How lookalike audiences work
You give the platform a source: customers you have uploaded, people who bought through the pixel, video viewers, list members. It looks at what those people have in common across the signals it holds, then assembles a much larger audience of people who resemble them. You choose how closely they must resemble the source — a tighter setting gives a smaller, more similar group, a looser one gives a bigger, vaguer group.
The quality of the result depends almost entirely on the source. A source built from your best customers describes something worth copying. A source built from every website visitor, or from a list too small to hold a real pattern, describes noise, and the audience produced from it is a bigger pile of the same noise.
Why lookalike audiences matter — and why that is now contested
They mattered because for years they were the main way to find new people who behaved like existing buyers, at a time when the alternative was guessing at interests.
That position is now genuinely disputed. A number of practitioners and commentators argue that as delivery systems improved at finding buyers on their own, a lookalike mostly acts as a fence: it removes people the system would have found and does not improve the ones it keeps. Others still see value, especially for niche products, small markets or businesses whose buyers are unusual enough that broad delivery struggles to find them.
I would treat this as a judgement call rather than a settled fact. It is not something the platforms have confirmed, and the evidence on both sides is largely account experience rather than published proof. Anyone telling you flatly that lookalikes are dead, or that they are essential, is overstating what is actually known.
Common mistakes with lookalike audiences
The first is building them from a weak source. A short list, a mixed list, or a list of leads that never became customers produces an audience that resembles the wrong people.
The second is stacking many bands as separate ad sets. Those bands overlap heavily, so they compete with each other in the auction and each one gets too little spend to learn from.
The third is treating a lookalike as a quality guarantee. It describes similarity, not intent — the people in it are not looking for you today, and the creative still has to earn the click.
The fourth is deciding the question by opinion. This is precisely the sort of claim that should be settled inside your own account, not by whoever spoke loudest at a conference.
How to act on it
Test rather than adopt a position. Run a lookalike against the same offer and creative on broad targeting, give both enough budget and enough time to leave the learning phase, and compare cost per qualified result rather than click-through rate. Change one thing at a time so the comparison means something.
Where you do use them, invest in the source first: your best customers rather than all customers, refreshed as the custom audience behind it grows. Keep the number of bands small, exclude existing customers so you are actually prospecting, and re-test the decision periodically, because the platforms keep changing underneath it. If you want a second opinion on what your own account data says, that is the sort of question a Meta ads consulting session is for.