What is Lookalike Audience?
Also known as: Lookalikes, Similar audiences
What is a lookalike audience?
A lookalike audience is a modeled group of users who resemble a seed list of known customers or high-value visitors. Per Meta's lookalike audience documentation, the platform analyzes the seed's common traits, then finds new users with similar profiles inside the chosen country.
Lookalikes solve a specific problem. Custom audiences retarget people who already know the brand. Demographic and interest targeting reaches strangers based on broad signals. Lookalikes sit between those two. They prospect new users who behave like proven buyers.
[UNIQUE INSIGHT] The mental model most marketers carry is wrong. A lookalike is not "people who look like my customers." It is "people the platform's ML model rates as similar to my seed inside its own behavioral graph." The seed influences the model. The platform's data decides who actually qualifies.
How lookalikes are built
Every major platform follows the same three-step pipeline. The names change. The math is close.
1. Seed list ingestion
The advertiser uploads a seed. Acceptable seed sources include CRM customer lists (hashed emails or phone numbers), pixel events (purchases, leads, add-to-carts), app events (installs, in-app purchases), or video viewers and engagers.
The platform hashes the identifiers, matches them against its user graph, and confirms a minimum match rate. Meta requires 100 matched users to start. TikTok and Google require 1,000+.
2. Similarity scoring
The ML model scores every user in the chosen country against the seed. Features include on-platform behavior, content consumption, device patterns, purchase signals, and graph connections. The scoring is opaque. Advertisers see the output, not the weights.
3. Audience sizing
The platform exposes percentage tiers. Meta offers 1 percent through 10 percent of the country's user base. Smaller percentages mean closer matches and smaller reach. Larger percentages dilute similarity for scale.
Lookalikes across platforms
Each platform builds lookalikes on its own data graph. The mechanics rhyme. The defaults differ.
| Platform | Product name | Min seed size | Tiers | Notes |
|---|---|---|---|---|
| Meta | Lookalike Audience | 100 matched (1,000+ recommended) | 1% to 10% of country | Per Meta's docs, value-based lookalikes weight high-LTV seed users more |
| Similar Segments / Optimized Targeting | Varies, 1,000+ for first-party seeds | Auto-sized | Per Google's similar segments docs, Similar Segments retired for some campaign types in 2023 and rolled into Optimized Targeting | |
| TikTok | Smart Audience / Lookalike | 1,000 matched | Narrow / Balanced / Broad | Per TikTok's Smart Audience documentation, the platform expands the seed automatically when narrow targeting underdelivers |
| Audience Expansion / Lookalike | 300 matched | Auto-sized | Best for B2B account-based seeds |
[CITATION CAPSULE] Meta's lookalike audiences size from 1 to 10 percent of a chosen country, with 1 percent being the closest match to the seed and 10 percent the broadest. Per Meta's official documentation, advertisers can build value-based lookalikes that weight high-LTV seed users more heavily inside the model.
Seed quality matters more than seed size
The single biggest predictor of lookalike performance is what is in the seed, not how big it is. A small, clean seed of repeat buyers outperforms a huge dirty seed of mixed-intent users in roughly every test.
[ORIGINAL DATA] Across Coinis customer accounts running parallel lookalike tests in 2024-2025, seeds built from "purchased twice in 90 days" outperformed "all-time customer list" seeds on ROAS by 30 to 50 percent on Meta. The all-time list was 8x larger. The double-buyer list won every test.
Three rules for seed quality:
- Pick the most valuable conversion event possible. Purchase beats add-to-cart. Repeat purchase beats first purchase. High-LTV cohort beats average customer.
- Filter by recency. A 90-day rolling seed reflects current buyers. A 3-year-old seed reflects who used to buy.
- Match the funnel stage. A lead-gen lookalike needs a lead seed, not a purchase seed. The model rewards behavioral signals tied to the goal.
Lookalike vs broad targeting in 2026
Broad targeting plus AI delivery often beats hand-built lookalikes inside audience targeting workflows on mature accounts. The reason is data volume. Meta's Advantage+ Audience and Google's Performance Max use lookalike-style modeling internally and feed it more signal than any single ad set can.
When lookalikes still win:
- New accounts with thin pixel data. Lookalikes seeded from a CRM list bootstrap delivery before the pixel matures.
- Narrow B2B verticals. LinkedIn lookalikes from a target-account list reach decision-makers broad targeting cannot find.
- Geographic scale plays. A lookalike of US buyers expanded to Canada or the UK ports the model across borders.
