Detection: 1% lookalike at scale — broader audience worth testing

Key: lookalike_size_suboptimal Severity: Medium Confidence: 70%

What this detection looks for

We flag an ad set as lookalike-size-suboptimal when all of these are true:

The ad set is active
The ad set's targeting includes a 1% lookalike audience
The audience's approximate size is at or below 5 million users
The ad set has spent at least $1,500 in the audit date range

Why this matters

Lookalike audiences in Meta are sorted by similarity to the seed audience. A 1% lookalike targets the top 1% of users most similar to the seed; a 3% lookalike targets the top 3%. The 1% audience is more "similar" but smaller. The 3–5% audience is larger but only marginally less similar.

For accounts that spend at scale ($50/day+) against a 1% lookalike of a country of 100M users, the effective audience is 1M people. Meta's delivery system will exhaust the high-intent portion of that 1M quickly and start serving impressions to whichever sliver remains. Frequency rises, conversion rate falls.

A 3% or 5% lookalike trades a small amount of similarity for 3–5× the audience size. In published Meta case studies and practitioner reports, broader lookalikes at the same spend often deliver CPA within 5–10% of the 1% with much more headroom — and in many cases the broader lookalike outperforms the 1% because the delivery system has room to optimize.

This is one of the cheapest, lowest-risk tests in the playbook. The finding is intentionally informational — surfaced at 70% so the user knows it deserves a test, not a same-day pause.

How we calculate confidence

This detection is surfaced at 70% confidence by design. The underlying pattern is observational (audience size × spend), not a measured outcome on this specific account. Below 90%, so a "what would change our mind" section is required.

How we calculate the estimated monthly cost

We treat 5% of the ad set's spend as the at-risk efficiency gap that broader-lookalike testing would resolve, projected to a 30-day month.

monthly_recoverable = (ad_set_spend × 0.05) × (30 / days_in_range)

This is intentionally conservative. The real upside is more often "scale headroom" than "lower CPA today," and headroom is hard to monetize precisely. We surface a small dollar number so the finding ranks correctly against more decisive ones rather than dominating the top-of-list.

What would change our mind

This finding can be a false positive in a small number of cases:

The 1% audience is already winning at scale. If CPA on this ad set is the best in the account and it has been stable for a month, do not test broader. The audience is delivering the highest-similarity users and that is exactly what you want.
You have already tested broader lookalikes recently and they underperformed. Lookalike performance shifts with seed quality. If you tested 3% within the last 60 days and it lost, the finding is informational only.
Small seed audience. If the seed list has under 1,000 people, the 1% lookalike is built from a noisy signal and broadening it amplifies the noise. Build a larger seed audience first.

How to fix it

Build a parallel 3% lookalike of the same seed audience in Meta Audiences.
Duplicate the current ad set. Replace the 1% lookalike with the 3% lookalike. Keep everything else identical — creative, budget, placement, optimization event.
Run both for 14 days with equal budgets. Compare cost per purchase, not cost per click.
If the 3% delivers CPA within 10% of the 1%, shift budget to the 3%. The volume headroom is worth the small efficiency cost.
If the 3% loses meaningfully, keep the 1% and try a 5% or a different seed audience instead. The 1% is a real winner.
As a general practice, always test broader lookalikes when an ad set on a 1% audience grows past ~$1,500/month, even before this detection fires.

What we look at to make this detection

effective_status on the ad set
The ad set's included_custom_audience_ids cross-referenced to snapshot.audiences, looking for entries with audience_type == "LOOKALIKE" and lookalike_ratio == 0.01
approximate_size on the lookalike audience metadata, fetched from the Meta audiences endpoint
spend on the ad set, summed across the audit date range

Source

This methodology page is generated from apps/api/app/services/detections/lookalike_size_suboptimal.py. The detection code is open for inspection. We do not have hidden rules.