion.
2. Parcels returned unpaid excluded entirely. Modelling on these actively finds more people who will do the same.
3. Wholesale, staff and test accounts removed. They distort value heavily and there are always more of them than anyone expects.
4. Only consented records. A seed is a customer match upload and needs the same lawful basis.

Percentage, expansion and where to spend

Most platforms let you choose how broad the resulting audience is — commonly expressed as a percentage of a country's population.

• Start at 1 %. It is the closest match and the easiest to judge.
• Widen only when 1 % exhausts, measured by frequency rising and cost per acquisition drifting up.
• Do not compare a 1 % and a 10 % audience on ROAS alone — they are doing different jobs, and the wider one should be judged on incremental reach.
• Exclude existing customers from the resulting audience. A lookalike will happily include the seed's own members if you let it, and paying to acquire people you already have is the most common quiet waste in this format.

Refresh, and judge it against a real alternative

A seed uploaded once ages: it stops containing your newest good customers and keeps people who have since lapsed. Refresh monthly, as a replacement rather than an addition.

Then test it honestly. The comparison is not lookalike against nothing — it is lookalike against the platform's own broad targeting with the same creative and budget. Broad targeting with good product data has closed much of the gap in recent years, and shops running lookalikes out of habit sometimes find the seed is doing very little.