How modelled conversions estimate the unseen

Why modelling exists#

Opt-outs, aggregation thresholds and platform restrictions hide a growing share of real outcomes from direct observation. Ignoring the hidden share would under-report systematically and starve learning; modelling estimates it from observed patterns, consented cohorts and historical behaviour instead. The alternative to estimates is not truth but a guaranteed undercount. Start at the Meta Ads hub for the family view.

Expect modelled share to grow over time as privacy defaults tighten, which makes reading discipline more valuable yearly rather than less. Accounts that reconcile against order books stay calm; accounts that worship native totals swing with every methodology tweak. Meta own modelled conversions guide describes the current method.

Where modelled counts appear#

Modelled fills blend into standard reporting columns rather than arriving in a separate bucket: reported conversions mix observed events with estimated ones, and breakdowns rarely flag which is which. AEM traffic, skipped attributions and small-cell aggregations are the heaviest modelled zones. Assume any opted-out or aggregated slice carries estimates.

This blending is why day-level reactions misfire: modelled portions settle and revise as aggregation completes, so early reads wobble. Judge in weekly cohorts and let the estimates land. Our AEM guide maps the aggregated zones.

Reading estimates honestly#

Treat modelled counts as ranges, not scalpel figures: direction and magnitude across weeks deserve trust, single-day decimals do not. Validate against your order book monthly and track the platform-to-truth ratio as its own metric; a stable ratio means healthy estimation, a drifting one names a configuration fault. Never optimise daily budgets on modelled wiggles.

Segment reads by traffic type: consented observed slices read precisely, aggregated slices read approximately, and blended totals sit between. Reporting that names the mix earns stakeholder trust that flat totals never will. Our mismatch guide reconciles the platforms.

Modelled plus observed in reports#

Build reports that separate what you know from what is estimated: observed event counts, modelled share estimates, finance reconciliation and the resulting confidence band. Stakeholders accept estimates gladly when labelled; they revolt when estimates pose as certainties and later revise. Honest columns cost nothing and save quarterly arguments.

Hold creative and audience tests to observed-friendly designs: geo holds and conversion-lift studies measure incrementality without leaning on modelled fills. Conversion Lift settles what modelling can only suggest. Developer modelling context is at AEM developer docs.

Improving the inputs#

Models are only as good as their fuel: full CAPI cover, clean dedup, ranked AEM slots with value sets, and consistent event parameters all sharpen estimates. Every observed event you restore through server signal is one fewer outcome left to guesswork. Measurement investment pays twice under modelling, once in counts and once in estimate quality.

Prioritise the fixes in order: server cover first, honest dedup second, parameter discipline third. Each lifts observed share and tightens the modelled remainder.

Shops with catalogues gain twice: matched content IDs sharpen both observed matching and the modelled estimates built on top. Feed health keeps the IDs honest.

What modelling cannot do#

Modelling cannot invent buyers, recover events never sent, or rescue broken plumbing: it estimates the unobservable, not the unreported. Accounts with missing CAPI, failed dedup and unranked AEM slots get confident-looking estimates built on sand. Fix the pipes before debating the estimates.

It also cannot settle incrementality: modelled credit still credits, and only holds prove causality. Pair modelled reporting with periodic lift tests and finance reconciliation. Attribution windows shape the credit, and our team reviews estimation health pre-peak.

Frequently asked questions#