When lookalike modeling first began gaining traction on media plans, advertisers and agencies alike were excited about the possibility of impacting campaigns by learning more about their existing customer base. However, we’ve seen limited success with this strategy in the direct response (DR) space.
Then there are predictive audiences. Many demand-side platforms (DSP) offer a proprietary predictive audience solution that gives more responsibility to the platform to find who it thinks is most likely to convert.
Lookalike modeling surfaces a blueprint of users based on a brand’s first-party data.
The PRO: It helps inform what target audience we should test to reach new but similar users in effort to drive incremental conversions.
The CON: It takes time to build a large enough data sample to run the algorithm. Our DSP partner, The Trade Desk, has a minimum threshold of unique cookies to model from to ensure statistically significant recommendations. While this approach can produce strong models, it also inhibits our ability to launch this tactic right away for new clients.
Predictive audiences surface who a potential new customer looks like, giving the reins to the DSP to search the internet universe and dynamically pinpoint audience segments it determines as likely to convert.
The PRO: There’s more data to ingest and model from, since it isn’t tasked to input only one brand’s 1st party data. This tactic can also be launched immediately for new clients.
The CON: Runs the risk of a programmatic trader “setting and forgetting” the tactic. The algorithm may also choose some pretty unrelated audience segments to test that are both expensive and ineffective.
At IMM, we maintain involvement in the process by removing very expensive or underperforming segments in the predictive audience. We also assess the top-performing segments and will oftentimes translate that across other strategies like private marketplace deal buys, native ad buys, programmatic skin buys and more.
- In the wireless category, we’ve seen varying success for both lookalike modeling and predictive audiences. For a large national wireless brand with 20,000+ unique cookies in the purchaser pool, lookalike modeling drove a 27% efficiency in online sales CPA compared to predictive audiences.
- For a smaller national wireless brand with 7,000+ unique cookies in the purchaser pool, predictive audiences drove a 64% efficiency in online sales CPA compared to the lookalike model and 13X the volume of sales.
- For a VoIP brand, predictive audiences drove a 29% efficiency in online sales prospecting CPA and contributed to 87% of prospecting sales.
- In the restaurant category, predictive audiences contributed to 77% of site action conversions, driving even more conversions than retargeting strategies.
In general, lookalike models have performed well for large brands with existing strong market awareness and daily online sales. It can also be a useful tool for small brands seeking to understand their current customer. But for smaller brands focused on growing conversions, predictive audiences have proven to be more lucrative.