Lookalikes Were Supposed to Die on Meta. Why Are They Still Beating Broad in Some Accounts?

Lookalikes are dead for scaled DTC but alive in three account types: small-seed niche brands, high-ACV B2B, and strong-identity DTC categories. The shared pattern is signal density — the seed contains behavioural signal Andromeda cannot infer from broad. Classify the account against the three exceptions before accepting the broad-wins default.

Lookalikes stopped working for scaled DTC because Meta's current ranking system (Andromeda) already incorporates the behavioural, demographic, and engagement signals a lookalike audience was built to represent. Feeding a lookalike into a broad campaign that already has access to the same signal just restricts the delivery system from exploring segments outside the lookalike boundary. For accounts above roughly $5k monthly spend with addressable audiences over 100,000 people, broad consistently beats lookalikes on CPA and incremental ROAS. This is one of the structural reasons broad targeting dominates for scaled DTC accounts in 2026, and the generalised rule emerged from this account class specifically.

Signal density is the amount of behavioural information per customer in the seed audience that is genuinely rare in the broad population. A scaled DTC seed of 50,000 customers who bought generic consumer goods contains little signal that Andromeda cannot already infer from browsing behaviour. A seed of 300 B2B customers who work at mid-market SaaS companies and visited the brand's docs three times contains signal that is behaviourally rare and extremely hard to extract from broad targeting. When signal density in the seed is high relative to broad availability, a lookalike extends that signal to a larger lookalike-pool audience in a way broad targeting cannot replicate because broad does not have the seed to anchor on. This is the only condition that matters for whether lookalikes still work.

1) Small-seed niche brands with fewer than 500-1000 qualified customers. The customer base is specific enough that broad targeting cannot find more like them, but the seed is still large enough to anchor a 1-3% lookalike. 2) High-ACV B2B accounts where each customer in the seed represents genuinely rare behavioural intent. A seed of 400 customers of a $50k ACV SaaS product contains more signal density than a 50,000-customer seed of a $30 DTC product. 3) Strong-identity DTC categories (gaming hardware, home brewing, specialty cycling, specific craft hobbies) where buyers cluster behaviourally around shared platforms, publications, and purchase patterns broad delivery cannot infer from general browsing alone. Across all three, the pattern is the same: signal density high enough that the seed teaches Meta something broad does not already know.

High-ACV B2B accounts need lookalikes because the broad targeting pool contains too few qualified decision-makers for Meta's optimisation to converge on them. A product selling to compliance officers at mid-market financial services firms has maybe 15,000 qualified people in the US. Broad targeting across 200 million US adults wastes impressions on 199.985 million wrong people before finding the right ones. A 1-5% lookalike seeded on existing customers narrows the delivery pool to roughly 2-10 million people with behavioural similarity to known buyers, which Meta can optimise against. Meta's B2B CPM at around $4 versus LinkedIn's $25-60 only translates to real cost advantage when targeting narrows enough to deliver qualified impressions, and in high-ACV B2B that narrowing almost always requires lookalikes.

The useful seed size depends on signal density, not raw count. For small-seed niche brands and high-ACV B2B, 300-1000 customers in the seed is often enough if the seed is clean — recent buyers only, no bulk imports, no stale emails. For strong-identity DTC categories, 1000-5000 customers works well because the identity signal compounds across a larger seed without diluting. For most accounts, 1% lookalikes beat 5% because the smaller percentage preserves more of the seed's signal density. The 10% lookalike rarely wins in 2026 — at that pool size, the lookalike is effectively broad with a nominal seed anchor, and Andromeda's broad delivery outperforms it.

1) Classify the account against the three exception types above. If none apply, default to broad and skip the test. 2) If one applies, build a fresh lookalike using the cleanest seed available (recent buyers, verified emails, no stale data). 3) Run a 28-day paired test: one broad ad set, one 1-3% lookalike ad set, matched budget and creative. 4) Measure blended CPA at day 28, not platform-reported ROAS. 5) If lookalikes beat broad by more than 15 percent on blended CPA, keep lookalikes as the primary targeting layer. 6) If broad beats lookalikes or ties, abandon lookalikes for this account and revisit only if the signal density of the seed changes (new product line, new customer segment).

Lookalikes are not universally dead. They are structurally dead in the account class the broad-wins rule tested on and alive in the three exceptions: small-seed niche brands, high-ACV B2B, and strong-identity DTC categories. The practical next step is to classify the account against the three criteria before adopting the broad-wins default. In Advize's experience, two in ten accounts fall into one of the exceptions and are being under-served by the universal advice to abandon lookalikes.