Conversion modeling uses Google's machine learning to estimate conversions it cannot directly observe. It analyzes conversions from users who consented to cookies, then applies those patterns to visitors who declined, using anonymous cookieless pings as the input signal. The result is an estimated count added to the Conversions column, alongside real, observed data.
Below: the 4-step modeling process, how Consent Mode feeds it, and how accurate the estimate actually is.
How Does Conversion Modeling Work? (The Short Answer)
Conversion modeling uses Google AI to analyze how consented users behave, then estimates the conversions from unconsented users that cookies could not record. Google Ads Help describes the mechanism this way.
Conversion modeling uses observed conversions to predict unobserved conversions without identifying any one individual.
The model exists because two forces strip cookies and identifiers out of the data. Browsers are deprecating third-party cookies, and visitors are declining consent under GDPR-style opt-in rules. Without modeling, both losses would show up as a drop in reported conversions, even though the underlying sales or signups still happened.
Modeled data does not just patch a report. It also feeds Google Consent Mode, which in turn feeds Smart Bidding. Google's automated bid algorithms use the blended observed-plus-modeled total to keep optimizing toward a realistic conversion count, instead of an artificially low one.
How the Conversion Model Works, Step by Step
The model turns known behavior into an estimate for unknown behavior in four stages. Each stage takes the output of the one before it and narrows the gap between observed and total conversions.
Step 1: Separating Observed and Unobserved Conversions
Every ad interaction starts in one of two pools. Observed conversions carry a cookie or identifier that directly links an ad click to a conversion. Unobserved conversions have no such link, usually because the visitor declined cookies, blocked third-party tracking, or converted on a different device.
This split happens automatically as data arrives. A visitor who accepts the cookie banner and later checks out lands in the observed pool. A visitor who rejects the banner and checks out anyway lands in the unobserved pool, with no direct trail connecting the two events.
Step 2: Grouping Observed Conversions by Shared Traits
Google divides the observed pool into subgroups based on shared characteristics. The standard grouping variables are location, time of day, device type, and browser. The model then calculates key metrics, such as conversion rate, for each subgroup.
A subgroup might be "mobile visitors from Germany converting between 6 and 9 PM." Say that subgroup shows a 4% conversion rate among observed users. That rate becomes the reference point for step 4.
Step 3: Matching Unobserved Interactions to Those Groups
Unobserved ad interactions get sorted into the same subgroup framework used in step 2. The sort relies on whatever signals are still available, such as country, device, and time of day. No cookie or identifier is needed here, because the sort uses aggregate traits, not individual tracking.
An unobserved click from a mobile visitor in Germany at 7 PM slots into the same subgroup as the observed example above. Google never linked that specific click to a specific outcome, yet the aggregate traits still place it.
Step 4: Predicting the Missing Conversions With Machine Learning
Google's machine learning applies each subgroup's known conversion rate to its unobserved interactions to estimate how many likely converted. The result populates the Conversions column as a modeled conversion, blended alongside the observed count.
Take the "mobile Germany evening" subgroup converting at 4% among observed users. If that subgroup has 10,000 unobserved clicks, the model estimates roughly 400 additional conversions from that segment alone. Google only reports the estimate when confidence is high enough; sparse subgroups produce no modeled figure at all.
Where Consent Mode Fits: The Signal That Makes Modeling Possible
Google Consent Mode is what makes modeling possible in the first place. When a visitor denies cookies, advanced Consent Mode still sends anonymous cookieless pings instead of going silent. Google then compares consented and unconsented behavior to estimate the gap. According to Google's consent mode modeling documentation, consented users are typically 2 to 5 times more likely to convert than unconsented users. That differential is exactly what the model has to account for.
Consent management decides what Google is allowed to collect in the first place. The consent choice a visitor makes therefore shapes which pool, observed or unobserved, their interaction lands in. The ad_storage signal specifically controls whether advertising cookies can be written at all.
The gap is largest in opt-in consent regions like the EEA. There, many visitors reject ad cookies by default, pushing a larger share of traffic into the unobserved pool that modeling has to fill.
Cookieless Pings: The Data Google Models From
A cookieless ping is a signal Google's tags send to Google's servers when a visitor denies cookie consent. It replaces the cookie-based measurement that would normally fire. Each ping carries two kinds of data.
- Functional information added passively by the browser: timestamp, user agent, and referrer.
