Some of those conversions never happened

Open Google Ads, look at the Conversions column, and pick a number out of it. Some of what you are looking at was measured: a click happened, a tag fired, the two were joined, and a sale exists somewhere at the end of it. Some of it was not. Some of it is an estimate of how many conversions probably occurred among the people the platform could not observe, produced by a model, and added into the same column as the rest.

This is not a scandal and it is not hidden. Google documents conversion modelling thoroughly, explains where it is applied, publishes the thresholds an account has to clear before it runs, and gives a reasoned account of why estimating is better than reporting zero. Every large platform now does some version of it. Given that consent banners, tracking prevention and cross-device journeys have made a substantial slice of the path genuinely unobservable, modelling is a sensible response to a real problem.

What is a problem is the reporting convention that follows from it. The estimate and the observation share a column. They share a decimal place. They are exported together, charted together, and pasted into a board slide together, and by the time the number reaches somebody making a budget decision there is nothing left in it to say which part was seen and which part was inferred. A finance director asking can you show me one of these is asking a reasonable question, and for a modelled conversion there is no answer.

A conversion you cannot point at is a statistic. A conversion you can trace to an invoice is evidence. Both belong in a report; only one of them belongs in a column labelled Conversions with no further comment.

What conversion modelling is actually doing

Conversion modelling estimates the conversions that could not be observed by measuring the relationship between users who can be observed and users who cannot, then applying that relationship to the unobserved population.

The logic is straightforward and, in isolation, hard to argue with. Suppose a hundred people click an ad. Sixty of them consent to advertising cookies and can be followed to a purchase; forty do not, and vanish at the click. If the sixty convert at four per cent, reporting the campaign's conversions as only what the sixty produced understates the campaign — not because the model says so, but because the other forty were real people who did not stop existing when they declined a cookie.

Modelling closes that gap by estimating what the forty did, using the observable behaviour of comparable users, adjusted for the ways the two groups are known to differ. It is applied in more places than most advertisers realise:

• Cross-device journeys, where the click is on a phone and the purchase is on a laptop and no shared identifier ties them together.
• Browsers that block or time-limit the cookies conversion measurement depends on, so the path exists but the record of it expires before the sale does.
• Consent mode, where the user declined advertising or analytics storage and no tag was allowed to write anything.
• App journeys where the device identifier is unavailable, with platform APIs and aggregated signals used in place of the missing join.

That final sentence is the one worth sitting with. The revenue attached to a modelled conversion is itself modelled — predicted from the values of the conversions that were observed. So a channel that happens to attract higher-value customers among the people who refuse consent will be understated, and one that attracts lower-value customers among them will be overstated, and neither error is visible from inside the platform, because the platform has no way to see the values it could not observe.

Consent mode, and the number that appears where a gap was

With consent mode implemented, tags still fire when consent is refused but do not read or write cookies; the resulting gap in measured conversions is then filled by modelling, provided the account is large enough to qualify.

Consent mode is the mechanism that connects a cookie banner to an ad report. When a visitor declines, the tags adjust their behaviour rather than disappearing: no advertising or analytics cookies are read or written, and cookieless signals are sent instead. Those signals are what the model then works from.

Crucially, this only happens above a threshold. Modelling requires enough observed behaviour to learn from, so small accounts and small markets get no modelled conversions at all — they simply get the gap.

Seven hundred ad clicks in seven days, per country and per domain grouping, is a serious volume bar. It means the same consent banner produces two entirely different measurement realities depending on the size of the account behind it, and — more awkwardly — two different realities within one account across its markets.

Read across the middle row and the reporting problem is obvious. In the large market the channel looks like it is working; in the small market the identical performance looks like it is failing; and the difference between them is a volume threshold in somebody else's system. A multi-market advertiser comparing country performance from platform data alone is, in part, comparing eligibility for modelling.

If two markets run the same campaigns with the same banner and one reports far better returns, check whether the difference is the market or the threshold before reallocating anything.

Analytics models too, and it is honest about the limits

Behavioural modelling in Google Analytics estimates the activity of users who declined analytics cookies from the behaviour of users who accepted, under its own thresholds — and the modelled portion is deliberately excluded from several places, including the raw data export.

The same idea appears one layer down in analytics, and this is where the shape of modelled data becomes clearest, because Google spells out where it does not work.

That last exclusion is not an oversight, and it is the most useful thing in the whole documentation set. Modelled data cannot be exported to a warehouse because there is nothing to export. A modelled conversion has no event, no timestamp, no identifier and no customer. It is a quantity that exists at the level of a report and dissolves the moment you ask for the rows underneath it.

Which explains an argument that happens in a lot of companies and is usually mistaken for a tagging fault: the platform reports one number, the warehouse reports a smaller one, and everybody goes looking for the broken tag. Frequently nothing is broken. The two numbers are measuring different things — one includes an estimate of the unobservable, and the other, by construction, cannot.

Why this matters more than it used to

As the observable share of the customer journey shrinks, the modelled share of reported performance grows — so the same reporting habit that was harmless when modelling filled a few percent becomes load-bearing when it fills a quarter.

It is worth being fair to the platforms here. They did not choose this. Consent requirements, tracking prevention and the general collapse of the third-party cookie removed a large part of the measurable path, and an advertiser who insisted on observed conversions only would be looking at a systematically understated account and would cut spend on channels that are working. Modelling exists because the alternative is worse.

But three consequences follow, and they are the ones that should change how the report is built rather than how the platform is configured.

