How big should your seed audience be

Quick Answer

The minimum Meta-recommended seed size for a Facebook lookalike is 100 people, but 100 is not usable in practice. The real minimum is 500 people for a stable lookalike. The sweet spot is 1,000 to 50,000 people — large enough for statistical reliability, small enough to preserve signal concentration. Beyond 50,000 you start diluting quality. If your seed is smaller than 500, widen the source (e.g., 180 days instead of 90) or fall back to a different seed type.

The numbers Meta gives you and what they actually mean

Meta's documentation says you can build a lookalike from as few as 100 people. Technically true. In reality, anything under 500 creates an unstable lookalike — the model cannot find reliable patterns from so few data points, and performance varies wildly.

Here is how seed sizes actually perform:

• Under 100: Lookalike cannot be built or will be unusable
• 100 to 499: Builds, but unstable. Avoid if possible.
• 500 to 999: Usable. Moderate signal quality.
• 1,000 to 10,000: Sweet spot. Strong signal, statistically reliable.
• 10,000 to 50,000: Still strong. Slightly more signal dilution.
• 50,000 to 500,000: Diminishing returns. Signal starts to blend with the broader population.
• Over 500,000: Too broad. The model has too much noise to find tight patterns.

Why seed size matters so much

Meta's lookalike model works by finding patterns that distinguish your seed from the general population. Tiny seeds produce tiny patterns — or worse, random ones. Huge seeds contain everyone from power users to drive-by browsers, so the patterns become muddy.

Think of it as a signal-to-noise ratio. Below 500 people, the noise dominates. Between 1K and 10K, the signal is clean. Above 50K, signal starts to fade as you dilute the most distinctive features.

Step-by-step seed sizing

1. Count your current seed. Open Audiences in Ads Manager. Look at the size column for your custom audience source.
2. If under 500: Widen the time window (from 90 days to 180 or 365), combine multiple seed sources with OR logic, or use a broader event type (e.g., AddToCart instead of Purchase).
3. If 500 to 1,000: Usable. Expect the lookalike to work but not to be world-class. Keep building the seed over time.
4. If 1,000 to 50,000: You are in the sweet spot. Focus on quality filtering rather than size. Is every person in the seed a person you want more of?
5. If over 50,000: Consider tightening. Filter to top-quality segments — high LTV customers, recent purchasers, most engaged users — to sharpen the signal.
6. If over 500,000: Almost certainly too broad. Aggressively filter down to your best cohorts.

Real examples

Brand new DTC skincare brand. 112 purchasers. Built a lookalike anyway. Performance was erratic — some days CPA was £12, other days £60. Rebuilt the seed 60 days later with 380 customers. Still unstable. At 620 customers, the lookalike finally stabilised at £18 CPA.

Established ecom brand. 340,000 purchasers in the last 180 days. Lookalike of all purchasers: CPA £32. Filtered to top 10% by LTV (34,000 customers): CPA £22. Same seed source, 10x smaller, much better performance. Bigger was not better.

B2B SaaS. 68 paying customers. Too small for a stable lookalike. Widened to include qualified sales opportunities (closed-won + stage 3+ opportunities) — 310 contacts. Still below 500. Widened further to include all demo requesters from qualifying companies — 540 contacts. Finally big enough. CPL stabilised within 14 days.

When small seeds actually work

Small seeds can work in two cases:

1. Lookalike size is widened. A 500-person seed might fail at 1% but work at 3% or 5% because the lookalike pool is big enough to deliver consistent budget.
2. Multi-country lookalikes. Combining countries gives you a bigger effective lookalike pool even from a small seed.

Neither is ideal. They are workarounds for when you cannot grow the seed.

FAQs

What is the absolute minimum to build a lookalike?
Meta allows 100 people minimum, but practical minimum is 500 for stable performance.

Can a seed be too big?
Yes. Over 500,000 you start seeing signal dilution. Over 5 million, the lookalike is barely different from broad targeting.

Should I include old customers in the seed to grow it?
Only if they still represent your current ICP. Customers from 3 years ago might behave differently or be different kinds of buyers. Test including them vs excluding.

Is it better to have 300 very high-quality customers or 2,000 average customers in the seed?
2,000 average customers. Below 500, statistical reliability wins over signal quality.

Does Meta refresh the seed automatically?
Meta refreshes the lookalike automatically — it updates the people in the lookalike based on your seed. But the seed itself (your custom audience) only updates if you rebuild it.

Quality seed + quality creative = winning ads

A well-sized seed is only half the job. The other half is creative that resonates with the people Meta finds for you. Pix-Vu tests ad creative against the audiences your seed produces so your ads convert the prospects you worked hard to target. Try Pix-Vu free at https://pix-vu.com.