CBCesar Baz
Insights Arquitectura de Adquisición
Arquitectura de Adquisición

How to spot when Meta
is optimising for the wrong event

The system reports conversions, CPL looks steady, but the sales don't add up. The algorithm is learning from data that doesn't represent what actually happened.

Cesar Baz·Mayo 2026·9 min

The hardest thing to diagnose in an acquisition system is not when everything looks bad. It is when everything looks good and the business still doesn't grow. The Meta dashboard shows conversions climbing, CPL steady, the algorithm "active" — and yet the sales team reports poor-quality leads and revenue doesn't move.

That is almost always a sign that the algorithm is optimizing toward the wrong event. It isn't making a mistake — it is doing it because you told it that was the right goal.

Why it is hard to spot from the inside

The problem with optimizing for the wrong event is that the system becomes genuinely good at producing that event. If the goal is "form filled", the algorithm will build a refined model of audiences that fill out forms easily. Event Match Quality rises, signal coverage looks healthy, conversions climb week over week.

Everything says the system is working. Except that it isn't producing the result that matters.

To catch it you have to leave the Meta dashboard and triangulate with real-world data: CRM, sales, operations.

The 4 diagnostic signals

Señal 01
Gap between Meta conversions and CRM records above 25%

Meta reports 400 conversions this week. The CRM receives 290 leads. The 110 difference is not imperfect attribution — it is the conversion event firing in situations that produce no real lead. Common causes: the confirmation page is reachable by direct URL, the form counts the start of filling as a conversion, or there is Pixel+CAPI duplication without proper deduplication.

Señal 02
Contact rate below 35%

When the sales team calls the day's leads and fewer than 35% pick up or reply, there is an audience quality problem. The algorithm is attracting people who completed the form but have no real intent to buy — perhaps because of misleading creative, perhaps because the form is too easy to fill without understanding what is being requested. Low contactability is the first operational feedback that the system is bringing leads that don't exist in practice.

Señal 03
Platform conversions grow but revenue does not follow

If over the last 60 days Meta-reported conversions grew 40% and revenue grew 8%, the system is getting more efficient at producing conversions that generate no business. That disconnect accelerates: the more the algorithm refines its model toward the wrong event, the further the audience profile drifts from the real customer.

Señal 04
Event Match Quality is high but the real conversion rate is low

An EMQ of 8.5 means Meta can match conversions to real profiles very well. A high EMQ with low real conversion is a specific signal: the system knows exactly who completed the event, but that event does not represent the people who buy. The problem is not technical — it is strategic. The event that is right technically is the wrong one for the business.

The diagnosis in one question: How many of the "customers" Meta reports as converted show up as paying customers in your billing system or CRM for the last month? If the answer is "I don’t know" or there is a large gap, the conversion event is not measuring what you think it measures.

The three most common failure scenarios

Scenario 1 — The confirmation page is reachable by direct URL. The Pixel fires on /gracias.html whenever a user lands there. If that page has no access restriction (a check that the visit follows a completed form), any direct visit — from an internal email, from a bookmark — counts as a conversion. Easy to detect: look for conversions recorded outside working hours or from known IPs.

Scenario 2 — The form counts premature events. The "lead" event fires when the user clicks submit — before the server confirms receipt. If the form throws a validation error, the conversion is already recorded even though the lead never arrived. The right fix: fire the conversion event from the server (CAPI) only when the CRM confirms receipt, not from the browser on button click.

Scenario 3 — Duplicate firing without deduplication. Pixel and CAPI both active, with no shared event_id between them. The same lead is recorded twice. The algorithm learns that this person converted twice — and weights them more heavily in the model. The result: the system overvalues certain user profiles that in truth reflect a technical error, not double intent.

What to do once the diagnosis is confirmed

The order of correction matters. Nothing changes all at once:

  • First — fix the technical tracking: Deduplication, restricted confirmation page, correct event firing. Without this, any change of conversion event is still built on dirty data.
  • Second — define the right event: Which event in the funnel correlates best with a real customer? In telecom: port-in started. In retail: purchase completed. In services: a qualified call booked. That event needs enough volume (at least 50 a week) for the algorithm to optimise against it.
  • Third — accept the degradation period: The algorithm loses the refined model it had for the previous event and re-enters learning. CPL will rise for 2–3 weeks. That cost is justified if the new event represents real customers.

Do your Meta conversions match your real sales?

If there is a gap, the system is optimizing against a false target. I review the tracking, the conversion event and its coherence with the CRM before touching budget or audiences.

View diagnosis process

References

Primary documentation for the platforms cited on this page:

Case figures refer to the periods and accounts described in each one; they are not promises of replicable results.

Por — Performance marketing, data and AI consultant

Publicado el · Actualizado el