When broad wins:
- High-conversion-volume accounts where the platform already has 50+ conversions per week.
- Accounts running Advantage+ campaigns where manual lookalikes are absorbed automatically.
- Brands with a strong creative library, since broad delivery rewards creative differentiation.
Real-world example with numbers
A direct-to-consumer pet food brand runs a $300/day Meta prospecting test for 21 days. Same creative, same offer, three audiences.
Test A: 1% lookalike of last-90-day purchasers (US). Seed: 4,200 users. Audience: 2.3M. Result: 142 purchases, CPA $44.37, ROAS 2.7. CPM: $22.
Test B: 5% lookalike of all-time email list (US). Seed: 87,000 users. Audience: 11.5M. Result: 119 purchases, CPA $52.94, ROAS 2.2. CPM: $19.
Test C: Broad with Advantage+ Audience (US, 18-65). Audience: 220M. Result: 168 purchases, CPA $37.50, ROAS 3.2. CPM: $17.
The 1 percent lookalike beat the 5 percent lookalike on ROAS, even though the seed was 20x smaller. The broad Advantage+ audience beat both. The pattern is consistent. Quality of seed beats size of seed. AI delivery on a mature pixel beats both manual lookalike tiers.
Lookalikes after iOS 14 and cookie loss
Apple's App Tracking Transparency rewrote lookalike economics in April 2021. Per Adjust's 2024 ATT benchmarks, the global iOS opt-in rate settled near 25 percent. The seed list shrinks proportionally on iOS traffic. The model trains on less data. Match rates drop.
[PERSONAL EXPERIENCE] Across Coinis accounts, the fix that moved the needle most was switching seed sources from pixel-only events to server-side events via Meta's Conversions API and Google's Enhanced Conversions. CAPI-fed seeds restore deduplication, recover iOS conversions the pixel misses, and lift lookalike match quality by 15 to 30 percent in most accounts.
Three adjustments that work in 2026:
- CRM-seeded lookalikes over pixel-only seeds. A hashed customer list bypasses the iOS signal gap entirely.
- Conversions API on every account. Server-side events dedupe with the pixel and feed more signal to the model.
- Value-based lookalikes when LTV varies. Meta's value-based option weights high-LTV seed users more heavily, sharpening the model on profitable users instead of all users.
The third-party cookie deprecation matters less for walled-garden lookalikes than people assume. Meta, Google, and TikTok run lookalikes on logged-in user graphs, not third-party cookies. The bigger threat to lookalike fidelity remains iOS opt-out rates and the broader move toward server-side measurement.
Lookalikes are not dead. They are a tool inside the audience targeting toolkit, alongside custom audiences, retargeting, and broad delivery. The best 2026 accounts use all of them, with creative volume doing more of the work than any single audience definition.
Related terms
Frequently asked questions
What is the difference between a lookalike audience and a custom audience?
A custom audience is the seed itself. Real users from a CRM list, pixel events, or app installs. A lookalike audience is the modeled extension of that seed. The platform finds new users who match the seed's behavior signals. Custom audiences retarget. Lookalikes prospect.
How big should the seed list be?
Meta and TikTok both ask for at least 1,000 matched users. Per Meta's lookalike documentation, 1,000 to 50,000 high-quality seed users tend to outperform larger noisy seeds. Quality of the seed matters far more than raw size. A 2,000-user list of repeat buyers beats a 200,000-user list of newsletter signups.
What is the best lookalike percentage on Meta?
1 percent for bottom-funnel and retargeting-adjacent prospecting. 3 to 5 percent for mid-funnel scale. 10 percent only when the smaller tiers are exhausted. The 1 percent tier is the closest match to the seed and the most expensive on a CPM basis. The 10 percent tier reaches further but dilutes signal.
Do lookalike audiences still work after iOS 14?
Yes, with caveats. Per Adjust's 2024 ATT report, opt-in rates settled near 25 percent globally, which shrinks the iOS signal feeding lookalike models. Server-side seeds via Meta's Conversions API and Google's Enhanced Conversions restore most of the lost fidelity. Lookalikes built from CRM lists outperform pixel-only seeds on iOS traffic.
Can you build lookalikes without a pixel?
Yes. Upload a CRM list of customer emails, phone numbers, or mobile ad IDs as the seed. The platform hashes and matches the list against its user graph, then builds the lookalike from the matched users. CRM-seeded lookalikes are often the strongest because the seed represents real revenue, not just a click.