| Contrary to the belief that advertising is less data-driven, the complexity and dependency on feature-rich data sets has increased over time. Redundant targeting sometimes happens due to the ad objective of maximizing expected revenue over cost. However, it's more nuanced in practice. Advertisers aim to optimize expected revenue over cost within a 'payback period', which is typically 1-3 years for large advertisers. This is calculated through customer acquisition cost, retention, and incrementality (the probability of an ad causing conversion). Only advertisers themselves can effectively calculate incrementality due to access to their specific conversion/retention metrics. This incrementality, along with the optimal mix of channels and ad spend, is the ongoing challenge for sophisticated advertisers. It involves multi-objective optimization across millions of ad assets, campaigns, and targeting criteria. Privacy regulations since 2015 and subsequent laws like GDPR and CCPA have led to more reliance on probabilistic modeling for targeting. The entire pipeline of targeting, engagement, conversion, and retention forecasts are now based on probabilistic models. While ad networks offer simplified scaling solutions like 'target roas' and 'campaign budget optimization', they're more useful for average advertisers with limited internal resources - eg, they can't justify hiring ML SWEs, quant traders, technical PMs, etc. Advertising has become even more data-driven and arbitrage gains for sophistication have increased. Profound gains can be made with investments in marketing and forecasting science, similar to the operations of a quant trading firm. Source: I've managed $10B+ through automated ad spend systems since 2012. |
You already see this with some goods so its certainly at least somewhat effective, e.g. luxury watch ads in beverly hills contrasted with military recruitment billboards in south LA.