Modelled conversions in Google Ads - what they are and why they matter

When a user takes an action on your website that should register as a conversion - completing a form, making a purchase - that event may or may not be measurable depending on the user's privacy settings, browser type, consent choices, and whether the session crosses multiple devices. Google's response to these gaps is to use machine learning to model what conversions it cannot directly observe, based on patterns from the conversions it can. The result is a conversion count that is partly observed and partly estimated.

How Google models conversions

Google uses aggregated, anonymous patterns from users who do consent to tracking to estimate the conversion behaviour of similar users who do not. If users with a particular set of behavioural characteristics - device type, session depth, engagement pattern - convert at a certain rate when observable, Google applies a similar estimated rate to unobservable users with matching characteristics. The modelled conversions are then added to your observed conversions to produce the total conversion figure in your reports.

What this means for your data

Your conversion numbers in Google Ads are not a direct count of people who completed an action. They are a blend of directly observed conversions and statistical estimates. For most accounts the ratio of modelled to observed conversions is relatively small, but in sectors with high cookie refusal rates - news, healthcare, finance, legal services - the modelled proportion can be significant. This does not mean the numbers are wrong, but it does mean they carry inherent uncertainty that pure observed data does not.

Smart bidding and modelled data

Smart bidding strategies use your conversion data - including modelled conversions - as their optimisation signal. Google argues that including modelled conversions improves bidding accuracy because it gives the algorithm a more complete picture of conversion behaviour than observable data alone provides. The counterargument is that relying on modelled data introduces estimation error that can compound within automated bidding systems. Both positions have merit. The practical implication is that unusually volatile conversion data - spikes or drops that do not correlate with any known campaign or business change - may reflect changes in the modelling rather than actual changes in conversion behaviour.

What to do

Implement Enhanced Conversions to reduce the proportion of modelled data in your account. Enhanced Conversions use hashed first-party data from form submissions to directly match conversions to Google accounts, reducing the dependency on modelling for those specific conversion paths. The more first-party data you pass back to Google, the more of your conversions are observed rather than estimated, and the more reliable your bidding signal becomes.