Est. reading time: 7 minutes
Somewhere in the last few years, “lookalikes are dead” became a safe thing to say in the marketing world. Then we open an account, look at what the lookalikes were seeded with, and find the entire customer list dumped in as one source, untouched since 2023. The audience didn’t stop working. It was never given anything worth modeling.
A lookalike is a pattern-matching system, and it can only find patterns that exist in the seed. Feed it a seed full of one-time discount buyers mixed with your best repeat customers, and it will faithfully find you more of that blend. The output degrades exactly as much as the input did. That’s the diagnosis behind most “lookalikes don’t work anymore” accounts, and it’s fixable without touching a single ad.
How the model actually uses your seed
When you build a lookalike audience, Meta takes your source audience, identifies the shared traits and behaviors across those people, and ranks the rest of the eligible population by similarity. The percentage you choose sets how deep into that ranking the audience reaches. A 1% lookalike takes the closest matches, and expanding toward 5% or 10% trades precision for scale.
Two things follow from that mechanism. First, the seed is the only meaningful input you control. The similarity model is Meta’s, the population is Meta’s, and your entire influence on the output is the quality of the list you hand over. Second, the percentage is a dial, not a setting to get right once. The correct value depends on how much you’re spending and how quickly you exhaust the tight end of the ranking.
There’s also a floor worth knowing. Meta requires at least 100 people from a single country in the source, and its own guidance points to sources in the range of 1,000 to 5,000 people as the sweet spot. Below that, the model is inferring a pattern from noise.
The default mistake: seeding with everyone
The most common seed we find is the full customer export. Every buyer, every order value, every acquisition source, uploaded as one list. It feels rigorous because it’s complete. It’s actually the weakest version of the input, because averaging your best customers with your worst produces a profile that describes neither.
The fix is to seed with the customers you want more of, not the customers you happen to have. In practice that means building the source from your top tier by lifetime value, your repeat purchasers, and your most engaged subscribers, and leaving out the one-time bargain hunters entirely. If you haven’t worked out where that value tier sits, Shopify’s guide to calculating customer lifetime value covers the math, and an ESP like Klaviyo can segment the list by purchase frequency and order value before export.
Meta also supports value-based lookalikes, where you include a value column in the upload and the model weights high-value customers more heavily instead of treating every row as equal. If your order values vary widely, this is the difference between asking for more customers and asking for more revenue.
The tradeoff to respect
Narrowing the seed pushes against the size floor. A top-5%-of-LTV segment from a young brand might be 300 people, which is technically enough and practically thin. When the high-value segment is small, widen the definition (top 20% instead of top 5%, or repeat buyers regardless of value) rather than padding the list with weak customers. A slightly looser definition of “good” beats a precise definition diluted with noise.
The plateau: one seed, running forever
The second pattern shows up in accounts that got real results from a lookalike and then never touched it again. Performance erodes over months, and the erosion gets blamed on creative or on the platform. What actually happened is that the campaign worked through the closest matches, and the seed was never varied to open new territory.
The fix is to treat seeds as a portfolio rather than a monument. Different sources model different traits, and each one reaches a different slice of the population:
- Seeds split by product category find buyers who resemble that category’s customer, which matters when your catalog serves distinct use cases.
- Seeds split by acquisition source (checkout buyers versus newsletter signups, for example) model different levels of intent.
- Seeds split by geography help when your customer profile genuinely differs by region, and add nothing when it doesn’t.
Run these as separate lookalikes with separate budgets so you can see which seed produces buyers, not just clicks. Judge them on cost per acquisition and ROAS over a real purchase cycle. CTR will tell you which audience is curious, and curiosity is not the metric you’re paying for.
When you don’t have enough buyers to seed with
New brands and low-frequency categories hit a genuine constraint here. If you have 80 customers, you don’t have a purchase-based seed, and no amount of segmentation will manufacture one. The instinct is to wait until the customer list grows. The better move is to seed from engagement instead.
High-intent behavior exists upstream of purchase, and Meta can build lookalikes from it. Visitors who cleared a meaningful time threshold on site, viewers who reached 75% of a product video, and people who engaged with your page or left reviews all carry signal about who your buyer will be. These seeds are weaker than a purchase seed, because engagement is a proxy and proxies are noisy. But a decent proxy at sufficient volume beats a perfect signal at insufficient volume, and you can graduate to purchase-based seeds as the customer list matures.
Where lookalikes sit in the current Meta account
One more piece of diagnosis, because it changes how much weight this tool should carry. Meta has spent the last several years pushing delivery toward broader, system-driven targeting. Audience expansion means the platform will often reach beyond your defined lookalike when it predicts conversions elsewhere, and Advantage+ setups lean on your conversion data more than your audience definitions.
That doesn’t make seeding irrelevant. It changes what the seed is for. Your lookalike is less of a fence and more of a starting instruction, and your conversion signal does the rest of the steering, which is why tracking quality sits underneath all of this. It also means targeting can no longer compensate for what happens after the click. The most precisely modeled audience in the account still bounces off a landing page that sends visitors back to Google, and a lookalike that gets served tired creative fatigues like any other audience. We covered the creative side of that problem in the shift that makes retargeting ads feel fresh again, and the same logic applies to prospecting.
The rebuild, in order
If your lookalikes have gone quiet, work through this sequence before concluding the tool is spent:
- Audit the current seeds. Check what each lookalike was built from, how old the source is, and whether it mixes customer tiers that should be separated.
- Build one value-weighted seed. Top LTV tier or repeat buyers, exported fresh, with a value column if order values vary.
- Stand up two or three seed variants. Split by product category or acquisition source, and let them compete on CPA at equal budgets.
- Test percentage as a scaling lever. Start tight, and widen only when the tight audience shows frequency creep or rising costs.
- Refresh on a schedule. Seeds built from live segments in your ESP stay current. Static CSV uploads decay quietly, so date them and rebuild.
Lookalikes reward the brands willing to be deliberate about inputs, and punish the ones who set them up once and move on. The accounts where they “stopped working” almost always skipped straight to the verdict without ever auditing the seed. Run the audit first. It’s a few hours of work against a conclusion that’s currently costing you scale, and if you’d rather have us look at the account alongside you, that’s a conversation we’re glad to have.
