Smart Bidding

How Smart Bidding Uses Structured Learning

As Patrick Gilbert explains in Never Always, Never Never, Google's smart bidding algorithms rely on structured learning. Unlike unstructured learning systems that explore data without predefined labels, smart bidding works with clearly labeled historical data. The system analyzes your conversion history, matching specific inputs (user demographics, device type, location, time of day, search query) with outputs (conversions and conversion values). This structured approach allows the algorithm to identify patterns and make predictions about future auction outcomes.

The algorithm processes hundreds of signals simultaneously. Beyond basic targeting criteria, it considers factors like browser type, operating system, past site interactions, and even the competitive landscape of each individual auction. This comprehensive analysis happens in milliseconds, with the system calculating the optimal bid for each opportunity based on its learned understanding of what drives conversions for your specific account.

The Learning Phase Challenge

Every smart bidding campaign begins in a learning phase, a critical period that many advertisers misunderstand. During this phase, the algorithm actively explores different bidding strategies, testing hypotheses about which combinations of signals predict conversions. Performance appears volatile because the system is experimenting, not optimizing.

The learning phase is the period when an algorithm is actively gathering data, testing hypotheses, and figuring out which combinations of inputs drive the best outcomes. It's exploration mode.

Patrick Gilbert, Never Always, Never Never

Gilbert uses the analogy of Plinko from The Price Is Right to illustrate this process. Launching a smart bidding campaign is like dropping a disc from the top of the board. You set initial parameters, but the algorithm must bounce through various data signals, testing different paths before finding an optimal strategy. The disc might land in a good slot, but it might not be the best possible slot. This is why smart bidding sometimes requires periodic resets when market conditions change or performance plateaus.

Common Smart Bidding Strategies

Target CPA (tCPA) focuses on acquiring conversions at a specific cost per acquisition. The algorithm increases bids for users it predicts are likely to convert and decreases bids for those less likely to convert, maintaining your target average CPA across all conversions. This strategy works best when you have clear cost constraints and consistent conversion values.

Target ROAS (tROAS) optimizes for conversion value, not just conversion volume. The system calculates the expected value of each potential conversion and bids accordingly. If your products have varying profit margins or transaction values, tROAS helps focus spend on higher-value conversions. However, it requires accurate conversion value tracking to function properly.

Maximize Conversions attempts to drive the most conversions possible within your daily budget. Without a specific CPA target, this strategy can be more aggressive, especially useful during the early stages when you're still determining appropriate target costs. Maximize Conversion Value applies the same logic but prioritizes total conversion value over conversion volume.

The Importance of Campaign Liquidity

Smart bidding performance improves dramatically with what the Interactive Advertising Bureau calls "liquidity." Gilbert references their definition: when machine learning identifies the most valuable impressions by allowing every dollar to flow to the most valuable opportunity. This happens when humans remove constraints and let the system optimize freely.

Campaign consolidation directly impacts smart bidding performance. Instead of running multiple small campaigns with individual budgets, combining similar campaigns provides the algorithm with more conversion data and budget flexibility. Shared budgets and broad targeting give smart bidding systems the volume they need to make confident predictions and exit the learning phase successfully.

Confidence vs. Accuracy in Smart Bidding

Gilbert identifies a critical distinction between algorithmic confidence and accuracy. A smart bidding algorithm might be highly confident in its predictions while being completely wrong about actual outcomes. This typically occurs when conversion tracking is misconfigured or when the algorithm optimizes based on flawed data signals.

The most dangerous scenario is high confidence with low accuracy. Imagine a conversion pixel placed on an "add to cart" page instead of the purchase confirmation page. The algorithm confidently optimizes for users who reach the cart, believing it's driving purchases, while actual sales remain flat. This overconfidence can drain budgets quickly while delivering poor business results.

Monitoring the quality of data feeding smart bidding algorithms is crucial. Accurate conversion tracking, proper attribution windows, and clean conversion definitions ensure the system learns from meaningful signals rather than optimizing toward vanity metrics that don't drive business growth.

Related Terms