All terms

Lookalike Audiences: How Seed Quality Determines Performance

A lookalike audience finds new users who share characteristics with your best existing users. Learn how to build high-quality seeds, choose the right similarity percentages, and what changed after iOS 14.5.

Jay Ma
8 min read
Lookalike audience building: how seed quality determines paid media performance
On this page

A lookalike audience is a group of users identified by an ad platform that shares behavioral and demographic characteristics with a seed audience you provide. Instead of targeting users based on interests or demographics, you give the platform a list of your best users and ask it to find more people who look like them.

This is one of the highest-leverage levers in paid acquisition -- when the seed is right. The logic is direct: if you can identify what your best users have in common, the platform can use that pattern to find new users with similar characteristics. The quality of the output depends entirely on the quality of the seed. A lookalike audience built from your best 1,000 retained users is a fundamentally different tool than one built from your 50,000 total app installers, even if both technically qualify as "lookalike audiences."

The Seed Determines Lookalike Performance More Than Any Platform Setting

Most teams build lookalike audiences from all converters -- everyone who installed, signed up, or purchased. That is a reasonable starting point but a weak seed. When you include low-value converters in your seed, you dilute the signal with users who converted once and never returned. The algorithm then finds users who look like that mixed population, which includes significant churn risk baked in from the start.

High-performing teams build seeds from specific behavioral cohorts:

High-LTV users: Your highest-value cohort by 90-day revenue. Cohort data consistently shows these users have a different profile than average payers -- different device usage patterns, different engagement frequency, different acquisition channels. Seeding a lookalike from this cohort consistently outperforms a general payer seed by 20-40% on downstream LTV, because the algorithm is looking for users likely to become high-value rather than just likely to convert once.

Retained users: Users still active after 30 days. Retention is a strong LTV proxy when detailed LTV data is not yet available. Playco used 30-day retained players as their seed and saw CPI drop 28% compared to their day-1 install lookalike, with higher session depth and purchase rates in the two weeks post-install. The retained player cohort was smaller but cleaner -- it excluded all the immediate churn that inflated their general install list.

Subscription renewers: For subscription products, users who have renewed at least once have demonstrated willingness to pay repeatedly. This cohort makes a better seed than first-purchase users because it selects for the behavior that actually drives LTV: renewal.

Recent high-value purchasers: For e-commerce and in-app purchase scenarios, segmenting by recent purchase value rather than all purchasers filters out promotional buyers who only convert during sales events and then disappear.

Minimum seed size: Meta recommends at least 1,000 users for reliable lookalike modeling. In practice, 5,000+ gives more stable results, and 10,000+ produces the most consistent performance. Smaller seeds can work for products with genuinely small user bases but tend to produce less stable audiences that need longer observation windows before optimization decisions can be made with confidence.

Meta Lookalike Percentages: What the Scale Actually Means

Meta's lookalike audiences use a 1-10% similarity scale based on the total adult population of a target country. A 1% lookalike in the United States contains roughly 2.2 million users most similar to your seed. A 10% lookalike contains approximately 22 million.

The tradeoff is precision vs. scale:

1% lookalikes work best for performance campaigns where conversion quality matters more than reach. They are smaller audiences, so they exhaust reach faster, but match quality is higher and CPMs are typically lower because competition for smaller audiences is more limited.

3-5% lookalikes are the most common starting range for prospecting campaigns. They balance similarity with enough scale to allow the platform algorithm to optimize effectively. The algorithm needs volume to learn; a 1% lookalike in a small market may not provide enough impressionable users for the system to gather meaningful signal within a reasonable timeframe.

7-10% lookalikes work for brand awareness and reach campaigns where volume is the goal. Conversion rates are lower, but CPMs are usually also lower, which can produce acceptable CPAs for upper-funnel campaigns.

In practice, running multiple lookalike percentages as separate ad sets simultaneously lets you understand where quality drop-off happens for your specific product and market. Some categories show minimal performance differences between 1% and 5%; others show sharp quality drops above 2%. Testing is the only way to know where the line sits for your audience.

How iOS 14.5 and ATT Changed Lookalike Signal Quality

Before Apple's App Tracking Transparency rollout in 2021 and its continued enforcement through 2026, Meta's lookalike modeling drew on deterministic user-level data from a large proportion of iOS users. After ATT, a significant portion of iOS users opted out of tracking, which means Meta cannot directly observe their in-app behavior or attribute their actions to specific ad exposures.

The result: lookalike audiences built today include modeled estimates for opted-out iOS users rather than direct behavioral observation. The platform still functions and lookalike audiences still outperform cold interest targeting, but the signal is noisier than it was before 2021.

This makes seed quality more important, not less. Seeds built from first-party data matched through Meta's Conversions API -- using email address, phone number, or customer ID hashes -- are more reliable than seeds built from pixel-only tracking that has gaps for opted-out iOS users. The Conversions API matches directly against Meta's user graph using identifiers you control, without depending on browser-level tracking that ATT has degraded.

Hellyeah's Managed Growth platform addresses this through server-side data integration. AIMA connects your CRM and MMP data to Meta's Conversions API, so audience seeds are built from matched user records rather than incomplete pixel data. BeFreed, the AI reading platform, rebuilt their lookalike seeds using server-side CRM data after their pixel-based audiences degraded post-ATT. Their 90-day LTV lookalike audiences, rebuilt on cleaner Conversions API data, delivered 40% higher 60-day ROAS than their previous pixel-based equivalents -- not because the lookalike mechanism changed, but because the data feeding it was significantly more accurate.