- Aggregate, non-identifying information: whether the page URL contains ad-click information such as a GCLID or DCLID, the boolean consent state, and a random number generated on each page load.
Google draws a hard boundary here, stated in the same Google Ads documentation cited above.
Consent mode cookieless pings are never used to track individual users across apps or websites, build remarketing lists, or generate user profiles.
They exist to calibrate the model, not to identify anyone. This is the mechanism most competing explainers skip entirely. It is also the answer to the most common objection: modeling does not track the people who said no.
Why Advanced Consent Mode Recovers More Than Basic
Basic Consent Mode blocks Google tags from loading until a visitor interacts with the banner. If consent is denied, no data reaches Google at all, not even the consent status itself. According to Google's consent mode overview, that produces a "general model": less detailed, because it has no signal from denied-consent visitors to calibrate against.
Advanced Consent Mode loads tags immediately when the page opens and sends cookieless pings even when consent is denied. That extra signal produces an "advertiser-specific model," Google's own term for the more detailed version. The difference between basic and advanced Consent Mode decides how much data the model actually gets to work with.
Observed vs Modeled Conversions: What Is Actually Recovered
Observed and modeled conversions differ in how the link is made, not in whether the underlying conversion happened. The table below breaks down the mechanism side by side.
| Attribute | Observed conversions | Modeled conversions |
|---|---|---|
| How the link is made | Cookies and other identifiers connect the ad interaction to the conversion directly | Machine learning statistically links unobserved interactions to likely conversions |
| Data used | First-party and third-party identifiers with explicit consent | Aggregate subgroup patterns plus cookieless ping signals |
| Identifies individuals? | Yes, at the identifier level (consented) | No, aggregate estimation only |
| Where it appears | Conversions column, tagged as observed | Conversions column, blended with observed data |
| Confidence level | Direct measurement | Estimate, shown only above a confidence threshold |
Modeling only runs where there is enough data to support it. When a subgroup has too few observed conversions to calculate a reliable rate, Google does not model that segment at all. The unobserved interactions in it simply go unreported.
Where Conversion Modeling Runs: Google Ads vs GA4
The term "conversion modeling" actually covers two related Google systems that run on different thresholds. Google Ads conversion modeling recovers the attribution link between an ad click and a conversion event. Google Analytics 4 behavioral modeling, by contrast, estimates on-site behavior and channel attribution for visitors who declined analytics cookies. It feeds GA4's own reporting rather than Ads bidding directly.
| Attribute | Google Ads conversion modeling | GA4 behavioral modeling |
|---|---|---|
| What it estimates | Ad-click-to-conversion journeys lost to consent decline | On-site behavior and channel attribution for analytics-declined users |
| Minimum data needed | 700 ad clicks over 7 days, per country and domain | 1,000 denied events per day for 7 days, plus 1,000 granted daily users for 7 of the last 28 days |
| Where results show | Conversions column in Google Ads | Blended reporting identity in GA4 |
| Feeds | Smart Bidding | GA4 channel and conversion reports |
Both systems run independently and use their own eligibility thresholds. That independence is one reason Google Ads and GA4 conversion counts for the same site can diverge. Microsoft runs its own equivalent through Microsoft Consent Mode, and Meta models conversions from missing or partial data too. Each platform's modeling stays inside its own ecosystem. Pairing Consent Mode with server-side tagging can strengthen the raw signals every one of these systems receives.
What You Need for Conversion Modeling to Work
Conversion modeling only activates once a site clears a minimum data threshold; below it, Google shows observed data only, with no modeled figure added. The requirements differ by platform.
| Requirement | Google Ads | GA4 |
|---|---|---|
| Consent signal | Correct Consent Mode setup or IAB TCF v2.0 implementation | analytics_storage consent signal present on every page |
| Volume threshold | 700 ad clicks over 7 days, per country and domain grouping | 1,000 denied events/day for 7 days AND 1,000 granted daily users for 7 of the last 28 days |
| Tag timing | Tags must load and send cookieless pings even on denial (advanced mode) | Tags must load before the consent dialog appears, in all cases |
| Reporting view | Conversions column | Blended reporting identity |
The Ads threshold comes from the same Google Ads consent-mode-modeling documentation cited earlier; the GA4 threshold comes from GA4's behavioral modeling documentation. The tag-timing requirement trips up more implementations than the volume threshold does. Suppose Consent Mode is configured so tags only fire after a visitor grants consent. No cookieless pings ever reach Google from the denied-consent group, so the advertiser-specific model has nothing to calibrate against. Setting up Consent Mode correctly, so tags fire and send signals regardless of the visitor's choice, is the prerequisite everything else depends on. Configuring Consent Mode v2 so denied states still send cookieless pings is the specific setting that enables advanced modeling.
