Last September I sat with a client whose Meta ROAS had dropped from 4.1x to 2.6x in eight weeks. No changes to creative, budget, structure, or targeting. The platform looked broken. We dug in. Their dev team had pushed a tag-manager update that quietly broke 31% of their pixel events. Andromeda had been making decisions on a third of the signal it used to have. The auction wasn’t broken. We were starving it.
That’s the part of the AI-in-performance-marketing story almost nobody is telling. The smarter the system, the more its performance hinges on the quality of what you feed it. Bad signal used to give you mediocre delivery. Now it gives you a collapse, and you’ll spend three weeks blaming the platform before you find the real cause.
Conversion API, pixel quality, and signal density
The Conversion API isn’t optional anymore. I’d put it more strongly: if you don’t have CAPI properly implemented with deduplication against the pixel, you are giving Meta a deliberately worse picture of your customer than your competitors are. The model performs accordingly.
The threshold I look for on a healthy account is an Event Match Quality score of 8.0 or higher on Purchase events, with at least 95% of events matched between pixel and CAPI. Below 7.0 and you’re leaving 15–25% of attainable performance on the table, not because the events aren’t firing, but because Meta can’t confidently associate them with the users who triggered them. That uncertainty propagates into the model. Lower-confidence conversions get weighted less in optimisation. The system learns slower. Everything that happens downstream is worse.
The fix is unglamorous and the agencies who don’t do it can hide that fact for months. Audit your EMQ. Check parameter coverage: email, phone, name, IP, user agent (all hashed). Verify deduplication. Most accounts I audit are leaving easy gains on the table here, and the work pays for itself in a fortnight.
Why first-party data has gone from useful to essential
Direct answer in 50 words: first-party data (email lists, CRM data, lifetime value signals, in-app event streams) used to be a nice-to-have audience building tool. Under Andromeda, it’s a core delivery input. The system uses your CRM signals to find lookalike behaviour, model LTV, and prioritise users likely to convert at the value tier that matters to your business.
The brands winning on this are the ones treating their customer file like the strategic asset it is. They upload customer lists weekly, not quarterly. They segment by purchase recency, AOV, repeat behaviour. They feed back actual revenue and gross margin per customer where they can, not just purchase events. The system uses all of that to decide which user is worth bidding for, and how much.
The brands losing on this typically have a customer list of 40,000 buyers that was last uploaded eight months ago. Their LTV signal is whatever Meta can infer from the pixel, which is a fraction of the truth. Their lookalike modelling is anchored on stale data. The competitor with fresher signal is bidding more confidently on the same users and winning the impressions that matter.
The attribution shift you have to actually act on
Modeled conversions now make up a meaningful and growing share of what Meta reports. On most of my accounts, between 18% and 30% of attributed purchases are modeled rather than directly tracked. That’s not Meta being shady. It’s the inevitable consequence of iOS, browser privacy changes, and consent loss. The choice isn’t whether to live with modeled data. It’s how well-built your inputs are so the modelling works in your favour.
The shift means a few uncomfortable things. Your platform-reported ROAS will increasingly diverge from your GA4 and shopify numbers, and that’s not necessarily wrong. Meta sees post-view conversions GA4 won’t credit. GA4 sees direct-channel conversions Meta misses. The right answer isn’t to pick one platform as truth. It’s to look at MER (total revenue divided by total ad spend) and let the platforms argue about credit underneath.
The other thing this means: blended numbers are now the only numbers that matter for decision-making. Platform ROAS is useful for diagnosing the system, not for setting strategy. If your Meta ROAS goes from 4.2x to 3.8x but your MER holds at 3.1x, the platform isn’t underperforming. The modelling is just shifting attribution between platforms. Don’t pull budget for that. You’d be solving a measurement artefact, not a business problem.
How signal quality affects delivery and scaling
This is the part that surprises people most. Signal quality doesn’t just affect what gets reported. It directly affects what the system delivers next.
Andromeda treats high-confidence conversions as stronger optimisation signals than low-confidence ones. An event with a 9.5 EMQ score, full parameter coverage, and CAPI-verified delivery gets used to find more users like that one. An event with a 6.5 EMQ score and three missing parameters might still count toward your ROAS reporting, but it does less work in shaping where the system spends your next pound.
Across the accounts I’ve audited that had real signal problems (broken CAPI implementations, missing parameters, stale CRM uploads) the average lift from cleaning signal up properly has been 12–18% on blended CPA inside 30 days. That isn’t a tactic. It’s giving the system the data it needs to do the job you’re already paying it to do.
The other implication: scaling capped by signal quality looks identical to scaling capped by demand. Both manifest as performance softening as budget rises. Most brands diagnose this as a demand problem, pull back budget, and miss the actual cause. Before you accept that an account has hit a ceiling, audit the signal stack. The ceiling often isn’t real. It’s just where the data got too sparse for the model to keep up.
Where to put the work
The work isn’t sexy and it doesn’t make a good case study. Audit CAPI implementation and Event Match Quality. Push parameter coverage above 95%. Re-upload your customer file weekly with the richest segmentation you can produce. Stop arguing about platform-by-platform ROAS and start looking at MER. Trust modeled conversions enough to act on them, but only after you’ve made the inputs as good as they can be.
The brands who do this work will look like they have an unfair advantage on Meta in 2026. They won’t. They’ll just have hardened the inputs everyone else is still ignoring.
If you suspect your signal stack is the thing capping scale and you’d like a proper audit, I’m easy to reach.
