Why Lookalike Audiences Stop Working, and What to Seed Instead

For a year or two, the lookalike audience was the growth engine. Seed it with your customers, let the platform find more people like them, scale the budget, watch the customers arrive. Then, somewhere along the way, the engine went quiet. The dashboards still look fine. The attributed ROAS is still respectable. But blended cost per new customer has climbed for two or three quarters straight, and nobody can point to what changed.

Often, nothing obvious changed. That is the problem. Three quiet mechanisms can degrade lookalike performance while the visible settings stay the same, and they do not always show up in platform reporting.

Seed drift: the audience models the wrong customers. A lookalike is a similarity engine; it finds people who resemble the seed you gave it. Many lookalike seeds are built from all purchasers, or all converters, which means the seed is shaped by whoever converts most often, not necessarily whoever is worth most. As covered in our guide to LTV segmentation, the frequent discount buyer and the quiet full-price customer look identical in a conversion count and very different on margin. Seed from the count and the system can replicate volume-led economics: more of your cheapest customers, delivered efficiently. The lookalike may be working as designed. It was pointed at the wrong target.

Convergence: the new audience stops being new. As a lookalike runs alongside your other campaigns, its delivery can overlap with people you already reach: existing customers, site visitors sitting in retargeting pools, brand searchers. Attributed performance can hold up, because these people convert well; some were converting anyway. Incremental reach quietly shrinks. This is the same reporting blind spot we cover in platform ROAS vs incremental ROAS: attribution records contact, not cause, so an audience that harvests existing demand can look strong while contributing less new revenue than the dashboard implies. The economics of chasing an audience you already own get harder to defend, for the reasons argued in the Growth pillar.

Dilution: widening the net weakens it. When volume drops, the reflex is to widen the similarity percentage: 1% becomes 3%, becomes 10%. Each step can trade away the similarity that made the audience work in the first place. At the widest settings, a lookalike can behave like broad targeting with a lookalike's name, and you inherit broad targeting's economics without having chosen them.

All three mechanisms can share a signature: platform metrics hold or improve while blended numbers deteriorate. Attributed conversions increasingly come from people who were close to converting regardless, so the platform's report card can stay healthy. The number to inspect is blended cost per genuinely new customer: total spend against new customers acquired, from your own records, with existing customers excluded. If that line has been climbing while the dashboard stayed flat, you may be watching harvest replace prospecting.

An invented illustration of the shape: spend flat at £15,000 a month, attributed CPA steady at £45, while new-to-business customers fall from 280 to 190 a quarter. The dashboard says nothing changed. The P&L says the machine may be increasingly re-buying the audience you already had.

The fix starts with what you feed the system, and it is the same discipline the rest of this cluster is built on:

Seed from margin-inclusive high-LTV customers, not all purchasers. The LTV segmentation method produces the list: rank by margin-inclusive lifetime value, take the top tier, and make that the seed. A smaller, value-led seed pointed at the right target can serve you better than a large seed pointed at the wrong one.

Feed value-weighted conversions. Audience seeds tell the system who to look for; conversion values tell it what to optimise towards on every auction. Import values that reflect margin, or at least tier differences, so the delivery system has a signal for worth rather than volume. Without this, even a well-seeded audience can drift back towards the easiest converters.

Exclude the audience you already own. Suppress existing customers and active retargeting pools from prospecting campaigns, so the lookalike has a clearer job: finding people you have not reached. This exclusion can reveal the audience's true prospecting performance, which is the number you actually needed.

Re-seeding, value imports, and exclusion architecture are standard workstreams in our customer acquisition growth engagements, because they are where audience expansion usually goes wrong first.

Even a well-fed lookalike is one expansion route, not a growth strategy. When it plateaus, the durable moves are the ones in the Growth pillar: profile your best customers on pre-purchase attributes, map adjacent segments with challengeable hypotheses, and test them in stages with unit-economics pass marks. Intent terms, trigger occasions, and segment-specific propositions expand reach through evidence rather than through a wider similarity dial.

And before re-scaling anything, clean the account. Expansion layered onto existing waste can scale the waste alongside it; the pre-scale waste audit is the fifteen-minutes-per-check place to start.

Audience products are steadily absorbing lookalike logic into broader automated delivery, where the audience decision can happen inside a black box you cannot inspect. That shift makes the inputs more important, not less: the seed lists and conversion values you feed the system are increasingly among the few audience controls that remain. Getting them right is not busywork; it is one of the remaining steering inputs.