How Accurate Is Conversion Modeling?
Google validates its models with holdback testing. According to Google's own announcement, it reports recovering more than 70% of ad-click-to-conversion journeys lost to consent decline, though results vary widely by consent rate. Modeled figures are estimates, not observed counts, so a sudden large swing in modeled conversions is a signal to sanity-check, not celebrate.
Holdback validation works by applying the model's methodology to a subset of observed conversions and comparing the prediction against the real, known outcome. If the predictions consistently match the holdback subset, Google treats the wider unobserved estimate as trustworthy. If they diverge, the model does not ship a modeled figure for that segment.
Accuracy varies because the underlying consent rate varies. A site where most EEA visitors reject cookies has a large unobserved pool and more room for the model's estimate to drift from reality. A site with high consent rates has a smaller gap to fill and a tighter estimate. Advertisers watching PPC forums have reported modeled conversions jumping by unusually large margins after enabling advanced Consent Mode. Treat any sudden, outsized jump as a prompt to check the underlying consent rate and traffic mix. It is not a number to report at face value.
Conversion Modeling in Practice: A Worked Example
Consider an EEA-based online store running Google Ads with advanced Consent Mode enabled, where roughly 40% of visitors reject ad cookies at the banner.
The store clears Google's eligibility bar: it logs well over 700 ad clicks per 7-day window in its main market and domain. Every visitor who accepts cookies produces an observed conversion, cookie-linked back to the ad click that brought them in. Google groups those observed conversions into subgroups by device, region, and time of day, then calculates each subgroup's conversion rate.
The 40% who reject cookies still trigger a cookieless ping because advanced Consent Mode keeps tags loaded regardless of consent. Their ad-click interactions get sorted into the same device, region, and time-of-day subgroups as the observed data. Google's model then applies each subgroup's known conversion rate to that unobserved slice and adds an estimated conversion count to the Conversions column.
The store's Google Ads dashboard now shows a blended total: real, cookie-linked conversions plus a modeled estimate for the 40% who said no. Smart Bidding uses that fuller total to keep optimizing spend, instead of bidding as if 40% of the store's actual buyers did not exist.
FAQs
Is conversion modeling the same as enhanced conversions?
No. Enhanced conversions match a site's own hashed first-party data, such as an email address, to a logged-in Google user. Conversion modeling statistically estimates the conversions still missing after that matching happens. The two work together: enhanced conversions recover some observed data first, and modeling estimates whatever gap remains.
Does conversion modeling track users who declined cookies?
No. It uses aggregate, non-identifying cookieless pings and applies patterns learned from consented users. Google states cookieless pings are never used to identify individuals, build remarketing lists, or generate user profiles.
Why don't my Google Ads and GA4 conversion numbers match?
Google Ads and GA4 run separate modeling systems with different thresholds, attribution windows, and reporting dates. Conversion modeling is one documented reason the two totals diverge for the same site and date range.
How long does conversion modeling take to start?
It activates automatically once a site clears the eligibility threshold. For Google Ads, that is 700 ad clicks over 7 days per country and domain; for GA4, it is the 7-day behavioral-modeling windows. Below those thresholds, Google reports observed data only.
Can I turn conversion modeling off?
There is no manual per-account switch for eligible accounts. Google applies modeling automatically once the eligibility thresholds are met, and the modeled figures appear in the Conversions column alongside observed data.
Does conversion modeling work with Meta or Microsoft Ads?
Yes, but each runs its own separate system. Meta models conversions for missing or partial data and validates against ground truth. Microsoft runs its own equivalent through Microsoft Consent Mode. This article covers the Google Ads and GA4 mechanism specifically.
Do US-only advertisers get conversion modeling?
Modeling runs wherever the eligibility thresholds are met and consent signals exist, since thresholds are set per country and domain. Its largest impact is in the EEA and UK, where opt-in consent rejection is common. A US site with meaningful denied-consent traffic can qualify too.
Understanding the mechanism only helps if the consent signals feeding it are actually configured correctly. Consently ships Google Consent Mode v2 support enabled by default, managing the ad_storage, analytics_storage, ad_user_data, and ad_personalization signals that conversion modeling relies on.
