Half your conversions are estimates. (now what?)

Between 20% and 50% of the conversions inside your ad dashboards were never observed. No pixel recorded them, no click ID came back, and no matching row exists in your order database. They are statistical estimates that Google and Meta calculate for journeys they can no longer track directly — printed in the exact same column, in the exact same font, as real verified purchases.

This is not deceit. It is the platforms' necessary response to privacy changes like Apple ATT and cookie restrictions. But when a major share of your reported conversions are educated guesses, evaluating marketing on dashboard numbers alone will lead you astray.

Why ad platforms started guessing

The tracking systems ad platforms built during the 2010s assumed a third-party cookie could follow a buyer across the web from first impression to checkout. Between 2017 and 2024, four privacy shifts broke that mechanism:

• Apple ITP restricted browser cookies to seven days, and to just 24 hours after ad-clicked redirects.
• App Tracking Transparency (ATT) made cross-app tracking opt-in on iOS, with opt-in rates hovering around 25%.
• Consent regulations in Europe and elsewhere meant millions of visitors decline tracking cookies entirely.
• Browser privacy defaults in Safari and Firefox blocked third-party cookies outright.

Ad networks faced two choices: report only what their tags could physically see — causing reported conversions to plummet — or infer the missing sales statistically. They chose to estimate.

What Google, Meta and GA4 actually model

Each platform fills in measurement blind spots using its own machine-learning models, but none of them separate estimates from observed sales in everyday reports:

Because modeled numbers are baked into the primary conversion column, you cannot simply filter them out. You have to understand how they behave.

Three ways modeled data distorts your reports

1. Row-level reconciliation is impossible by design

The most common mistake in marketing analytics is trying to match every Meta or Google conversion to an individual order row in your warehouse. When an analyst finds that 30% of platform conversions have no matching order ID, they assume the pixel is broken.

It is not broken. Those conversions correspond to no single order because they represent an aggregate statistical projection. You can only reconcile at the total volume and ratio level.

2. Small campaign slices become wildly inaccurate

Modeling works well on large populations because the law of large numbers smooths out errors. But on a niche campaign with only 15 conversions a week, modeled conversions are guesses applied to a tiny sample.

Rule of thumb: treat conversion counts as directional whenever a reporting slice has fewer than 100 conversions. Only rely on them for fine-grained budget shifts above that volume.

3. Traffic shifts mimic tracking regressions

Because modeled conversions depend on how much traffic is unobservable, changes in your audience mix change your discrepancy ratio even when your tracking tags are working perfectly.

Four rules for reporting estimated numbers

1. Let ad algorithms bid on their own numbers. Smart Bidding and Advantage+ need these modeled signals to evaluate auctions. Do not starve the bidding algorithm of its own data.
2. Report to leadership using database ground truth. Never send unadjusted ad-platform totals to the board. Use MER (total revenue ÷ total ad spend) and blended CAC. Both come from your bank account and warehouse, where modeling cannot enter.
3. Track the overclaim ratio as a health indicator. A platform reporting 1.3x your database orders is normal. What you are watching for is sudden week-on-week movement.
4. Check device mix before investigating tags. If the gap widens, inspect iOS share and consent opt-out rates before assigning developers to rewrite tracking code.

How to expand your verified conversions

You cannot turn off platform modeling, but you can dramatically reduce the share of your traffic that requires estimation:

To expand your first-party verified floor:

• Implement server-side conversion APIs. Meta Conversions API (CAPI) and Google Enhanced Conversions pass hashed customer data directly from your server, bypassing ad blockers and browser cookie limits.
• Capture customer identity earlier. Gathering an email address during lead capture or account creation provides a match key even if the buyer finishes their purchase on another device.
• Store landing-page click IDs on order rows. Capturing gclid and fbclid parameters at landing and writing them directly to your order records creates a permanent attribution key that privacy updates cannot erase.

Frequently asked questions

Can I turn off modeled conversions in my ad account?

No. There is no switch in standard Google Ads or Meta Ads reports to display only physically observed conversions. The platforms integrate modeled data directly into their reporting columns. To view unmodeled conversions, build first-party attribution tracking in your own warehouse.

Are modeled conversions real orders?

In aggregate, yes. They represent real transactions that took place, but which the ad platform could not directly trace to an individual ad click. However, modeling cannot prove incrementality — whether the ad caused the purchase, or whether the customer would have bought anyway.

How do modeled conversions affect view-through attribution?

View-through conversions are already weak attribution signals because they credit ad impressions rather than clicks. Modeling compounds this by estimating unobserved impressions. Always separate click-based conversions from view-through conversions in your weekly reports.

Should we stop using ad platform conversion counts?

No. Ad platform numbers remain the best signal for intraday campaign optimisation, ad set scaling, and creative testing. Just do not use them to calculate executive unit economics or board-level ROI.