Meta Andromeda: why signal quality decides your CPA now

Meta rebuilt its ad selection around an AI retrieval engine called Andromeda, rolled out through 2025. The practical consequence for advertisers is blunt: the quality of the conversion signal you feed Meta now matters more for your CPA than any targeting decision you can make by hand.

What is Andromeda?

Andromeda is the retrieval engine inside Meta's ad delivery system — the layer that scans an enormous pool of candidate ads and decides which handful are worth ranking for a given person at a given moment. Advantage+ is the surface you interact with; Andromeda is the engine underneath it. You don't configure Andromeda. You feed it.

And what you feed it is conversions. The system learns who converts, on what, after seeing which creative, from the conversion events your account reports. That's the entire interface between your business reality and Meta's machine.

Why does it amplify signal problems?

Because it's a feedback loop, and feedback loops compound their inputs. Andromeda trains on the conversions it receives. If that signal is thin, delayed, duplicated, or unmatched, retrieval gets worse — it surfaces your ads to the wrong people — which produces fewer real conversions, which means even less training signal. Bad signal doesn't just misreport your results anymore. It actively degrades delivery.

The failure modes, concretely:

• Thin signal: browser-only tracking loses conversions to iOS, Safari, and ad blockers. The optimizer learns from a biased sample of your buyers — the ones whose browsers happened to cooperate.
• Delayed signal: events that arrive hours late (batchy server setups, flaky retries) land outside the window where they're most useful for learning.
• Duplicated signal: Pixel and CAPI both firing without shared event IDs teaches the system that low-intent behavior converts twice as often as it does. See deduplication done right.
• Unmatched signal: a server event with no hashed email or phone may be delivered but never connected to a person. Meta counts it weakly or not at all — a ghost conversion.

Under the old system these were reporting annoyances. Under a retrieval engine that learns continuously from your events, they are a delivery problem with a CPA attached.

What's the practical checklist?

Four things, in order of impact:

1. Pixel + CAPI redundancy, deduplicated. Browser Pixel for the visits it can see, Conversions API from your server for everything, the same event_id on both so Meta keeps exactly one copy. Redundancy without dedup is worse than either alone.
2. One prioritized conversion event. Optimize campaigns toward a single, clean Purchase (or Lead) event. Multiple overlapping custom events split and muddy the learning signal.
3. High Event Match Quality. Send SHA-256-hashed email and phone on server events so Meta can match the conversion to a person. EMQ is the difference between a conversion that trains the model and one that evaporates — we break down what moves the EMQ score, and your live score for each event type is on /event-match-quality.
4. No double-counting. Audit that sent events and Ads-Manager-counted events actually line up. A gap in either direction means the optimizer is learning from a distorted picture.

Is "creative is the new targeting" actually true?

Half true. Creative diversity genuinely matters more under Andromeda — retrieval engines reward having many distinct candidates to match against many distinct people, so a wide creative portfolio outperforms three polished variants of one concept.

But the half everyone skips: the optimizer scores creative *by the conversions it attributes to it*. If your conversion signal is thin or mismatched, Meta can't tell which creative actually sells — it's grading your ads with a broken scale. Teams that pour budget into creative volume while running leaky tracking are testing into noise. Clean signal is what makes creative testing mean anything. You need both, and signal comes first, because it's the measurement layer for everything else.

This is the layer Caply runs for you: our snippet plus managed servers send Purchase and Lead to Meta CAPI with hashed email and phone, deduplicated by event_id against your Pixel, with retries so events arrive on time. Every event gets a Match Forecast score before Meta grades it, and Ghost conversions shows you the sent-vs-counted gap continuously. Run the free scanner to see how your current signal would look to Andromeda — it takes about a minute — or compare plans on pricing.

Andromeda didn't change what good tracking looks like. It changed the price of bad tracking.