Why Your Lookalike Audiences Plateau at 2% — and What Seed Quality Has to Do With It

You built a lookalike audience off Meta’s “Interested in Puzzle Games” segment. It worked at 1%. You expanded to 2%. Performance dropped. By 3%, it looked like broad targeting with extra steps.

This is the lookalike plateau, and almost every mobile gaming UA team hits it. The question nobody asks: is the problem your expansion percentage, or your seed?

How Lookalike Expansion Actually Works

Meta’s lookalike algorithm finds users who resemble your seed audience. The quality of the output is bounded by the quality of the input. If your seed is “people who clicked a gaming ad once,” your lookalike is a statistical expansion of ad-clickers — not spenders, not retained players, not high-LTV users.

Interest categories are inferred signals. “Interested in Puzzle Games” might mean someone:

• Clicked a puzzle game ad
• Liked a Facebook page about Candy Crush in 2019
• Watched 3 seconds of a puzzle game video

None of these behaviors predict whether someone will install your game, play past day 7, or spend money.

What Behavioral Seed Data Looks Like

Contrast that with a seed built from actual behavioral data:

• 122M device IDs of users who actively play puzzle games — not inferred, observed
• Scored by real in-app behavior: session frequency, spend patterns, app diversity
• Filtered by quality tier: high spenders, power engagers, premium device owners

When you seed a 1% lookalike with 50,000 users who demonstrably spend money in puzzle games on premium devices, Meta’s algorithm has a fundamentally different signal to expand from. The statistical model isn’t extrapolating from ad clicks — it’s extrapolating from actual spending behavior.

Why This Matters at 2-3% Expansion

At 1%, even a mediocre seed produces a decent lookalike because the algorithm is being highly selective. The quality gap shows up at expansion:

• Interest-category seed at 2%: The algorithm runs out of strong matches fast because the original signal was weak. It starts pulling in users who vaguely resemble ad-clickers.
• Behavioral seed at 2%: The algorithm has a richer signal to extrapolate from. Users who resemble actual spenders look different from users who resemble ad-clickers — even two layers of statistical expansion out.

This is why UA teams see their CPAs spike at 2-3% with interest-based seeds but hold steady with behavioral seeds. The expansion curve is flatter when the starting signal is stronger.

The Practical Test

Here is a straightforward way to test this:

1. Take your current best-performing lookalike (probably seeded with your own first-party purchaser data or an interest category)
2. Build a parallel lookalike seeded with behavioral audience data — actual puzzle game spenders, filtered by quality tier
3. Run both at 1%, 2%, and 3% expansion
4. Compare CPA curves, not just CPA at a single expansion level

The gap between the two curves is your seed quality tax. For most puzzle and casual game publishers, this gap is 15-40% at the 2% expansion level.

Where to Get Behavioral Seed Data

This is exactly what Onmo Signal provides. 273 pre-built audience segments built from 300M+ device IDs scored by real app behavioral data across 170,000+ apps.

For puzzle game UA specifically:

• 122M puzzle game players (US primary)
• Segmented by spend behavior, engagement intensity, device quality
• One-click upload to Meta, TikTok, Snapchat, or Google as Custom Audiences
• Use as direct targeting or as lookalike seeds

Browse the catalog and upload your first audience at signal.onmoai.com. Free right now — no upload limits during the launch period.

---

Your lookalikes are only as good as what you seed them with. Stop expanding from guesses. Start expanding from signal.