Modeled conversions are purchases, sign-ups and other actions that an advertising platform estimates rather than counts. They cover buyers a tag could not see: someone who rejected cookies, a Safari visitor whose click identifier expired, a person who clicked on a phone and paid on a laptop. The platform studies the conversions it did observe and adds an estimate of the ones it missed to its reported totals. Google, Meta and others use the technique because privacy rules and browser limits now remove a measurable share of conversions from view, and automated bidding systems need a complete count to set prices.

Why some conversions cannot be counted

An observed conversion depends on a chain of identifiers. A click on a Google ad appends a click identifier, the GCLID, to the landing page address. A tag stores it in a first-party cookie, and when a purchase happens the tag "reads this cookie and sends it back to Google Ads with the conversion information", according to Google's help pages. If any link breaks, the sale still happens but the platform cannot connect it to the ad.

Three breaks dominate. Consent: under the ePrivacy Directive and the General Data Protection Regulation (GDPR), a European visitor who declines the cookie banner leaves no identifier behind. Browsers: Apple's Intelligent Tracking Prevention (ITP), introduced in Safari in 2017, shortened the life of script-set cookies and later began stripping click identifiers from some URLs. Devices and apps: since App Tracking Transparency (ATT) took effect with iOS 14.5 in April 2021, most iPhone users have not shared the advertising identifier that once tied an app ad to a later action.

Modeling is one of two repairs. The other, matching hashed customer data to signed-in accounts, produces observed conversions; modeling produces estimates.

How the models work

The system starts with conversions it can observe and attribute, then uses aggregate patterns from comparable observed traffic to predict how many unlinked clicks converted. Google's help page on modeled conversions lists the gaps it covers: third-party cookie limits in Safari and Firefox, first-party cookie limits, users in the EU who decline consent, iOS traffic under ATT, the removal of the Google Play advertising ID for App campaigns, and conversions that cross devices.

Consent mode supplies the inputs for the consent case. Launched by Google in September 2020, it lets tags load before a visitor decides and then change behaviour based on the answer. When ad_storage is denied, tags can send cookieless pings: hits that carry no cookie but record that a page view or conversion occurred. Google says the pings are "never used to track individual users across apps or websites, build remarketing lists, or generate user profiles".

The implementation choice matters. In advanced consent mode, tags load before the banner and send pings for refusers. In basic consent mode, tags are blocked until consent is given, so no pings arrive, and Google recommends pings "for more accurate modeling". Eligibility thresholds also apply. Google's consent mode modeling page, which dates the feature from April 2021, requires "700 ad clicks over a 7 day period, per country and domain grouping" before models enter a training period. Google Analytics 4 (GA4) requires at least 1,000 daily events from non-consenting users over seven days, and 1,000 daily consenting users on at least seven of the previous 28 days, for its behavioural modeling to run.

The output is aggregate: a fractional addition to a campaign's count on a given day, not a reconstructed record of a person. Google says it includes modeled conversions in reporting only "when we're highly confident" in them and uses "techniques like holdback validation", withholding observed conversions and checking whether the model predicts them.

Where modeled numbers appear

In Google Ads, modeled and observed conversions share one column. "In the 'Conversions' column, Google reports both modeled and observed conversions," according to Google's help page. The same blended totals are tied to optimisation and bidding, so an estimated sale influences Smart Bidding in the same way as a counted one. Modeled conversions "can take up to 5 days to fully process and stabilize", according to Google, which makes the most recent days look weaker until estimates arrive.

Google's enterprise tools follow the same logic for Floodlight, the conversion tag used by Display & Video 360 (DV360) and Campaign Manager 360. A DV360 release note told buyers that "starting the week of September 28, all attribution models will be opted into conversion modeling by default", a change Google said would produce "a low double-digit increase in conversions for most advertisers". The undated note matches September 2020. On iOS, Google has applied modeling to App campaigns whose journeys end in mobile web browsers, where ATT removes the identifier.

Meta reached the same place by another route. Ahead of ATT, Meta told developers on December 16, 2020 that "statistical modeling will be used for certain attribution windows and/or metrics to account for less data availability from iOS 14 users", and that "in-product annotation will communicate when a metric is modeled". It paired the change with Aggregated Event Measurement for web events from iOS devices.

Origin and evolution

Estimating unseen conversions predates privacy regulation. On October 1, 2013, Google AdWords introduced Estimated Total Conversions, which added modeled cross-device conversions to observed ones. Cross-device estimates were later folded into the standard Conversions column, with a rollout that began on September 6, 2016, according to Google. Accounts of the origin differ: PPC Land's 2020 primer says modeled conversions "were first created by Google in 2017", the year Safari's tracking limits began.

Privacy limits then turned modeling from extra insight into gap-filling. On April 15, 2021, Henrique de Freitas, a product manager at Google, announced conversion modeling through consent mode in Google Ads, writing that "on average, conversion modeling through Consent Mode recovers more than 70% of ad-click-to-conversion journeys lost due to user cookie consent choices". The same post added that "results for each advertiser may vary widely". Google has not published the method behind the average.

