How Our Predictive Models Reduced Wasted Ad Spend by 34%

We recently ran an internal analysis that we think every advertiser should see. We compared campaigns using our predictive bidding models against campaigns using traditional fixed-CPM bidding strategies across the same time period, same advertisers, same budgets, same targeting. Over 200 campaigns across display, native, in-app, and CTV over a six-month window.

The headline result: AI-optimised campaigns wasted 34% less budget on non-converting impressions. But the details behind that number tell a more important story about how programmatic advertising actually works and where the money goes.

What We Mean by Wasted Spend

Wasted spend is the portion of your budget that goes toward impressions which had near-zero probability of driving the outcome you care about. Every campaign has some of this. Not every impression will convert. That is expected.

But there is a difference between buying an impression that had a reasonable chance of converting and did not, versus buying an impression that was never going to convert in the first place. A banner ad served to a bot. A placement with 8% viewability where nobody could possibly see it. A thirteenth impression shown to a user who has already been served your ad twelve times without engaging once.

In a fixed-CPM model, you pay the same price for every impression regardless of quality. A high-value impression on a premium publisher costs the same as a low-value impression on a site with questionable traffic. The platform has no mechanism to differentiate between them. Your budget gets spread across all of them equally.

Our predictive models work differently. They score each impression individually and only bid on the ones with a meaningful probability of driving the desired outcome. Low-value impressions get skipped entirely. The budget that would have been spent on them gets redirected to better opportunities.

How We Measured It

We compared matched pairs of campaigns. Same advertiser, same budget, same targeting criteria, same creative, same timeframe. The only difference was the bidding strategy: one campaign used our predictive bidding models, the other used a traditional fixed-CPM approach with standard optimisation rules.

We ran this comparison across 200+ campaign pairs spanning e-commerce, SaaS, fintech, gaming, and consumer brands. The campaigns covered all four channels we support: display, native, in-app, and CTV. This was not a cherry-picked sample. We included every matched pair we could find over the six-month window regardless of vertical or budget size.

The Results

The numbers were consistent across verticals and channels:

• 34% reduction in spend on impressions that never led to any engagement or conversion event

• 22% lower effective CPA across the predictive bidding cohort

• Predictive campaigns reached 18% fewer unique users but generated 27% more conversions from those users

• The performance gap widened over time as models accumulated more campaign-specific data, starting at 15% in month one and reaching 40%+ by month six

That last point is particularly important. The longer a campaign runs on predictive bidding, the better the model gets at distinguishing high-value from low-value impressions for that specific advertiser. Fixed-CPM campaigns show no such improvement over time because there is no learning mechanism built into the bidding logic.

Where the Savings Came From

Three areas drove most of the efficiency gains:

Low-viewability placements. Our models learned to avoid placements with poor viewability history, even when those placements technically met all targeting criteria. A placement that matches your audience parameters but only achieves 15% viewability is essentially invisible inventory. Fixed-CPM campaigns kept buying these placements at full price because they had no mechanism to assess placement quality at the bid level.

Audience segments with low intent signals. Not all users in a targeting segment are created equal. Two users might both match your demographic and interest criteria, but one is actively researching products in your category while the other is casually browsing headlines. Our models learned to differentiate between these users based on contextual and behavioural signals, bidding more on high-intent users and less on low-intent ones. Fixed-CPM campaigns treated both equally.

Frequency waste. This was one of the biggest sources of savings. Predictive campaigns naturally pulled back spend on users who had already been exposed to the ad multiple times without engaging, reallocating that budget to fresh prospects who had not yet seen the creative. Fixed-CPM campaigns continued serving impressions to over-exposed users at the same rate, burning through budget on people who had already demonstrated they were not going to convert.

What This Means for Advertisers

A 34% reduction in waste does not mean 34% lower total spend. It means 34% more of your budget goes toward impressions that actually have a chance of driving a result. For a brand spending $50,000 per month on programmatic, that is roughly $17,000 per month redirected from dead impressions to productive ones. Over a year, that is $204,000 in budget that is working instead of being wasted.

The 22% CPA improvement is the downstream effect. When more of your budget reaches the right people in the right environments, your cost to acquire each customer drops meaningfully. And because the models keep improving over time, that advantage compounds quarter after quarter.

This is why we built Adxe around predictive models rather than rules-based bidding. Every impression is an investment decision, and our system treats it that way. If you want to see what the difference looks like on your specific campaigns, we are happy to run a comparison.