Conversion Tracking and Attribution

What the Conversions column actually contains

The number in the Conversions column of a Google Ads report is not a count of events observed to happen. It is a blend of conversions Google observed and conversions Google estimated, and Google's documentation states it plainly: in the Conversions column, Google reports both modeled and observed conversions.

That changes what the figure can carry. A count of observed events is a record of things that happened, subject to the ordinary questions about how they were recorded. A modeled conversion is an inference about something that was not observed, produced by a machine-learning process whose inputs the advertiser never sees and whose method is not published in auditable form. Both arrive in the same column, added together, and no control in the standard interface separates them.

So wherever a conversion count appears — a damages calculation, a missed cost-per-acquisition target, an allegation of inflated results — the figure is a blend and has to be described as one. Whether that matters to the claim is a question for counsel. What I can describe is what the number is, how it was produced, and which parts of it a person with account access chose.

The measurement chain, and the five places a person configured it

A conversion is the output of a chain, and every link in it is a configuration choice made by a person with account access:

1. A user clicks an ad; Google appends a Google Click ID (GCLID) to the landing page URL.
2. A tag fires on the site: the Google tag (gtag.js), or a Google Tag Manager container.
3. On a designated action, a conversion tag reports back against a conversion action the advertiser configured.
4. That conversion action carries its own settings: whether it counts every conversion or one, its conversion window, value, category, attribution model, and whether it is included in the Conversions column at all.
5. Google reports the result, plus modeled conversions where it could not observe the link.

Two of those settings do most of the damage. A conversion action set to count every conversion rather than one turns ten form fills from one person into ten conversions. A conversion action excluded from the Conversions column vanishes from the headline metric while still appearing under All conversions, so a single account reported two ways produces two totals, each correct to its own definition.

These are settings, not facts about the business. Where a report is offered as evidence of lead volume, the questions are what the conversion actions were set to do, and whether they held across the whole period compared.

Reported and measured are two different figures

The most common error I find in a production is not fabrication. It is a figure taken from one column and described as though it came from another. Four numbers get confused:

Google describes a modeled conversion as using data that does not identify individual users to estimate conversions it cannot observe directly, and describes the method as machine learning: observed conversions are divided into subgroups sharing characteristics such as time and browser, unobserved interactions are sorted into the same subgroups, and the known conversion rates of the observed subgroups link unobserved interactions to conversions where appropriate. That is an estimation procedure, described as one by the company that runs it. See Google Ads Help, About modeled online conversions (read August 14, 2026).

Where Google says it models, and where it says it does not

Google names five situations in which it models: journeys that cross devices; browsers that restrict or disallow cookies; consent-required regions where a user has not consented; App Tracking Transparency withholding a device identifier; and removal of the Google Play advertising ID. Modeling is not universal, and Google states that offline conversion imports and accounts without enough weekly conversions might not include modeled conversions at all.

A confidence threshold applies before anything is reported:

We only include modeled conversions in our reporting when we're highly confident that conversions actually occurred as a result of ad interactions.

That cuts both ways. Above the threshold the reported figure contains estimates; below it, real conversions are missing because Google declined to model them. Neither condition is visible in the standard interface. The published figures governing this behavior:

Below that click threshold an account gets no consent-mode modeling at all, so unconsented conversions are absent rather than estimated. Google also describes consented users as typically two to five times more likely to convert, so a report is not simply understated by the share of users who declined.

Why the Google Ads figure and the Analytics figure never reconcile

The two systems disagree by design, and Google publishes the reasons:

• Different credit. Google Ads uses the last Google Ads click; Analytics uses the last click across all channels.
• Different date. Google Ads reports a conversion against the date and time of the click; Analytics uses the date and time of the conversion. A month-by-month table from one will never tie to a month-by-month table from the other.
• Different units. Google Ads counts clicks; Analytics counts sessions. Several clicks inside one session count once in Analytics, and Google Ads filters invalid clicks Analytics may still record.
• Modeling. Google names modeled conversions as a likely reason Google Ads shows more conversions than Analytics.

Google also lists structural causes to rule out before treating a gap as meaningful: auto-tagging switched off, redirects stripping the parameter before the tag fires, a bookmarked URL carrying a stale GCLID, server delays, accounts unlinked mid-range, several ad accounts on one property, filters, and missing tags. See Google Analytics Help, Data discrepancies between Google Ads and Analytics (read August 14, 2026).

