Creative Lookalike Audience Signals Your Team Should Use
Creative lookalike audience signals worth feeding into seed selection, beyond raw installs, and how to check they still hold.
Most lookalike audience seeds are built from whoever converted, regardless of which creative brought them in. That is a reasonable default when a campaign is new and there is no other data to build from. It stops being a reasonable default once a portfolio has months of creative-level performance data sitting unused, because creative lookalike audience signals, the behaviour patterns tied to specific hooks and concepts rather than to the campaign as a whole, usually carry more useful information than an undifferentiated seed pool does.
The distinction matters because different creative concepts do not just convert at different rates, they tend to convert different kinds of players. A gameplay-first hook and a narrative-first hook for the same title can produce installer populations with meaningfully different retention and spend patterns even when their headline CPI looks similar, and a lookalike model trained on the blended seed pool from both concepts inherits neither pattern cleanly.
What "creative signal" means in a lookalike context
A creative lookalike audience signal, in practice, is any behavioural pattern that correlates with which creative concept or hook a player was exposed to before converting: which specific ad variant they saw last before install, how long they engaged with a playable's interactive moment before converting, or which of several hook framings (competitive, narrative, puzzle-mechanic) drove the click. As covered in this desk's earlier piece "Creative segmentation by player type, not just genre," these patterns often map more closely to actual player motivation than genre or demographic targeting does, which is exactly the kind of signal a lookalike model is meant to exploit rather than average away.
Building seeds from creative cohorts, not campaign totals
The practical shift is building separate lookalike seeds from each creative concept's converters, rather than one seed from the whole campaign, at least for concepts that have run long enough to generate a statistically usable population. A concept-specific seed lets the platform's modelling find players who resemble the people a specific hook actually reached, rather than an average of several different hooks' audiences that may not describe any single real player well. This only works once a concept has enough conversion volume to build a seed of meaningful size, which is the main practical constraint: a new or lower-spend concept may need to borrow from a broader pool until it earns its own seed.
Checking that the signal still holds
Creative lookalike signals are not permanent. A hook that reliably identified high-retention players in the spring can lose that correlation once the concept has been running long enough that its audience saturates or the concept itself gets iterated into a materially different execution. Three checks keep this from going stale:
- Re-validate each concept-specific seed's downstream retention against a fresh cohort at least once a quarter, not just at the point the seed was first built.
- Watch for a concept whose CPI stays flat while its post-install quality metrics decline, since that pattern often means the lookalike audience built from its early converters is no longer describing who the concept is actually reaching now.
- Retire a concept-specific seed once the underlying creative itself has been meaningfully reworked, rather than assuming the old seed still applies to the new execution just because the concept name did not change.
Creative lookalike audience signals: a worked example
Take a hypothetical mid-size publisher running two concepts for the same puzzle title over a six-week flight. Concept A is a competitive, leaderboard-framed hook; Concept B is a relaxing, no-pressure framing of the identical core mechanic. Both convert at a similar blended CPI, around $2.10, which is the number a campaign-level dashboard would report if the two concepts were never separated. Split by concept, Concept A's converters show a 21% day-7 retention rate and an average of $0.38 in day-14 in-app spend, while Concept B's converters show a 29% day-7 retention rate but only $0.19 in day-14 spend. Building a single blended lookalike seed from both concepts' converters produces a model chasing an audience that resembles neither pattern particularly well. Building two separate seeds, one from each concept's converters, lets the publisher scale Concept A's lookalike audience specifically when the priority is monetisation and Concept B's specifically when the priority is retention and volume, a distinction the blended seed could never make. The publisher in this example is hypothetical and the figures are illustrative, but the shape of the divergence, similar CPI masking materially different downstream behaviour, is the pattern this piece is arguing teams should expect to find once they split their own seed data by creative concept.
What to watch next
Two changes are worth watching as this practice matures. Ad networks are increasingly building their own creative-level audience modelling directly into automated bidding tools, and a network offering that natively could reduce how much manual seed-splitting a buyer needs to do, though as of this piece none of the major networks has published a creative-level lookalike feature that fully replaces the manual approach described here. Privacy-driven data limitations, particularly on iOS, continue to constrain how much creative-exposure data is available at the individual level for seed-building, which means the geo- and network-level workarounds teams have already adopted for other measurement gaps will likely need to extend to lookalike seed construction as well.
Where this fits into a media buying calendar
None of this replaces platform-side automated audience expansion, which most networks now push hard by default. It sits alongside it, as an input a buyer controls rather than a black box the platform manages alone. A team that feeds creative-specific seeds into its campaigns, and re-checks those seeds against fresh cohort data on a fixed schedule, gets more out of automated expansion than a team that lets every concept draw from the same undifferentiated pool and hopes the platform's model finds the distinction on its own.
