Conversion Modeling: How Google Estimates the Conversions Consent Is Costing You
7 min read · AstraLoop Studio
You launch a campaign on Google Ads. The cookie banner pops up, and a sizeable share of your visitors (in Italy, often 20% to over 40%) deny tracking consent. From that moment on, Google can no longer observe what those people do: whether they buy, fill out a form, or call you. The result in your dashboard is fewer conversions than actually happened, and the real risk is cutting budget from campaigns that are, in fact, working.
Conversion modeling is Google's answer to that gap. These are estimates produced by a machine learning model to reconstruct conversions that genuinely happened but that the system couldn't measure directly. They aren't invented numbers, but they aren't 100% observed data either. Understanding the difference, and knowing when to trust them, is now part of the job for anyone managing ad budget.
What modeled conversions are
An observed conversion is one Google sees from start to finish: ad click, identifier set, purchase or lead tracked, all tied back to the same user. A modeled conversion, on the other hand, is a conversion that, in all likelihood, happened, but is missing one of those observable steps. The model estimates it and attributes it to the correct campaign, keyword, or ad.
Google fills the gap statistically: it takes the subset of traffic it can observe in full, measures its actual behavior, and applies those same rates to the portion of traffic it can't see. Modeled conversions don't land in a separate report — they flow into the standard "Conversions" column and feed everything downstream, including the Smart Bidding that sets your bids.
A concrete example: out of 100 real sales generated by your ads, Google might directly observe 70, while the other 30 stay invisible because the user denied consent or switched devices. Without modeling, you'd see 70 conversions and conclude the campaign performs worse than it actually does. With modeling, Google estimates those missing 30 and puts them back where they belong.
Why the gap exists
"Invisible" conversions don't come only from denied consent. There are four main causes:
- Denied consent. Under GDPR and Consent Mode, if the user refuses cookies, Google can't set identifiers or read the full journey.
- Browsers without third-party cookies. Safari (with ITP) and Firefox have restricted cookies for years, breaking attribution even for users who do give consent.
- Cross-device journeys. The user clicks the ad on their phone and buys from their computer: two devices, one purchase that's hard to stitch together.
- Handoffs between networks and apps. The path between apps, in-app browsers, and the website fragments signals further.
The practical consequence is that a share of the sales generated by your ads no longer reaches the report directly. Which is exactly why tracking conversions properly on Google Ads isn't enough on its own anymore: you also need to understand how the system reconstructs what it doesn't see.
How the model works, behind the scenes
Conversion modeling follows a precise logic, worth knowing so you don't treat it as a black box:
- Isolate observable traffic. Google starts from users who gave consent and whose journey is fully visible.
- Learn the patterns. It calculates conversion rates by geography, device, time of day, campaign type, and other signals.
- Quantify the relationship between consenting and non-consenting users — how much the behavior of one predicts the behavior of the other.
- Apply and re-attribute. It estimates the conversions from unobservable traffic and assigns them to the campaigns that generated them.
This is where Consent Mode v2 comes in. In advanced mode, even when a user denies consent, Google's tags send anonymous, cookieless "pings": they contain no personal data, but they signal that a relevant event occurred. These pings are exactly what makes the model far more accurate, because they feed it signals specific to your account instead of a generic model.
One caveat: the model doesn't kick in at just any volume. Google requires minimum thresholds — for example (per official documentation, subject to revision), roughly 700 ad clicks in 7 days per country/domain combination for Consent Mode modeling. Thresholds for GA4's behavioral modeling are higher (on the order of several thousand events and users per day). Below those numbers, modeling may not start, or it may switch off — and it's an ongoing requirement: if traffic drops, so does the modeling.
When to trust modeled data (and when not to)
The golden rule: modeled conversions are reliable in aggregate, much less so in detail. A statistical model is precise at scale and shaky when you probe it on a single case. Here's how to calibrate:
| Trust it more when | Trust it less when |
|---|---|
| You're looking at aggregated account or campaign data | You drill down to a single keyword or a single ad |
| The window is wide (30 days or more) | You're analyzing a single day or a few hours |
| Conversion volume is high and steady | Volume is low or highly seasonal |
| The setup has been stable for weeks | You've just changed tags, consent, or tracking |
| You're comparing the trend over time | You're comparing the absolute number against another platform |
The most common mistake is comparing Google Ads' conversion count (which includes modeled ones) with the raw number from another source and crying foul over "inflated" data. These are two different measures with different boundaries. The value of modeled conversions isn't a perfectly precise absolute number — it's giving Smart Bidding enough signal to optimize instead of flying blind.
