Skip to content
whitead

Modeled conversions

glossary google ads updated 2026.05.20 8 min read
// contents · 9 sections
  1. What modeled conversions are
  2. How the model works
  3. What changed in 2025-26
  4. Common misconceptions
  5. Can you opt out of modeled conversions?
  6. Are modeled conversions the same as GA4 modeled rows?
  7. Does Smart Bidding optimise on modeled conversions?
  8. Related concepts
  9. Sources

Modeled conversions are the statistical inference layer Google Ads and GA4 use to fill in conversions they cannot directly observe — iOS ATT opt-outs, consent-denied EEA traffic, and cross-device journeys where the deterministic identifier is missing. By 2025-26 they are the default measurement substrate underneath Consent Mode v2 Advanced, Enhanced Conversions, and GA4 reporting (Google Ads Help, 2025-09-18).

What modeled conversions are

A modeled conversion is a count (and value) inferred by Google's machine-learning systems for a conversion event that occurred but could not be tied to an ad click via deterministic signals — typically because the user denied consent, opted out of ATT, used a different device, or used a privacy-preserving browser that suppressed the click identifier. The model uses observed, fully-attributed conversions as training data and extrapolates to similar unattributed traffic (Google Ads Help, 2025-09-18). Modeled rows appear in the same Conversions column as observed conversions — there is no separate "modeled" column in standard Google Ads reporting, though GA4 surfaces a "Conversion modeling" row in the Attribution path explorer (Google Analytics Help, 2025-08-12).

Three flavours run in production today:

  • Consent Mode modeling — fills the gap left by EEA users who denied analytics or ads consent under Consent Mode v2 Advanced.
  • Cross-device modeling — attributes a conversion on Device B to an ad click on Device A when the user was not signed in to a Google account on both.
  • App campaign / iOS ATT modeling — recovers SKAdNetwork-only signal for iOS app installs and in-app events.

How the model works

Google trains a per-advertiser behavioural model using your observed converters (users who completed the conversion AND whose click was deterministically attributed) as positive examples. The model learns which click, query, ad, device, and time-of-day patterns precede conversion in your account, then scores anonymous unattributed traffic against those patterns. When the aggregated score on a traffic slice crosses a threshold, the system credits a modeled conversion to that slice (Google Ads Help, 2025-09-18).

Three preconditions gate the model:

  • Identity infrastructure — Consent Mode v2 implemented (granted + denied pings both flowing), Enhanced Conversions enabled, or GCLID/GBRAID/WBRAID captured for App.
  • Volume — Google publishes a threshold of at least 700 ad clicks over a 7-day period on the Search Network, per country and domain grouping, before Consent Mode modeling activates; sub-threshold accounts see zero modeled rows rather than wrong rows. Modeling thresholds for Display/YouTube and Enhanced Conversions modeling differ (Google Ads Help, "About consent mode modeling"). Independent operators including Simo Ahava have echoed that consent-denied modeling quality is dominated by whether the account clears that floor with a measurable consent-denied fraction (Simo Ahava, 2024-03).
  • Stability — frequent bid-strategy changes, conversion-action re-additions, or learning-phase re-entries degrade the training set and inflate volatility.

Modeled values are typically released with a 24-72 hour lag and may restate as more training data arrives.

What changed in 2025-26

Three shifts moved modeled conversions from "nice to have" to load-bearing in the past year:

  • EEA Consent Mode v2 enforcement — since March 2024 advertisers serving EEA traffic must run Consent Mode v2 to use Google audience and remarketing products; by 2025 modeling is the only way the Conversions column captures EEA users who denied consent (Google Ads Help, 2025-04-22).
  • GA4 modeled rows in Attribution reports — GA4 surfaced explicit "Conversion modeling" lines in path-exploration and acquisition reports during 2025, making the modeled share visible per channel (Google Analytics Help, 2025-08-12).
  • Smart Bidding optimises on modeled signal — Google confirmed Smart Bidding treats modeled conversions as training input alongside observed conversions, meaning bid behaviour reflects the model's confidence on consent-denied traffic (Google Ads Help, 2025-09-18).

