Skip to content
ENIT

The method

This page explains where the estimator’s numbers come from. You don’t need it to use the estimator: you need it to decide how far to trust it. The full technical detail is in the METODO.md file of the publicly available source code (in Italian).

In short

GA4 shows only the sessions of visitors who gave consent. About those who didn’t, it knows nothing: neither how many there were nor how they behaved. The estimator estimates the missing part like this:

Observed sessionsfrom the file×declines ÷ acceptancesfrom your rate×how often they returnassumption=Missing sessionsestimate× conversion rate × how well they convert= missing conversions

Three things come from outside the data: your consent rate, a few assumptions about how visitors who don’t accept behave and, if you have them, the real orders from your order system.

This leads to the most important point on the page: the range you see reflects how uncertain the assumptions are, not how noisy the data is. Uploading more days doesn’t narrow it. A more reliable consent rate, or real orders, does.

What is missing from the data

100 visitors60 accept26 decline14 ignore the bannerGA4 sees themGA4 doesn’t see them
Quantity Can it be derived from the GA4 export?
How many visitors accept No. You need the figure from your banner platform (CMP)
How often visitors who don’t accept come back No. Assumption
How well visitors who don’t accept convert, and how much they spend No, unless you have real orders
How much consent varies between channels No. Assumption, or CMP data
Conversion rate and average order value of visitors who accept Yes, with more noise where numbers are small

Only the last row comes from your data. Everything else is a stated assumption, with a range.

The estimator estimates only what is missing because of consent. Traffic GA4 misses for other reasons (ad blockers, technical errors) is left out: see Other losses.

The CMP rate is not a per-session rate

Dividing sessions by the consent rate would be correct only if visitors who accept and visitors who don’t made the same number of visits. Usually they don’t.

Ten visitors in rows: six accept and make three visits each, four don’t accept and make two visits each. 60% of visitors produce 69% of sessions.visitorvisits01020304050607080910
6 visitors out of 10 accept, 18 sessions out of 26 are counted (69%)

How the CMP calculates the rate also matters:

  • per visitor: visitors who accepted divided by unique visitors. This is the best fit;
  • per banner shown: acceptances divided by the times the banner appeared. Visitors who accept never see the banner again, while those who ignore it see it on every visit: the rate comes out lower.

Declaring the wrong measurement is costly: in the tests, a “per banner” rate declared “per visitor” inflates the estimated sessions by 19% (Validation). That’s why the estimator asks, and accepts “I don’t know” by widening the range.

How often they come back

This is the ratio between visits per person for visitors who don’t accept and for those who do.

How the CMP measures Starting value Plausible range
Per visitor 0.8 0.45 – 1.45
Per banner shown 0.6 0.27 – 1.30
I don’t know 0.7 0.27 – 1.80

The reasoning: many visitors who don’t accept are just passing through, so on average they make fewer visits each; but the effect can also reverse, and the range goes above 1. There are no solid public sources for this value: it’s a reasoned judgment.

How well they convert

This is the ratio between the conversion rate of a visit without consent and that of a visit with consent. Starting value 0.4, range 0.15 – 1.05.

The only quantitative indication comes from Google: users who consent “typically” convert 2 to 5 times more, which means a ratio between 0.2 and 0.5. It is stated in the context of Google Ads conversion modeling, with no published data or method, and the same page warns that the gap “varies widely” by industry, consent rate and conversion type. That’s why the range is wider and reaches parity. Academic studies on the GDPR confirm the direction (those who remain trackable are a selected, higher-value group), not a number.

If this assumption is wrong, orders and revenue are wrong: with a true value of 1.0, transactions are underestimated by 20%. The remedy is real orders.

How much they spend per order

The ratio between the average order value of visitors who don’t accept and that of visitors who do. Starting value 1.0, range 0.75 – 1.35: there’s no strong reason to expect a difference, and no source.