How to feed the model well
Here's the part almost no one tells you: the quality of the model depends on the quality of the data you give it. "Garbage in, garbage out" applies to Google's machine learning too. Here are the concrete levers, in order of impact:
- Advanced Consent Mode, not basic. Basic mode blocks tags until consent is given and relies on a generic model — lower accuracy. Advanced mode sends cookieless pings and also unlocks GA4's behavioral modeling.
- Enhanced Conversions. Enhanced Conversions send Google encrypted first-party data (like the customer's hashed email) to reconnect the click and the conversion even without cookies.
- Server-side tracking. Server-side tracking collects signals more reliably and more resistant to browser blocks.
- First-party and CRM data. Building a solid base of first-party data and importing offline conversions from your CRM tells Google what really happened after the click, even days later.
- Volume and stability. Consolidating tracking into a single Google tag, avoiding double counting and dirty events, and keeping the setup stable helps the model "learn" instead of starting from scratch every time.
Notice the common thread: every lever is a way to give the model more clean signal. You're not gaming the system — you're putting it in a position to estimate well.
Want to know if your tracking is really feeding Google's model well? Request a setup analysis: we'll check your consent, Enhanced Conversions, and CRM data flow.
Privacy, machine learning, and data quality: the real connection
It's worth connecting the dots, because this is where the topic turns strategic rather than merely technical. The sequence goes like this:
- Privacy (GDPR, consent, cookieless browsers) creates a structural, permanent gap in the data: the full tracking of ten years ago is never coming back.
- Google's machine learning fills that gap with estimates, and by now it's the only way to get a complete picture.
- But machine learning is only as good as the signals it receives. And in a cookieless world, the best signals are the data you own: first contacts, purchases, outcomes logged in your systems.
Put differently: in a world where Google observes less and less, your competitive edge shifts to the data you collect and organize in-house. Whoever has a well-kept CRM, clean events, and a flow that feeds the model real conversions (including offline ones) gets better estimates, sharper Smart Bidding, and — for the same spend — more customers. Data quality isn't a technical footnote; it's what separates those who optimize from those who guess.
Mistakes to avoid
- Sticking with basic Consent Mode thinking "something gets modeled anyway": you leave accuracy and GA4's behavioral modeling on the table.
- Reading modeled conversions at the daily or keyword level and making drastic decisions based on plain statistical noise.
- Not connecting your CRM: without offline conversions, the model ignores everything that happens after the form.
- Double counting and duplicate events that dirty the signal and worsen every downstream estimate.
In summary
Modeled conversions aren't a trick, nor a problem to eliminate: they're the new normal in an advertising ecosystem where privacy has shrunk what's observable. Your job isn't to distrust the model — it's to feed it well and read it with judgment: trust the aggregate, be wary of the detail, and invest in first-party data quality. If you want the full picture of how all this fits together, start with our complete guide to conversion tracking.
Frequently asked questions
Can you trust modeled conversions?
Yes, but at the aggregate level. On accounts and campaigns with high volumes and wide windows, the estimates are solid; at the single-day or single-keyword level, read them cautiously, since statistical noise increases.
What's the difference between an observed conversion and a modeled conversion?
An observed conversion is measured directly by Google from start to finish of the journey. A modeled one is estimated by a machine learning model when an observable step is missing, for example due to denied consent or a device switch.
How do you activate conversion modeling on Google Ads?
By correctly implementing Consent Mode (ideally in advanced mode) and reaching the minimum traffic thresholds. Google cites, for example, roughly 700 clicks in 7 days per country/domain combination — figures subject to revision and that need to be maintained over time.
Do modeled conversions inflate the numbers?
No, they don't inflate them: they recover real conversions that consent and browsers hide. The appearance of inflation comes from wrongly comparing Google Ads' figure, which includes modeled conversions, with the raw figure from another platform.
Do I need advanced Consent Mode, or is basic enough?
Both allow a minimum of modeling, but advanced mode sends cookieless pings that make estimates more accurate and unlock GA4's behavioral modeling. For most businesses it's the better choice, weighed carefully against privacy considerations.
Do modeled conversions affect Smart Bidding?
Yes. They flow into the Conversions column and become the signal Smart Bidding optimizes bids against. Feeding the model well therefore also means making automated bidding strategies work better.
Fewer lost conversions, more reliable estimates, budget invested better. Talk to us and let's build a data infrastructure together that feeds the model with clean signals.