A Custom Algorithm Is Only as Good as the Planning Around It

By Phil Cowlishaw, Managing Director, Adobe Advertising

Custom algorithms are becoming a more important part of programmatic buying as advertisers look for ways to make media optimization reflect their own business goals and data. A model built around an advertiser’s specific signals will make bidding more relevant to the outcomes the business values most.

That said, custom algorithms alone do not guarantee better performance. When a campaign disappoints, teams often scrutinize the model first. In practice, the cause may be a poorly defined goal, too little data, an unrealistic target, or a campaign setup that limits what the algorithm can accomplish.

To achieve the best results with custom algorithms, advertisers should work through six areas in sequence.

1. Define the outcome precisely

A model needs a clear objective before it can be evaluated fairly. “Drive purchases” is too broad. Teams should identify the exact conversion event, agree on the primary KPI and target CPA, and document where the measurement signal comes from.

The attribution window should also be settled before launch, along with the campaign variables that will remain fixed during the evaluation. If goals or measurement rules keep changing, it becomes difficult to isolate the model’s impact.

2. Make sure there is enough signal

Signal volume and history need to be assessed before launch. As a rough floor, advertisers should look for at least 100 organic goal events per day and about 45 days of goal-event history. If retargeting is part of the strategy, roughly 45 days of audience history is useful as well.

 Conversion timing matters too. The shorter the path from decision to action, the faster the algorithm can learn which signals are associated with a successful outcome and optimize accordingly. For products with longer purchase cycles, advertisers should consider identifying meaningful intermediate actions, such as qualified visits, applications or other high-intent behaviors, that provide the model with more frequent feedback while still pointing toward the ultimate conversion.

A lack of DSP-attributed conversions at launch does not necessarily prevent a model from getting started. A strong organic conversion signal can provide enough information for initial learning.

3. Pressure-test the target and campaign

An algorithm can only optimize within the opportunity available to it. Before launch, advertisers should assess whether the target CPA is achievable given the audience, budget, and campaign design.

That starts with the audience. Teams should understand its size and reachability and how recently it has been refreshed. They should also examine the balance between retargeting and prospecting and whether lookalike expansion could add scale.

The campaign can impose limits as well. Attribution events need to fire correctly and map to the intended goals. Creative should fit the formats being bought. Targeting should leave enough room for delivery. Budgets, pacing, and bids should support the desired scale. Inventory and deal availability also need to match expectations.

If those conditions do not support the CPA goal, that should be surfaced before the test begins. Adjusting the target or campaign design upfront produces a more useful evaluation than asking the algorithm to overcome a structural constraint.

4. Build a fair comparison

Head-to-head platform tests are useful only when the conditions are comparable. Audiences and geographies should be equivalent, creative and flight dates should be aligned, and budgets should generally remain within about 5 to 10 percent of each other. Both sides should use the same attribution and conversion windows.

Core variables should remain stable throughout the test. Changing the conversion event, CPA goal, channel mix, or fundamental targeting midway through makes the result harder to interpret.

 When adding a new platform or partner, the most important question may be whether it creates incremental value. Two platforms can appear to perform well while effectively competing for the same pool of likely converters. A meaningful test should examine whether the new partner is finding additional conversions and expanding the total opportunity available to the advertiser.

If there is no incumbent and the algorithm is being tested against a performance target, strict platform parity is unnecessary. The benchmark and measurement assumptions still need to be locked before launch.

5. Give the model time to learn

Custom algorithms begin working immediately, but the effect may take time to appear in reported performance. Conversion lag means impressions served today may not produce measurable outcomes for days or weeks.

In many campaigns, meaningful movement in CPA does not become visible for two to three weeks. Early data is useful for spotting delivery or measurement problems, but a formal judgment should come only after the agreed attribution window and post-spend conversion lag have passed.

6. Refine one constraint at a time

Once enough data has accumulated, optimization should become more targeted. Thin conversion volume may point to signal quality or coverage. Limited scale may require audience expansion or refinement. Delivery problems may call for changes to campaign configuration.

The key is to make changes deliberately. Altering several major inputs at once may improve performance, but it also obscures what caused the improvement. Changing one meaningful variable at a time creates a clearer learning loop.

Custom algorithms can give advertisers a more direct way to translate their own signals and business priorities into media decisions. Their value depends on disciplined execution.

The best results will come from treating custom algorithms as an ongoing operating discipline. Define the outcome, establish the conditions for learning, measure carefully, and refine based on evidence. That is what turns a custom algorithm into a repeatable performance advantage.