One further point helps both sides. No publisher with a stated sample and method has established a normal range for the size of that gap. The percentages in circulation come from vendor blogs with no sample, method or date. So a discrepancy claim needs a magnitude, a direction and a mechanism measured in the account itself, and an opposing expert calling a gap "within normal range" has no published benchmark either.

The attribution default changed, and older targets were not written against it

Four rules-based attribution models, first click, linear, time decay and position-based, are no longer supported, and conversion actions that used them were upgraded to data-driven attribution. Two models remain, last click and data-driven, with data-driven the default for most conversion actions. The retirement took effect in mid-October 2023, across both Google Ads and Analytics.

Data-driven attribution uses the account's own conversion data to calculate the contribution of each ad interaction across the conversion path, at the recommended volumes set out above. Below that volume the model still operates, with reduced accuracy.

This matters most where an agreement, a scope of work or an internal target was written when last click was the working assumption. After the switch, the reported numbers are not measuring what the target was written against. Google is on record, through trade coverage of its own statements, that fewer than 3% of conversions were using the retired models and that the switch typically produces about a 6% increase in reported conversions. A comparison spanning that date carries a measurement change as well. See Google Ads Help, About attribution models (read August 14, 2026).

Offline conversion imports, where the record starts on the advertiser's side

In a business with a long sales cycle, the conversions that matter close offline. The mechanism is the offline conversion import: the advertiser stores the GCLID with the lead and, when the deal closes, uploads it with the conversion details back into Google Ads. That data originates in the advertiser's own CRM, so its accuracy is a question about the client's or the agency's data handling rather than about Google's measurement.

• The upload deadline is hard. Offline conversions uploaded more than 90 days after the associated last click are not imported; for enhanced conversions for leads the window is 63 days. A deal that closed later is structurally invisible to Google Ads reporting however well the account was run — a common and unfair basis for a performance complaint.
• De-duplication turns on three fields. A conversion is identified by the unique identifier, the conversion name, and the date and time. The same conversion will not import twice, but several conversions of the same type against one click will import if their timestamps differ. That is how an import can be inflated: one sale, shifted timestamps, separate rows.
• Enhanced conversions send a hash, not raw data. Google applies SHA256, a one-way hashing algorithm, to first-party customer data and matches it against signed-in accounts.

What a conversion figure cannot establish

Every page on this site names what will not work. For this record:

• You cannot separate modeled from observed conversions in a standard report. No expert can honestly say that of some total, a stated number were observed.
• You cannot treat a conversion count as a count of sales without tying it independently to order or CRM records. Different populations, different systems, different dates.
• You cannot reconcile Google Ads to Analytics to zero. An expert claiming an exact reconciliation has either used a single shared conversion definition or is over-claiming.
• You cannot treat a view-through conversion as evidence an ad caused a sale. No interaction was observed at any point.
• You cannot restate history from the change record. It shows that an action was reconfigured or deleted; it does not recompute the numbers reported under the old configuration.
• You cannot verify the modeling. The model is not published in auditable form and the inputs are not available to the advertiser. Anyone asserting what share of an account's conversions was modeled, without account-specific evidence, is guessing.

How I rebuild a conversion figure that can be tested

Because the platform figure cannot be taken at face value, the work is reconstruction rather than retrieval:

1. Pull the conversion action configuration first, with whatever change record survives, so the definition behind every number is on paper before any number is compared.
2. Test whether the definition held. If an action was redefined, if counting moved from one to every, if the attribution model switched, or if modeling began mid-period, the two ends of a comparison are not the same measurement.
3. Tie the platform figure to the client's own records. Orders, invoices, CRM entries and call logs are the only population checkable independently of the platform, and the gap between the two counts is a finding rather than an error to paper over.
4. Describe a platform figure as a platform figure, with its date range, column, attribution model and known blend of observed and estimated events beside it.
5. State the direction of the uncertainty rather than inventing a share. Modeling adds estimates; a threshold not met removes real conversions; a missed import deadline removes real sales. Those push in different directions, and the account decides which applies.

That produces a figure with its provenance attached. It does not produce certainty about how many sales an advertising program caused, and a report claiming otherwise has stopped describing the data.

Frequently Asked Questions

Published 2026-08-14