A conversion rate divides conversions by visitors, and the visitor count your analytics reports is not the number of people who came.
Two mechanisms cut into it before any marketing question arises, and neither is under your control. They do not remove the same proportion from the numerator and the denominator, which means the ratio moves for reasons that have nothing to do with your site.
This page covers what is actually in that denominator, what the platforms admit about the gap, why the published benchmarks disagree, and what remains measurable.
The formula is trivial. The denominator is not.
Everyone can compute the ratio. Almost nobody can say what the bottom number counts.
What it is supposed to count. People who arrived on your site during the period.
What it actually counts. People who arrived, whose browser executed your analytics tag, and who did not refuse the tracking that tag depends on. Three conditions, of which you control one.
Why that is not a rounding error. The visitors removed by the first two conditions are not a random sample of your traffic. They skew by browser, by device, by technical sophistication and by geography, which means removing them changes the composition of the denominator, not just its size.
The consequence that matters. Your measured conversion rate is a ratio computed on a subset of your traffic, and the subset is defined by conditions unrelated to whether people were interested in what you sell.
What that does not mean. It does not mean the number is useless. It means it is a measurement of a specific population, and comparing it to anyone else’s requires that their population be defined the same way. It almost never is.
Layer one: the visitors who never enter the count
Blocking removes a slice of your traffic before your analytics ever sees it, and the independent measurements agree on the order of magnitude.
The most methodologically careful measurement. Malloy, McNamara, Cahn and Barford, published at IMC 2016, combined a two million person opt-in panel providing ground truth on blocker installs with telemetry from nearly two trillion web transactions collected over February 2016. Ad block penetration in the US: 18%, with a 95% confidence interval of 15.7% to 18.6%. Germany 32%, France 29%, Canada 22%, UK 16%.
A larger, more recent publisher-side measurement. Blockthrough’s 2022 report puts the average adblock rate at 21%, measured across more than 10 billion page views on 9,453 websites. The page does not state the measurement period or a country breakdown, which is worth knowing before quoting it precisely.
Where the variation sits. In the 2016 study, penetration by publisher vertical ran from 17.8% to 24.7%, and across ten large publishers from 17.95% to 25.27%. Your own rate depends on who your audience is, and technical audiences sit far above these figures.
One number you should not generalise. A widely circulated 58% figure comes from a single site during a three-day window when a post was trending on Hacker News and Reddit. It is a real measurement of a very specific audience, and it is not a general rate.
What this layer does to your ratio. It removes visitors from the denominator, and it removes their conversions from the numerator too, but not proportionally: a blocked visitor who converts through a server-side path may still be counted as a conversion while never appearing as a visit. The ratio can therefore be overstated.
Layer two: the visitors removed by consent
This is the larger of the two wherever consent banners are shown, and it has been measured against back-end records rather than estimated.
The comparison that matters, because it uses ground truth. Orbit Media compared analytics counts against conversions actually recorded in CMS and marketing automation back-ends, across 33 accounts and 60 comparisons: newsletter signups, demo requests, ecommerce revenue.
The headline gap. “11.2% of data is missing from websites that don’t use a cookie consent banner”, against “20.3% of data is missing from websites that use a cookie consent banner.”
The traffic-side figure, on one property. On a single GA4 property of roughly 115,000 users, “15.8% of traffic data is missing when users were not shown the consent banner” against “55.6% of traffic data is missing when users were shown the consent banner”.
A citation trap worth avoiding. At least one third-party calculator restates that last figure as “GA4 captured only 55.6% of actual traffic”, which inverts it. The original says 55.6% is missing. Cite the source, not the restatement.
What this layer does to your ratio, and it is the opposite of the first. Consent refusal removes a visitor from the denominator, and also removes any conversion that visitor makes from the numerator. But the two are not removed in equal proportion, because converting visitors interact more with the page and are more likely to have dismissed the banner one way or another. The ratio is usually overstated here too.
Why you cannot correct for it by hand. You would need to know the conversion rate of the people you cannot see, which is precisely the thing you cannot observe.
The modelling that is supposed to fix it has a floor
Platforms offer statistical modelling to fill the consent gap. The documentation is unusually clear about when it does not apply.
What modelling is. Not observation. Google describes it as using machine learning “to model the behavior of users who decline analytics cookies based on the behavior of similar users who accept analytics cookies”. The figures it produces are estimates, and they are labelled as such inside the interface.
The volume floor, stated exactly. Per Google’s documentation, a property qualifies when it “collects at least 1,000 events per day with analytics_storage=‘denied’ for at least 7 days” and “has at least 1,000 daily users sending events with analytics_storage=‘granted’ for at least 7 of the previous 28 days.”
Read those numbers against a B2B site. A thousand denied events per day, every day for a week, plus a thousand consented daily users. A site receiving a few thousand visits a month clears neither condition, ever.