Regulation made the inputs mandatory. Consent mode version 2, announced in November 2023 with the Digital Markets Act (DMA) in view, became a condition for personalised advertising on European Economic Area (EEA) and UK traffic from March 2024. From July 21, 2025, Google began disabling personalisation, remarketing and conversion tracking for accounts sending non-compliant European traffic. On June 15, 2026, consent mode's ad storage signal became the only control over advertising data flowing from GA4 to Google Ads, replacing the Google Signals setting.

Why it matters for advertisers

Bidding is the main consumer of modeled conversions: Smart Bidding and Meta's delivery system learn from conversion volume, and a campaign whose conversions vanish from the data receives less spend.

Configuration errors show the stakes. Consultant Mike Teasdale described a client whose Google Ads conversions fell by 90% overnight after consent signals stopped reaching its tags; once the setup was fixed, roughly 40% of the attribution data came back and the remaining 60% was lost. Google's data strength uplift metric, introduced on September 10, 2026, estimates conversions recovered by first-party data over 14 days, counts modeled conversions from enhanced conversions for leads, and is shown only at 1% or more.

Limitations and disputes

The central dispute is verification. The platform that sells the media also estimates the results, and the estimates cannot be audited from outside. Google publishes no error rates per account and does not separate modeled from observed conversions in the main column. Experiments are the usual check. Citing Facebook's published field experiments, a measurement product lead at Genentech who previously worked at Amazon and Walmart argued in September 2026 that standard observational methods miss the true effect "frequently by a factor of 3 or more", though the underlying papers give no single multiplier.

A second dispute is statistical. Models assume that people who refuse cookies behave like similar people who accept them. If refusers are systematically more or less likely to buy, the estimates carry that bias into bids. Alliance Digitale, the French digital advertising trade body, warned in January 2026 that attribution grows harder as signals fragment.

A third is legal. Cookieless pings still send data to Google about visits from users who declined consent. Whether that transmission itself needs consent under the ePrivacy Directive is argued among consent vendors and lawyers; no European regulator has ruled specifically on consent mode pings.

Thresholds add a practical limit. Low-volume accounts, offline conversion imports and properties below the bar may get no modeling, so the technique favours large advertisers.

Observed conversions are recorded events tied to an ad interaction through an identifier. Modeled conversions are estimates added beside them.

Enhanced conversions recover conversions by matching hashed first-party data, such as email addresses, to signed-in accounts. The output is observed, not modeled, as PPC Land's enhanced conversions explainer sets out.

Data-driven attribution shares credit for a conversion among ad touchpoints. It does not decide whether the conversion happened.

Conversion lift measures how many conversions an ad caused, by comparing exposed and held-out groups. Modeled conversions estimate how many attributed conversions occurred, not how many were incremental.

Recent developments

Google continues to widen the inputs. On October 16, 2026 it said that from November 2026 Floodlight would use Google signed-in data signals to recover previously unattributed conversions in DV360, Search Ads 360 and Campaign Manager 360 reporting; Google did not say whether those conversions count as modeled or observed. Version 1.7 of its Data Manager API, released on May 28, 2026, extended offline conversions to the same three products. Browsers move the other way. Safari 27 blocks identity firms that turn email addresses into tracking IDs, removing observed data and leaving more to be estimated.

Timeline

  • October 1, 2013 - Google AdWords introduces Estimated Total Conversions, including modeled cross-device conversions.
  • September 6, 2016 - Google begins including cross-device conversions in the Conversions column.
  • 2017 - Apple introduces Intelligent Tracking Prevention in Safari.
  • September 2020 - Google launches consent mode.
  • Week of September 28, 2020 - DV360 opts all attribution models into conversion modeling by default.
  • December 16, 2020 - Meta tells partners it will use statistical modeling for some iOS 14 metrics.
  • April 26, 2021 - App Tracking Transparency takes effect with iOS 14.5.
  • April 15, 2021 - Google announces conversion modeling through consent mode in Google Ads, claiming more than 70% journey recovery on average.
  • 2022 - GA4 adds behavioural modeling for consent mode.
  • November 2023 - Google announces consent mode version 2.
  • March 2024 - Consent mode v2 becomes a requirement for personalised ads on EEA and UK traffic.
  • July 21, 2025 - Google begins disabling conversion tracking for non-compliant EU traffic.
  • June 15, 2026 - Consent mode becomes the sole control over GA4-to-Google Ads advertising data.
  • September 10, 2026 - Google introduces the data strength uplift metric in Google Ads.
  • November 2026 - Floodlight begins using Google signed-in data to recover unattributed conversions.

Summary

Who: Ad platforms, chiefly Google (Google Ads, GA4, DV360, Campaign Manager 360) and Meta, build and report modeled conversions; advertisers consume them through reports and automated bidding; regulators, browser makers and Apple shape the gaps they fill.

What: Statistical estimates of conversions that happened but could not be linked to an ad, added to observed conversions in platform totals.

When: Google estimated cross-device conversions from October 2013; modeling became central after ITP in 2017, consent mode in 2020, ATT in 2021 and consent mode v2 enforcement from March 2024.

Where: Inside ad platform reporting, most heavily on European traffic where consent refusals are common and on Safari and iOS where identifiers are restricted.

Why: Because privacy rules and browser limits hide a share of conversions, and bidding algorithms need a full count; the cost is that the estimates are produced by the seller of the media and cannot be independently audited.