More Conversions, Lower CPAs: How Ad Algorithms Really Work

Google and Meta’s ad algorithms are machine learning systems. They predict which users are most likely to convert, what to bid for each auction, and where to place your ads. Those predictions are only as good as the data they’re trained on—and that data comes from your events.

More events means a more accurate model. A more accurate model means better targeting, lower CPAs, better ROAS, and more efficient spend. This isn’t abstract. Both platforms publish the thresholds and the numbers.

Every event is a training signal

Each time an event reaches an ad platform, the algorithm updates its model. “This user, with these characteristics, from this placement, at this time—converted.”

With few events, the model is still learning. It explores broadly, testing different audiences, placements, and times. Delivery is volatile and CPAs are high because the algorithm is paying to learn.

With more events, patterns emerge. The model identifies which people, devices, and times of day correlate with conversions. It shifts from exploration (expensive guessing) to exploitation (targeted delivery). CPAs fall, ROAS improves, and performance stabilizes.

The thresholds both platforms use

Meta’s Learning Phase exits when an ad set accumulates roughly 50 optimization events within a 7-day window. Until then, Meta marks the ad set as “In Learning” and explicitly warns that CPAs will be higher and delivery less stable. Ad sets that exit Learning achieve roughly 19% lower CPAs than those stuck in it.

The inverse is also true. Accounts with less than 20% of budget stuck in Learning see approximately 68% lower CPAs than those with more than 50% in Learning—a direct measure of how much fragmented campaigns cost you.

Google’s Smart Bidding follows the same pattern. Campaigns need roughly 30–50 conversion events to calibrate before Smart Bidding can optimize effectively. The system never stops learning, but the biggest efficiency gains happen in that initial calibration window. Campaigns with richer conversion history ramp up faster and hit target CPAs and ROAS more reliably.

What more events actually unlock

At low event volumes, algorithms have no choice but to guess. At higher volumes, the model can identify genuinely useful patterns—which audiences convert, which placements waste budget, which times of day produce high-value conversions.

This has a direct impact on targeting. Google’s Smart Bidding can identify nuanced signals across queries, devices, and user behavior and adjust bids in real time. Meta’s Andromeda algorithm can test and exploit patterns faster. Both platforms bid aggressively where conversions are likely and conservatively everywhere else.

More volume lets the algorithm allocate budget better, not just more broadly—which means scaling spend doesn’t mean proportionally higher CPAs.

The compounding problem with missing events

Here’s where this becomes a practical issue: a large share of actual events never reaches the ad platform.

Safari’s ITP deletes the cookies that link ad clicks to conversions after just 24 hours. Ad blockers prevent tracking pixels from firing. Cross-device journeys break the attribution chain. The result: ad platforms can miss 30–50% of your actual conversions.

For the algorithm, this is the same as those events never happening. It doesn’t know what it’s missing. It optimizes on the incomplete picture—and those invisible conversions can’t contribute to the learning model.

If 40% of your conversions are invisible to the algorithm, you’re not just missing in reporting. You’re effectively running on half the training data, which means slower learning, higher CPAs, and a model that’s built around the minority of conversions it can see rather than your full customer base.

How Aetra closes the gap for Segment users

For teams using Segment, the most common sources of missing conversion data are click IDs expiring before the conversion happens, first-party matching data that never gets attached to events, and identity resolution that breaks across sessions and devices.

Aetra captures click IDs server-side (bypassing Safari ITP’s expiration limits), collects first-party matching data from your existing Segment events, and enriches every conversion before it reaches Google, Meta, or any other ad platform supported by Segment. More of your actual conversions reach the algorithm, with the signal quality needed to train accurate models.

The difference shows up directly in performance. Teams using Aetra see 30–50% higher match rates and 25% lower CPAs—the result of feeding ad algorithms the complete conversion data they need to do their job.

After setting up Aetra, we doubled orders in Google Ads and saw a 36% increase in leads on Facebook
Stefan Harvalias, CMO, Tawkify

The algorithms can only optimize what they can see. Use our free calculator to estimate how many conversions you’re losing and what recovering them could do for your CPAs.