Definition
A platform can count a conversion directly only when it receives the conversion event and can connect it to an ad interaction. Anything that breaks that connection leaves a gap: the conversion happened, but the platform cannot attribute it. Modelled conversions are the platform's estimate of the gap, calculated from the behaviour of comparable users whose journeys it did observe, and reported as part of its conversion figures.
The term covers estimates rather than counts. A modelled conversion is the output of a statistical inference, not a recorded event with a time, an order number and a customer behind it. Google and Meta both document that they model conversions in these situations and describe the conditions under which they do so; neither publishes the model itself.
Where the signals come from
The gaps that modelling fills arise in a few recurring places.
- Consent choices. Where a visitor declines advertising storage, Google tags in consent mode's advanced implementation send limited, cookieless signals instead of full events. Those signals, combined with data from consenting visitors, are what Google's conversion modelling works from. The consent mode guide sets out what the basic and advanced implementations give the models and what eligibility requires.
- Signed-in and cross-device journeys. A person who clicks an ad on a phone and buys on a laptop can be connected where the platform sees them signed in, and estimated where it does not.
- Browser and app privacy restrictions. Limits on identifiers and tracking remove matches that used to be observable, and platforms describe modelling and aggregation as their response.
The guide to privacy, consent and measurement explains what each consent state removes from observation and why a market where almost everyone declines gives the models little to work from.
In practice: reading observed and modelled together
Modelled conversions usually appear in the ordinary conversion columns, distributed across campaigns under the account's attribution settings, rather than as a separate line. That has practical consequences for reporting.
- Mark changes in the implementation. A switch in consent setup changes how much is modelled, and so changes reported conversions, without any change in what customers did. The consent banner QA checklist includes briefing ad platform owners on the modelled conversions and reporting gaps to expect before a new banner goes live.
- Label figures that include modelling. The guide to building a measurement plan treats platform conversions, with their view-through and modelled components, as a different source from analytics and warns against presenting either as revenue.
- Read trends, not small slices. Modelled figures are reasonably stable in aggregate and unreliable when cut by many dimensions at once.
- Allow recent days to settle. Modelled estimates can arrive after observed data, so the latest days rise as the estimates land.
Why totals differ from analytics and orders
Modelled conversions are one of the main reasons ad platforms report more conversions than an analytics tool or the order system attributes to them. Analytics may model some behaviour where eligible and leave other reports blank; the ad platform models its own attributed conversions; the order system counts orders with no attribution at all. The comparison of platform attribution and analytics attribution lists unobserved conversions as one of the systematic differences between the two, alongside windows and view-through credit.
The right question is therefore not which figure is correct but whether the trends agree. If modelled platform totals rise while orders are flat, the model is not evidence of growth. For online shops, the e-commerce industry marketing guide notes that consent leaves some sales unobserved by the ad platforms, which therefore report a mix of observed and modelled conversions that the store's own order data has to anchor.
A surf school reads two lines instead of one
Swellhaven Surf School added a consent banner in May. Its analytics recorded 212 lesson bookings from paid search in April and 148 in June, while Google Ads reported 230 conversions in April and 221 in June. The booking system showed 290 paid-search-tagged bookings in April and 284 in June, on the same budget. The owner first read the analytics fall as a collapse in performance. The gap between the two platforms was mostly modelled conversions for visitors who declined advertising consent. The monthly report now shows three lines: analytics bookings, Google Ads conversions with a note that they include modelling, and bookings in the booking system, with May marked as the banner launch. Budget decisions follow the booking-system trend.
Related terms
Consent ModeA Google tag feature that adjusts what tags collect according to the visitor's consent choices. relays consent choices to Google tags and, in its advanced implementation, supplies the signals conversion modelling uses. A View-Through ConversionA conversion credited to an ad that was shown but not clicked, within a set view window. is another kind of platform credit that analytics never sees, and the two are often confused in reconciliations. Enhanced ConversionsA Google Ads feature that sends hashed first-party customer data to improve conversion matching. recover matches for consented conversions using hashed first-party data, improving what is observed rather than estimating what is not. AttributionAssigning conversion credit to touchpoints. rules distribute modelled conversions across campaigns in the same way as observed ones.
Common confusions
- Modelled conversions are tracked conversions. They are estimates and should be labelled wherever they appear.
- Modelling recovers individual journeys. It estimates totals; it does not reveal who converted.
- Every account gets modelling. Eligibility conditions apply, and smaller accounts or partial implementations may see none.
- Modelling and enhanced conversions do the same job. One estimates the unmeasured part, the other improves matching in the measured part.
- A higher modelled total means better performance. Without movement in orders, it means more estimation.