What happens below the floor, in Google’s own words. “If there isn’t enough consented traffic to inform the model, then events triggered by users who decline consent aren’t reported.”
That sentence is the whole point. Not estimated. Not flagged. Not reported. The visitors are gone from the count, and nothing in the interface tells you how many they were.
The asymmetry this creates between companies. A large advertiser clears the threshold and gets a modelled estimate of the gap. A small one does not, and gets a silently truncated count. The two then compare conversion rates as though they were measuring the same thing.
The consent rate you are shown has its own rigged denominator
This one is almost too neat, given the subject: the metric describing consent has the same problem the article is about.
The three definitions, from a consent platform’s own glossary. Didomi defines opt-in rate as “Opt-ins / Total banner displayed”, no-choice rate as “No-choice / Total banner displayed”, and consent rate as “Opt-ins / Total choices (Opt-ins + Opt-outs)”.
What that third definition excludes. Everyone who saw the banner and did nothing. They are removed from the denominator entirely, which mechanically raises the figure.
Why this matters when you read a headline number. A published “consent rate” of 80% and a published “opt-in rate” of 55% can describe exactly the same population. One divides by decisions, the other divides by people shown the banner. Only the second tells you what share of your visitors you can measure.
The US variant to watch for. Some published US acceptance rates combine opt-ins with no-choice, that is to say everyone who did not explicitly refuse. That figure will always look high, and it is not comparable to a European opt-in rate.
The practical instruction. When anyone shows you a consent figure, ask what the denominator is. If it is not banners displayed, it does not answer the question you are asking, which is what fraction of your traffic remains measurable.
Your denominator depends on an interface decision
The last layer is the least discussed and the most uncomfortable, because it is a design choice rather than a technical constraint.
The experiment. Nouwens, Liccardi, Veale, Karger and Kagal, published at CHI 2020, scraped the five most popular consent platforms across the top 10,000 UK websites, n=680, and ran a randomised field experiment with 40 participants across eight banner designs.
The compliance finding. “Only 11.8% meet the minimal requirements that we set based on European law.”
The effect size that concerns your denominator. “Removing the opt-out button from the first page increases consent by 22-23 percentage points; and providing more granular controls on the first page decreases consent by 8-20 percentage points.” The regression coefficient for removing the reject button was 0.220, with a 95% confidence interval of 0.149 to 0.290.
A second dataset in the same paper makes the point harder. The same research group validated its findings against telemetry from a consent platform covering 3 million user logs across 2,000 websites in February 2019. Where no checkbox was pre-selected, 5.59% of users accepted, on 1,135,090 logs. Where all boxes were pre-selected, 98.84% accepted, on 1,988,681 logs.
Read those two numbers together. Same metric, same platform, same month. The measurable share of an audience ran from one in twenty to almost everyone, decided entirely by whether a box arrived ticked.
What that means for your conversion rate. Change your banner design and your measured denominator moves by up to twenty-odd points. Your traffic did not change. Your site did not change. Your conversion rate did.
Why this is worth saying out loud internally. Two companies with identical traffic and identical websites will report materially different conversion rates if one nudges consent harder than the other. Any comparison between them is measuring banner design.
Why the published conversion benchmarks disagree
Two widely quoted US figures differ by a factor of about 1.25, and the reason is not sampling error.
The one that declares its sample and its statistic. Unbounce reports analysing “+41K Landing pages, +464M Unique visitors, +57M Conversions”, and states that “the median conversion rate across all industries is 6.6%”. A median, on landing pages, with a declared sample.
The one that declares neither. LocaliQ reports an average conversion rate of 8.18% across industries on Google Ads traffic. There is no methodology section on the page: no sample size, no date range, no statement of mean versus median, no treatment of outliers.
Why they cannot be compared. Different denominators, landing page visitors against paid search clicks. Different statistics, a stated median against a probable mean. Different traffic, all sources against one paid channel. Any of those three alone would break the comparison.
Why the mean-versus-median distinction is not pedantry here. Conversion rates are skewed. A mean sits above a median on a skewed distribution, which means the undeclared figure is likely inflated relative to the declared one purely by choice of statistic, before any difference in population.
What neither of them can do. Tell a B2B company with a few thousand monthly visitors whether its rate is good. Both describe populations defined by conditions that company does not share, measured through denominators it cannot replicate.
From what traffic does a landing page test conclude?
The question precedes any test, and asking it saves most of the tests from being run.
The standard calculation. For a two-arm test on a proportion, the sample per arm follows n = f(α, β) × [p1(100−p1) + p2(100−p2)] / (p2−p1)², where f(α, β) = [Φ⁻¹(α) + Φ⁻¹(β)]², a formulation published with worked examples by Sealed Envelope and referenced to Pocock’s Clinical Trials.