1. The modelled share is not stable. It moves with consent rates, browser policy, market mix and campaign mix. A year-on-year comparison of conversions can therefore move without any change in customer behaviour at all, and there is no line in the report that says so.
2. Modelling is calibrated on the observable population, which is not a random sample of your customers. People who accept cookies differ from people who refuse them — by age, by device, by market, by how much they care. Any systematic difference in purchase value between the two groups becomes a systematic bias in modelled revenue.
3. Every platform models separately, on its own data, with its own thresholds and its own definition of a conversion. Two platforms can both be right about their own estimate and still sum to more conversions than the business actually had. Nobody is double-counting deliberately; nobody is in a position to notice.

The third one is the one that reaches the boardroom. Add up the conversions claimed across the paid channels of a mid-sized advertiser and compare the total to the number of orders in the finance system. The gap is rarely small, and it is rarely explicable from inside any single platform.

Every platform is estimating the part it cannot see, and none of them can see each other. The totals were never going to reconcile, and only one of the systems involved has a legal obligation to be right.

What to reconcile a model against

The only external check on a modelled conversion count is the record of what was actually sold — the closed-sales file — compared over a period long enough for the tail of the sales cycle to land.

You cannot audit a modelled conversion individually. There is no row to inspect and no customer to call. What you can do is check the aggregate against a number that was not produced by a model, and there is exactly one of those in the business: what was invoiced, to whom, and when.

The comparison is not conversions against sales, which will never agree and are not defined the same way. It is a set of relationships, tracked over time:

• Total closed revenue against total reported conversion value across all platforms, per month. Not to make them equal — to watch the ratio, and to notice the month it moves.
• The share of closed sales that can be traced to a specific, observed ad interaction. This is a real measurement, it goes down when the observable path shrinks, and it is the honest denominator for everything else.
• The Direct / Unknown bucket: sales that match nothing. Its size is the actual scale of the measurement gap in your business, expressed in money you definitely received rather than in conversions somebody estimated.
• Consent rate by market, next to the modelling threshold. This tells you which markets are being modelled, which are simply going unreported, and which comparisons between them are meaningless.

Do that for two or three quarters and something useful appears: a stable relationship between platform-reported performance and money in the bank, per channel. That relationship is the thing to plan with. It is not a correction factor to apply to the platform's number — it is a calibration, and the point of it is that it is derived from your own outcomes rather than from a model built on somebody else's population.

This is the problem CloseRev exists for. It reconciles a closed-sales export against a lead or call-source export, matches sales to the interaction that produced them on normalised phone numbers and emails, and reports what was matched, what was matched weakly enough to need a human decision, and what matched nothing at all. Nothing is modelled and nothing is inferred: a sale either traces to an interaction or it goes to Direct / Unknown.

How to report a number that is partly an estimate

A defensible report separates conversions that were observed from performance that was estimated, states the reconciled revenue from the sales system on its own line, and never presents a modelled total as a measurement.

The point is not to stop using modelled figures. They are the best available estimate of something real, and refusing to look at them is its own kind of error. The point is that a number's provenance travels with it or it does not travel at all.

1. Reconciled revenue first: closed sales matched to a specific ad interaction, from your own records. This is the number that survives being questioned, and it belongs at the top rather than in an appendix.
2. Platform-reported conversions and conversion value second, labelled as including modelled data, with the consent rate for the period beside them so a reader can see how much of the path was observable.
3. Direct / Unknown as its own line, reported as revenue with no traceable source rather than being distributed across channels by any rule, modelled or otherwise.
4. The ratio between reconciled revenue and platform-reported value, tracked over time. Movement in that ratio is a finding in itself and is usually the earliest warning that measurement, rather than performance, has changed.
5. A one-line note on what changed in measurement this period — a consent banner redesign, a new market crossing the modelling threshold, a browser policy shift. These move the numbers as much as campaigns do and are almost never written down.

None of this requires disbelieving the platforms. Their models are built by people with far more data than any advertiser has, and for the question they are answering — roughly how many conversions did this campaign produce among people we could not follow — they are probably closer to the truth than any alternative available.

But that is a different question from the one a business is asking when it decides where next quarter's budget goes. That question is which of these channels produced money we actually received, and it is answered by joining two files that both belong to you. A model can estimate what you could not see. It cannot tell you who paid, and there is a file in your own systems that can.

Questions people actually ask

What is a modelled conversion?

A conversion that was not observed, but estimated. When the path between an ad click and a sale cannot be measured — because cookies were refused, blocked or expired, or because the journey crossed devices — the platform uses the behaviour of measurable users to estimate how many conversions the unmeasurable ones produced, and adds that estimate to the reported total.

Can I tell which conversions in my Google Ads report were modelled?

Not row by row. Modelled conversions appear in the same Conversions column as observed ones and flow into every downstream report that uses that column. Consent mode has its own impact reporting that shows the aggregate effect, but there is no per-conversion flag that lets you separate the estimated from the measured in a campaign report.

Is conversion modelling inaccurate?

It is an estimate, which is a different thing from inaccurate. Modelling is built on real observed behaviour and has to clear volume thresholds before it runs at all. But it produces a population-level figure, not a set of events, so it can be a reasonable total and still contain no specific conversion you can point at, name, or trace to an invoice.

Why does the modelled figure not appear in my data warehouse?

Because there is nothing to export. A modelled conversion is a number in a report, not a row with an identifier, a timestamp and a customer attached. Google states plainly that modelled data is not available in the BigQuery export from Analytics — which is why platform totals and warehouse totals disagree, and why the warehouse is the one you can audit.

What should I reconcile modelled conversions against?

Your own closed-sales records. The platform's estimate answers how many conversions probably happened; your sales file answers which sales definitely happened and what they were worth. Comparing the two over a long enough period is the only way to find out whether the model is running high, low, or about right for your business.

Found this useful?

• Share on X
• Share on LinkedIn
• Share on Reddit
• Share on Email