A lookalike audience is a copy of your best customers, drawn by an ad platform's model from a seed list you provide. The model is only as good as that seed. Feed it 5,000 high-value buyers and it finds more people like them. Feed it a thin, self-selected slice and it finds more of the slice, not more of your actual market.
Consent quietly wrecks seed quality in two ways. It shrinks the seed, and it skews it. Both matter, and most marketers only notice the first.
Google already made the call for you
Google Similar Audiences don't exist anymore. Google stopped generating new Similar Audiences segments in May 2023 and removed them from all campaigns and ad groups in August 2023, citing the phase-out of third-party cookies and rising privacy expectations. In their place, Google pushed advertisers toward automated alternatives: optimized targeting, audience expansion, and Smart Bidding.
The shift is worth understanding, because it tells you where the whole industry is heading. Optimized targeting finds people who resemble those converting on your ads right now, however you define a conversion, even if they look nothing like your original audience profile. You hand Google a seed and an objective, and its model does the expansion inside a privacy-safe boundary. You've lost the ability to inspect and hand-pick a similar segment. You've gained a model that leans on real-time conversion signal, which is exactly the signal consent degrades.
Meta lookalikes still exist, and still need a seed
Meta kept Lookalike Audiences. The mechanics haven't changed much: you give Meta a source audience, pick a country and a size from 1% to 10% of that country's population, and Meta builds the lookalike. The technical minimum for a source is about 100 people from a single country, but that floor produces weak results. Quality climbs with seed size, and practitioners generally want a source in the low thousands, with 5,000 or more producing the most stable lookalikes.
Here's the consent problem. If your source is a pixel-built audience of converters, it only contains people who accepted marketing cookies. So your 100 or 1,000 or 3,000-person seed isn't a random sample of your customers. It's the subset comfortable enough with tracking to click accept. Meta then models a lookalike off that subset. You end up targeting people who resemble your privacy-tolerant buyers, which may not be the same as people who resemble your most valuable buyers.
Small seed, biased seed
Break the damage into its two parts:
- Smaller seed, weaker model. Below a few thousand records, lookalike quality falls off. Consent can pull a healthy converter list under that threshold before you notice, especially for EU and UK traffic where reject rates run higher.
- Biased seed, wrong lookalike. Even a large consented seed carries a selection bias. It over-represents whatever demographic and behavioral traits correlate with accepting cookies. The lookalike inherits that bias. This one is invisible in the dashboard. Your audience looks fine. It just isn't aimed where you think.
The bias compounds when you tighten the lookalike. A 1% Meta lookalike is the closest match to your seed, so if the seed skews toward privacy-tolerant users, the tightest lookalike skews hardest in the same direction. Visitors sending a Global Privacy Control signal or browsing in regions where you default to reject are absent from the seed entirely, and those absences aren't random. They correlate with exactly the privacy-conscious segments you might most want to reach.
Seed from data you own
The durable fix is to stop seeding from pixel-built audiences and start seeding from first-party data you collected with proper consent.
- Upload a customer list. Hashed emails and phone numbers from your CRM, collected with marketing consent, make a stronger and less biased seed than a browsing-session pixel audience. They cover buyers who rejected cookies but consented to marketing at signup, so the selection bias is different and usually smaller.
- Use value-based seeds. Meta's value-based Lookalikes and Google's value-based bidding weight the seed by customer value, so the model chases revenue rather than raw conversion count. Garbage in still means garbage out, so the seed's completeness matters more than ever.
- Keep seeds fresh. A stale seed models last year's customer. Refresh on a schedule.
- Work with the automation, not against it. Google's optimized targeting and Meta's Advantage+ audience are built for a low-signal, consent-constrained world. A clean first-party seed plus broad automated expansion beats a hand-built segment starved of data.
Know your usable seed size
Before you build a lookalike, you should know what fraction of your converters actually consented, because that fraction is your real seed. CookieBeam's consent analytics report the accept and reject split by region and segment, so you can see how much of your buyer base you're allowed to seed from, rather than discovering it when the lookalike underperforms. The A/B testing tool helps lift that consented share by finding banner variants that convert more visitors to accept without manipulation. A bigger, less biased seed is the single biggest input to any lookalike, and it starts with consent you can actually see and grow.
For the companion problem of how consent shrinks the retargeting pools that feed these seeds, see how cookie consent shrinks your retargeting audiences. To go deeper on building durable owned data, read our first-party data strategy guide. And if you're weighing Consent Mode's advanced setup to keep modeling alive, compare advanced versus basic Consent Mode.