In short
- A consent rate is not a percentage of lost data. At 60%, you are measuring a sample whose composition no longer looks like your real audience.
- Refusals do not fall at random. They concentrate on certain browsers, certain devices and certain traffic sources, which makes your channel comparisons wrong.
- Modelling fills part of the gap, under conditions. Google sets volume thresholds, and modelled data stays aggregated: it never comes down to the individual user.
- Inflating consent with a deceptive interface is expensive. On 31 December 2021, the French regulator fined Google 150 million euros and Facebook 60 million on that ground alone.
A marketing manager cuts a channel’s budget because its cost per acquisition comes out 50% higher than another’s. Yet both channels produced the same number of sales for the same budget. The gap does not come from their performance: it comes from their consent rate.
This situation comes up regularly in audits and triggers no alert. The figures are consistent with each other, the dashboards fill in normally. The decision is simply taken on a sample that does not represent the population it claims to describe.
The regulatory framework is covered elsewhere, in our guide to audience measurement without consent. What follows is about something else: what consent changes in your figures, and therefore in your trade-offs.
A consent rate is not a percentage of lost data
The common reading is arithmetic: 60% consent, so 40% of traffic invisible, so you just scale back up. That reading assumes the missing 40% behave like the visible 60%. Nothing guarantees it.
A sample reduced at random stays usable: proportions are preserved, only precision drops. A sample cut on criteria correlated with what you are measuring is not. You are not losing precision, you are losing representativeness.
What to take away A small sample makes you hesitate between two close figures, and time fixes that: you wait for more data before deciding. A biased sample makes you certain of a figure that always leans the same way, and three more months change nothing.
Why refusals do not fall at random
A visitor arriving from your newsletter already knows you. A visitor arriving from a social ad discovers your brand at the moment the banner appears. Those two people do not answer the same way.
The same gap exists on hardware. The environments most restrictive about trackers (Safari, iOS, browsers with built-in blocking) also concentrate the users who care most about these questions. Screen size matters too: a banner that reads well on a desktop becomes a wall of text on a phone.
Your channels are therefore not measured with the same intensity, and that gap adds itself to the performance gap without anything separating the two in your report. Same fragility as in our article on tracking that degrades in silence: nothing flags the error.
The arithmetic of the bias, on a simple case
Take two channels that produce strictly the same thing: 1,000 euros of budget, 40 real sales, 25 euros cost per acquisition. Only their consent rate differs.
| Channel | Budget | Real sales | Consent | Measured sales | Cost per acquisition shown |
|---|---|---|---|---|---|
| Channel A | €1,000 | 40 | 75% | 30 | €33 |
| Channel B | €1,000 | 40 | 50% | 20 | €50 |
Channel B looks 50% more expensive when it is identical. Budget moves to A, and the decision will be confirmed the following month since the bias repeats itself. The error is stable, reproducible, and therefore credible.
Notice what this calculation says: at 50% consent on both channels, the comparison would have stayed correct despite losing half the data. It is not the level of the rate that breaks the decision, it is how unevenly it is spread.
What modelling fills, and what it does not
When consent is refused, Google’s consent mode sends a signal without an identifier, and the platforms estimate the missing conversions from the behaviour of consenting users. Since March 2024, this setup is a condition of access to Google Ads audience features in the European Economic Area.
That estimate is not automatic. Google makes behavioural modelling in Google Analytics 4 conditional on minimum volumes: in the order of a thousand daily events with consent denied over at least seven days, and as many daily users with consent granted. These thresholds can change: check them before building a decision on them.
Most small and mid-sized company sites never reach them: modelling never kicks in and the gap stays whole. It is also what decides the choice between basic and advanced Consent Mode, a subject we treat separately, with the published figures. For the others, the estimate assumes that visitors who refuse convert like those who accept, a fragile assumption since the two populations differ structurally.
| Use | What modelling brings | The limit to know |
|---|---|---|
| Conversion volume per campaign | An estimated figure if the thresholds are met | Aggregated: no individual journey is reconstructed |
| Raw export to a data warehouse | Nothing | Only genuinely observed events are exported |
| Remarketing audiences | Nothing usable | An audience is built on identified users |
| Comparing low-volume channels | Nothing below the thresholds | Which is exactly the case of the sites that would need it most |
The second row deserves attention if you feed a data warehouse: modelled events do not appear in the raw GA4 export.