Google Similar Segments and TikTok Lookalike Audiences

Every major ad platform offers a version of lookalike audience targeting, though the mechanics differ:

Google Customer Match and Similar Audiences: Google's Similar Audiences product was deprecated for most inventory in 2023 as part of third-party cookie deprecation preparation. In 2026, Google's first-party matching strategy is built around Customer Match, where you upload hashed email lists, phone numbers, or customer IDs, and Google matches them against logged-in Google users. From Customer Match audiences, Google's system can build reach expansions that function similarly to lookalikes. For app campaigns, Google's UAC (now rolled into Performance Max for apps) handles audience expansion automatically based on your app's install and engagement data.

TikTok Lookalike Audiences: TikTok offers a tier-based system -- Narrow, Balanced, and Broad -- analogous to Meta's 1%, 5%, and 10% ranges. TikTok's user base skews younger and the platform's conversion attribution is less mature than Meta's, which makes seed quality even more consequential. A seed built from TikTok-native converters -- users who both saw your ad on TikTok and converted within the TikTok ecosystem -- tends to outperform a seed imported from external sources, because the algorithm's model is built from the platform's own behavior data rather than cross-platform inference.

Apple Search Ads: ASA does not offer lookalike audiences in the traditional sense. Apple Search Ads Basic uses machine learning to identify users most likely to install your app based on your existing user base, but you do not provide an explicit seed. The mechanism is conceptually similar, but the control layer is entirely managed by Apple.

Building Lookalike Audiences Across Multiple Platforms

A common error in multi-platform campaigns is using the same seed audience on every platform. Your best users on Meta may behave differently from your best users on Google or TikTok because the platforms reach different populations and users who convert on each platform arrived through different discovery paths.

In practice, building platform-specific seeds based on users acquired from each platform tends to outperform a single cross-platform seed. A user who found your product through Google Search behaved differently in your app than one who came from a Meta prospecting campaign. Their in-app behavior, LTV trajectory, and engagement patterns often differ enough that a Meta algorithm trained on one cohort would benefit from a separate seed built from Meta-acquired users specifically.

Fish Audio, the AI voice platform, built separate high-LTV seeds for Meta and Apple Search Ads. Users acquired through ASA had different listening patterns and higher subscription renewal rates than Meta-acquired users, and those signals were invisible to the Meta algorithm when both cohorts were combined in a single seed. Building platform-specific seeds improved their prospecting efficiency on both platforms simultaneously.

Hellyeah's platform manages this cross-platform complexity. AIMA ingests acquisition data segmented by channel, builds platform-specific LTV cohorts, and uses those cohorts to inform audience strategy on each platform independently. The result is that each platform's lookalike system receives a seed that reflects its own audience dynamics rather than a blended average that fits no single platform well.

When Lookalike Audiences Deliver and When They Do Not

Lookalike audiences are worth prioritizing when you have a sufficient high-quality seed (at least 1,000 users, ideally 5,000+), when you have behavioral data to segment your seed by value rather than just conversion status, and when your product has enough user history for the platform's algorithm to identify reliable patterns.

They are worth deprioritizing when your seed is too small to model reliably, when your product has a user profile so specific that broad platform modeling cannot capture the right signal, or when you are in a market where your user base is demographically or behaviorally unusual relative to the general population the algorithm draws from.

For most performance teams running at meaningful scale, high-LTV lookalike audiences are one of the most reliable prospecting strategies available -- when built from the right seed and evaluated on downstream metrics like LTV and payback period, not just install rate or initial CPI.

Related resources: Audience Segmentation, First-Party Data, Retargeting, CAC, App Tracking Transparency.

Frequently asked questions

  • What makes a good lookalike audience seed?

    A good seed is a behavioral cohort, not just a conversion list. High-LTV users (your top 10% by 90-day revenue), 30-day retained users, or subscription renewers consistently outperform seeds built from all converters. Minimum recommended seed size is 1,000 users, though 5,000+ gives more stable results. The seed should reflect the users you actually want more of, not the users you happened to acquire.

  • How do lookalike audiences work after iOS 14.5?

    After Apple's App Tracking Transparency rollout, Meta can no longer directly observe the behavior of iOS users who opted out of tracking. Lookalike audiences built today include modeled data for these users, not directly observed behavior. This makes seed quality more important, not less: seeds built from first-party data matched via email or phone hash through the Conversions API are more reliable than seeds built from pixel-only tracking with iOS gaps.

  • What is the difference between a 1% and 10% lookalike audience?

    Meta's lookalike scale runs from 1% to 10%. A 1% lookalike contains users most similar to your seed -- smallest audience, highest match quality. A 10% lookalike is larger with lower precision. For performance campaigns where conversion quality matters, 1-3% is the standard starting range. For brand awareness and reach campaigns that need volume, 5-10% provides scale at the cost of similarity. Running both simultaneously as separate ad sets helps you understand where quality drop-off happens for your specific product.

Jay Ma

Co-founder

Co-founder of Hellyeah. Writes about building durable growth loops that compound over time.

Your next campaign, handled.

Hellyeah takes care of strategy, creative, and optimization.