What it gives at realistic conversion rates. At 95% confidence and 80% power, detecting a move from 2% to 3%, a 50% relative gain, takes roughly 3,800 visitors per variant, so about 7,600 in total. Detecting a more realistic 20% relative gain, from 2% to 2.4%, takes over 21,000 per variant, more than 42,000 in total.
What that means for a B2B site. At a few hundred visits a month, a landing page test cannot conclude. Not slowly, not eventually: by the time you accumulate the traffic, your seasonality, your traffic mix and your offer will have moved more than the effect you are trying to detect.
What people do instead, and why it fails. They stop the test when the numbers look favourable. On these sample sizes, favourable-looking gaps appear and disappear constantly, so stopping on one is a way of guaranteeing a false positive.
What to test instead at low traffic. Things measured on impressions and clicks, which arrive in volumes where statistics behave normally. Ad copy, headlines and offers can be tested conclusively on the same account where a landing page test cannot.
The one exception worth knowing. A change large enough to be visible at your volume is a change you probably did not need a test to justify: a page that answers a different question, or a form that asks for something nobody will give. Fix those on judgement, and save testing for where you have the traffic.
What to measure instead
Four instruments survive everything above, and none of them requires the denominator to be clean.
Absolute counts, not the ratio. Number of enquiries, number of qualified opportunities, number of meetings held. These come from your CRM and your inbox, not from a tag, so blocking and consent do not touch them. They are also the numbers your business actually runs on.
Your own ratio over time, at constant setup. A conversion rate compared with itself is informative as long as nothing changed in the measurement: same tag, same banner design, same consent configuration. Change any of those and you have started a new series, and you should say so.
Back-end reconciliation, once a quarter. Count conversions in the system of record and compare with what analytics reported. The gap is your undercount, it is specific to your site, and it is the only version of that number that means anything for you.
Server-side or first-party measurement, where the volume justifies it. It moves part of the count out of reach of client-side blocking. It does not solve consent, and anyone selling it as a complete fix is overselling.
What to write down when you report a conversion rate. The denominator definition, the date the measurement setup last changed, and the absolute counts underneath. Three lines, and they prevent most of the arguments the number otherwise generates.
What a rising conversion rate can hide
The ratio goes up. Before anyone celebrates, there are three explanations that have nothing to do with improvement.
Your traffic got narrower. Cut a broad campaign, or lose rankings on a generic query, and you remove visitors who were never going to convert. The denominator shrinks, the numerator holds, and the rate improves while your absolute enquiries fall.
Your measurement setup changed. A new banner design, a consent configuration change, a tag moved or reinstalled. Any of these starts a new series, and comparing across the change measures the change rather than the site.
Your traffic mix shifted toward brand. Visitors arriving on your brand name convert at several times the rate of anyone else. If brand traffic grows as a share, your blended rate improves without a single non-brand visitor behaving differently.
The check that settles all three in one line. Put the absolute number of conversions next to the rate. If the rate rose while conversions fell or held flat, nothing improved: the denominator moved.
Why this matters more than it sounds. A rising conversion rate is the easiest metric to present as a win, and it is the one most often produced by losing traffic. Teams have optimised themselves into smaller businesses this way, one quarterly review at a time.
Where to go next
Your counters disagree with each other. Why GA4 and Meta conversions don’t match.
You want to move measurement out of the browser. Server-side tracking.
Your page is the thing you want to fix. What makes a B2B landing page convert.
You are looking at the click rather than the conversion. Click-through rate benchmarks.
Your figure stops at the contact. Cost per lead.
You are working with few deals a quarter. CAC at low volume.
In short
- The formula is trivial, the denominator is not. Your visitor count includes only people whose browser ran your tag and who did not refuse tracking.
- Blocking removes a fifth of it before your tag runs: 18% in the US on a two million person panel, 21% across 9,453 sites and over 10 billion page views.
- Consent removes more. Measured against back-end records, 11.2% of data missing without a banner against 20.3% with one, and 55.6% missing on one property where the banner was shown.
- Modelling has a floor most small sites never clear: 1,000 denied events a day for seven days, plus 1,000 daily consented users. Below it, Google states the events “aren’t reported”.
- The consent rate you are shown excludes people who ignored the banner. Ask for the opt-in rate, which divides by banners displayed.
- Banner design moves the denominator by up to 22 points, per a randomised field experiment, and on 3 million consent logs the acceptance rate ran from 5.59% to 98.84% purely on whether boxes arrived pre-ticked.
- The published benchmarks disagree because they measure different things: a declared median of 6.6% on 41,000 landing pages against an undeclared average of 8.18% on paid search.
- A landing page test needs about 7,600 visitors to detect a 50% relative gain, and over 42,000 for a realistic one.
A conversion rate is not compared to a benchmark, it is defined before it is measured. Book a diagnostic, or see how we approach B2B web design.