What actually moves a consent rate
Four factors weigh, in the order of frequency observed in audits: when the banner appears, how many clicks it takes to refuse everything, how the purposes are worded, and the context of the visit. A banner that appears before the page has finished drawing gets reflex answers.
Wording matters more than people think. “Audience measurement and personalised advertising” is understood; a list of legal bases is not, and confusion pushes towards refusal. How to write that banner is the subject of a dedicated article: designing and writing a cookie banner.
We deliberately give no target rate. The averages published by consent platform vendors vary too much by sector and by calculation method to serve as a benchmark. Your point of comparison is your own rate last month.
The line not to cross
A rate can be managed, but not at any price. On 31 December 2021, the French data protection authority issued a 150 million euro fine against Google and 60 million against Facebook, on the ground that refusing took more clicks than accepting. On 12 January 2023, TikTok was fined 5 million euros for an identical reason.
The calculation is bad in both directions: on one side the risk of a fine and its publicity, on the other a data point that matches no intent. You are measuring a click of fatigue, not a signal of interest.
The mistake not to make Hiding refusal behind a second level, greying out the reject button, pre-ticking purposes. Those three practices are exactly what the fines above targeted, and they degrade the very signal you were trying to improve.
What we look at in an audit
The diagnosis fits in six indicators, all readable in your measurement tool and your back office.
| Indicator | What it reveals | Action to take |
|---|---|---|
| Overall rate, day by day | A sudden drop signals a technical incident, not a change in behaviour | Cross-check the dates against your releases |
| Rate by traffic source | The gap between channels becomes a fictitious gap in cost per acquisition | Reweight conversions before any budget decision |
| Rate by browser and device | A marked gap on mobile or Safari distorts how you read your audiences | Check that the banner is readable on those formats |
| Share of visits in unknown status | That is not a refusal: it is a signal that never arrives | Check the loading order between banner and tags |
| Gap between real orders and measured conversions | The size of the shortfall, all factors combined | Compare day by day over 30 days |
| Daily volume with consent denied | Whether or not you are eligible for modelling | Compare that volume with the thresholds Google publishes |
Five moves are enough to start this week.
- Export your daily rate over 30 days and spot the breaks in the curve.
- Segment that rate by traffic source, then by device.
- Recalculate each channel’s cost per acquisition, corrected by its own rate.
- Compare, day by day, your real orders and your measured conversions.
- Have three people outside marketing test your banner, and count the clicks needed to refuse everything.
If step 3 changes the ranking of your channels, you have found the origin of a questionable trade-off. If step 5 takes more than two clicks, you have a compliance subject, fixable in your consent management platform.
Frequently asked questions
What is a good consent rate in 2026?
There is no usable reference value: published averages vary threefold depending on the sector and the calculation method. What counts is how stable your rate is and how evenly it is spread across channels. A rate of 45% that is the same everywhere hampers a decision less than an average of 70% hiding 85% on one channel and 40% on another.
Can I correct my figures by dividing them by my consent rate?
Partly, on one condition: apply a rate specific to each channel, never the overall rate. The correction restores aggregate orders of magnitude, which is often enough to arbitrate a budget. It reconstructs neither individual journeys nor multi-touch attribution.
Does Google’s modelling replace the missing data?
It estimates it, it does not replace it. It requires daily volumes that most small and mid-sized company sites never reach. When it does kick in, it provides aggregated totals: those estimated conversions never come down to the individual user and do not appear in a raw export.
Does server-side tracking increase my consent rate?
No, these are two different layers. Server-side makes the transmission of already authorised events more reliable, it does not change the visitor’s decision. A setup that sent the data of visitors who refused would not be an optimisation but a breach.
Is it better to show the banner later to improve the rate?
Delaying it mainly creates a grey area: visits in unknown status, neither accepted nor refused, which damage how you read the data. As long as the choice has not been expressed, non-exempt purposes stay inactive. The apparent gain is paid for in uncertainty.
Deciding while knowing your sample
None of these checks will push your consent rate back up. They say something else: to what extent your figures describe your business, and on which channels they describe it least well. That is what makes a trade-off defensible.
Consented data is scarcer than it was in 2020 and it will stay that way. The work consists in knowing what you measure and what you do not, rather than pretending to see everything. Our data, analytics and performance engagements almost always start there.
Unsure what your consent rate is doing to your decisions? In a first meeting we look at your rate day by day over 30 days, how it splits by channel and by device, and the gap between your real orders and your measured conversions. Let’s talk.