Channels, devices, periods

A CMP usually gives a single rate. But visitors arriving from email or direct traffic generally accept more than those arriving from an ad. Without channel data, the estimator assumes each channel may differ from the overall rate in unpredictable ways while keeping the average: the total doesn’t change, but the range for each channel is wider.

This is the least solid part. In tests with very different rates across channels (30% to 85%), the total stays correct, but the session range for a single channel contains the true value less than half the time. The estimator doesn’t ask for rates by channel: almost no CMP provides them. If a channel probably accepts much less than the others, read its figure as indicative.

Different rates by device help calculate the overall rate more accurately (the export doesn’t contain the device). They matter mostly when desktop visitors accept more and also visit more often: with differences like those tested (50% consent on mobile and 75% on desktop, one and a half times the visits on desktop), the overall rate alone overestimates sessions by about 4%, even when declared precise; with bigger differences the error grows. With rates by device the error disappears; if visits are similar across devices, the overall rate is enough. A rate that changed during the period, for example after a new banner, is entered with the date of the change.

Users: why they don’t add up

  • Adding up each day’s users counts returning visitors several times: that’s why the estimator calls them user-days.
  • Unique users in the period can’t be derived from the daily rows: the estimator reads them from the export’s total row, if there is one and it’s consistent; otherwise you read them in GA4 and enter them by hand.
  • For visitors who don’t accept, “user” has never been measured by anyone: without an identifier there’s no way to know whether two visits come from the same person. The estimate of unique users without consent is therefore the most fragile.

Where the range comes from

Each uncertain quantity doesn’t have a single value but a range of plausible values. The estimator repeats the calculation 4,000 times, drawing different values within those ranges each time (technically: a Bayesian model with Monte Carlo simulation), and looks at how the results are distributed.

central estimate8 times out of 1019 times out of 20results of 4,000 simulations

Two safeguards for small numbers:

  • a channel with three conversions doesn’t have “a rate”: it has an uncertain rate. It is pulled toward the site average, the more so the less data it has;
  • observed data is never changed: the uncertainty applies only to the added part.

The same data and the same assumptions always give the same result.

How to read the range

  • “Likely between X and Y” is the range that contains 80% of the simulations: 8 times out of 10.
  • “Almost certainly between X and Y” contains 95%: 19 times out of 20.

It isn’t a figure of speech. On test data with a known answer, the true value falls in the first range about 8 times out of 10 and in the second about 19 out of 20: the Validation page shows it.

There’s a condition: it holds if the assumptions are right. If one is wrong beyond its range, the true value can fall outside. How often, and by how much, is measured on the same page.

Real orders

The difference between orders in your order system and GA4 transactions directly measures how many orders are missing. The estimator uses it to calibrate how well they convert (and, with revenue, how much they spend) instead of the starting assumption.

Effect of real orders
Orders and revenue Very strong: the range narrows a lot
Other key events Weak
Sessions and users None

Why none on sessions? Missing orders pin down a product (how many visits are missing × how well they convert), not the two factors: many visits that convert little and few visits that convert a lot explain the same orders. For sessions, the consent rate is what matters.

“Fewer (or more) orders are missing than your answers predict” appears when, to explain the missing orders, “how well they convert” would have to fall outside its plausible range: too many orders are missing, or too few. Usually the period or scope of the orders doesn’t match GA4’s. In that case real orders aren’t used: transactions, revenue and key events go back to the estimate from the assumptions alone, and the estimator says so. For each other answer about the audience it reruns the same estimate, with the same criterion, and shows which ones would make the numbers add up.

Orders GA4 misses for other reasons

Real orders tell you how many are missing, not why. Some are missing because of consent; others because GA4 wouldn’t have recorded them anyway: ad blockers and browsers that block Google Analytics, technical errors, untracked returns from the payment page.

Real orders from your order systemseen by GA4missing: consentother reasonsthis boundary depends on the audience type

Precise definition. “Other reasons” is the share of real orders that GA4 doesn’t record and that do not come from visitors without consent. In practice:

  • orders from visitors who accepted cookies but have Google Analytics blocked, or hit a technical error;
  • orders from visitors who never saw the banner, because the blocker hides it too: the CMP counts them neither among those who accept nor among those who decline.