Google's own framing emphasises rigorous accuracy: the company reports that conversion modeling through Consent Mode recovers more than 70% of ad-click-to-conversion journeys lost to user consent choices, with results varying widely by account (Google blog, 2021-04). Independent measurement is more cautious. Simo Ahava's Consent Mode v2 instrumentation guide notes that modeling quality depends heavily on which mode is configured (Basic vs Advanced) and on whether ad_user_data / ad_personalization signals are actually flowing — implementations that look correct in the UI can still starve the model (Simo Ahava, 2024-03). Improvado's 2026 governance write-up framed the operator-facing pain more bluntly: stacking GA4's model, Google Ads' separate model, and Consent Mode fill-in on the same raw event produces three different answers that cannot be reconciled inside the platform UIs without a warehouse-side governance layer (Improvado, 2026-05). The Google Ads Help threshold page makes the same point structurally — below 700 ad clicks per 7 days per country/domain, the model defaults to zero modeled rows rather than wrong rows, which is safer but invisible to operators who assume modeling is always on (Google Ads Help).

The honest read: Google's accuracy claims are defensible at the aggregate level for advertisers above the volume threshold. Below the threshold, or during periods of platform instability, modeled conversions are best treated as a directional signal that requires triangulation against server-side measurement before it drives a bid decision.

Methodology note. Whitead's audit engine evaluates modeled-conversion reliability on three signals rather than asserting an account-wide accuracy number. First, we check whether Consent Mode v2 is implemented in Advanced mode (both granted and denied pings observed in the last 7 days). Second, we measure trailing 7-day ad clicks per country/domain against Google's published 700-click threshold — accounts below the floor are flagged as "modeling not activated" regardless of how the Conversions column looks. Third, we cross-reference the Google Ads Conversions column against server-side or CRM-confirmed conversions for the same window, flagging deviations greater than 20% for review. The rule does not fire on modeling itself — it fires on the combination of consent-denied traffic, sub-threshold volume, and the absence of a server-side or offline ground-truth signal to validate against. We do not assert frequency claims pending sample-size disclosure in a later wave.

Common misconceptions

Modeled conversions look like regular conversions in the UI, which invites two related mistakes: trusting them at any volume, and assuming you can turn them off.

Can you opt out of modeled conversions?

Not granularly. Modeling is on by default for any conversion action that meets Google's preconditions and you cannot toggle "model this conversion" per action. What you can do: switch Consent Mode from Advanced to Basic (no consent-denied pings sent → no Consent Mode modeling), disable Enhanced Conversions (removes one input signal), or report the Conversions column net of modeled rows by exporting both Google Ads native conversions and a server-side ground-truth count and comparing. Most agencies that ask "how do I turn it off" actually want the third — a parallel ground-truth pull — not a true opt-out (Google Ads Help, 2025-04-22).

Are modeled conversions the same as GA4 modeled rows?

No, though they share methodology. GA4's "Conversion modeling" row models behavioural data for the GA4 property (cross-device journeys, consent-denied sessions) and feeds GA4 reports. Google Ads modeled conversions are computed separately by the Ads measurement stack and feed the Google Ads Conversions column. The two systems can disagree, and they frequently do — see the GA4 vs Google Ads conversion mismatch playbook for the reconciliation workflow (Google Analytics Help, 2025-08-12).

Does Smart Bidding optimise on modeled conversions?

Yes. Google confirmed in 2025 that Smart Bidding treats modeled conversions as training input alongside observed conversions; the strategy does not know which is which when it sets a bid (Google Ads Help, 2025-09-18). The practical consequence: if your account is below the modeling volume threshold, Smart Bidding is optimising on a sparser, more volatile signal than the Conversions column implies. Triangulate with server-side measurement before assuming bid behaviour is stable.

Modeled conversions are downstream of Consent Mode v2 (which generates the denied-consent pings the model needs), Enhanced Conversions (which supplies first-party identifiers as model input), and GCLID / GBRAID / WBRAID capture (which determines what counts as deterministically observed). They feed Smart Bidding and Data-Driven Attribution as training input. When modeled rows look wrong, run the GA4 vs Google Ads conversion mismatch and server-side GTM for Google Ads playbooks before re-tuning the bid strategy — the symptom is usually instrumentation, not the model.

Sources

// was this useful?
// anonymous · no personal data stored
← Back to Wiki