An order from a visitor who blocks GA and declines cookies is counted only once, as “missing because of consent”: ad blocker users decline more often, and counting them here too would count them twice. That’s why the share is lower than the percentage of visitors who block GA.

The estimator has no preset value: it asks what audience the site has.

Audience Central value Plausible range
Consumer, mostly mobile 7% 2.5% – 19%
Technical or professional, mostly desktop 20% 8% – 50%
Mixed / I don’t know 9% 2.5% – 32%
Server-side tagging or own proxy 2% 0.6% – 6.5%

The values start from the few public measurements available (6% to 15% of visitors block GA on general-interest sites, 25% to 58% on technical ones), reduced for the overlap with visitors who decline and increased by a couple of points for technical errors. They are old measurements, on single sites and by interested vendors: treat them as an order of magnitude.

With real orders, the total depends almost entirely on this answer. In the tests, with a true share of 20% and the answer “mixed / I don’t know”, the true value stays within the “almost certainly” range but the central estimate is 14% too high; with the wrong answer “consumer”, it falls outside two times out of three.

What about sessions? Ad blockers have an effect even without real orders: the estimated session total doesn’t include visitors GA4 wouldn’t see even with consent. In the tests it stays below actual traffic by 2% with 12% of visitors blocking GA, and by 9% with 35%.

Limits

  • It’s an estimate, not a measurement. Change the assumptions and the results change.
  • The range holds if the assumptions are right.
  • More days of data don’t make the estimate more precise.
  • It covers consent only. Ad blockers and technical errors aren’t included.
  • It assumes opt-in consent, where tracking starts only after the visitor accepts (as in the EU and UK). Under opt-out rules, common in the United States, the default assumptions haven’t been tested.
  • Don’t use it if GA4 is already modeling.
  • The channel breakdown is the least solid part, and small channels have wide ranges.
  • Declaring correctly how the rate is measured matters: getting it wrong shifts sessions by 12% to 19%.
  • User-days aren’t distinct people.
  • The traffic you don’t see isn’t the same as the traffic you see: on average it converts less.
  • With real orders, orders and revenue depend on the audience type you choose.
  • It isn’t suited to fine-grained decisions: comparing two campaigns, reading an A/B test, moving budget to a minor channel.
  • The starting values of “how often they come back”, “how much they spend”, the share of visitors who ignore the banner and the variation across channels are reasoned judgments, without a solid public source.

Sources

  1. Google Analytics Help, Behavioral modeling for consent mode: modeling thresholds and the features that include it.
  2. Google Analytics Help, Reporting identity: the “Blended”, “Observed” and “Device-based” options; changing them has no permanent effect on the data.
  3. Google Ads Help, About consent mode modeling: “2–5x”.
  4. Aridor, Che, Salz, The effect of privacy regulation on the data industry, RAND Journal of Economics, 2023.
  5. Lin, Strulov-Shlain, Choice architecture, privacy valuations, and selection bias in consumer data, 2023.
  6. Measurements of Google Analytics blocking: M. Saric (June 2020, three sites: 13% overall, 6% to 26% depending on the type of site, 21% on desktop and 8% on mobile); Plausible Analytics (August 2021, one article trending on Hacker News: 58%, 68% on desktop and 50% on mobile); Practical Ecommerce (2017, about 210,000 visitors: from 13% on B2C sites to 42% on travel sites). They are measurements on a few sites, made by vendors of competing tools: treat them as an order of magnitude. No more recent independent measurements were found; for context, GWI estimates that in 2025 about 30% of internet users use an ad blocker at least some of the time, which doesn’t mean blocking GA.
  7. Brave, Blocking cookie notices: the browser hides cookie banners by default.

Sources consulted on 15 Nov 2026.

Start