Why your website analytics and your ad platform disagree

Every owner running ads hits this. The platform says forty-two conversions. Analytics says nineteen. The register says eleven new customers. Somebody asks which is right, the agency explains attribution windows for ten minutes, and everybody leaves the meeting less confident than they arrived.

The explanation is usually accurate and completely useless, because it does not tell you what to do. Here is what to do.

What are the actual causes of the gap?

Five, and most disagreements are two or three of them stacked.

Attribution window. An ad platform credits a conversion that happens days after a click, sometimes after a view with no click at all. Analytics generally credits the session where the action occurred. A customer who clicks Monday and buys Friday counts in one place and not the other.

Credit rules. The platform gives itself the conversion. Analytics may assign it to whatever brought the final session, often organic or direct. Two systems, both counting the same sale, assigning it to different owners.

Modelling. When consent is refused or a cookie is missing, ad platforms estimate the conversions they believe occurred and include them in the reported total. That is a legitimate technique and it means part of the number is not observed. Analytics tools do their own version of this.

Blocking and consent. Ad blockers, browser privacy settings, and consent banners stop analytics tags from firing. The visit happened. Your analytics never saw it.

Definition drift. The two systems are counting different events. A platform conversion might include a phone tap, a form view, and a purchase. Your analytics key event might be the purchase alone. This one is fixable and it is the one worth checking first.

Only the last cause is a bug. The other four are the systems working as designed.

Why is definition drift so much bigger than it sounds?

Because two engineers measuring what they both sincerely call the same thing can produce numbers that differ by seconds or by multiples, and neither has made a mistake.

The cleanest published demonstration is not about conversions at all, which is why it works so well as a warning. Kohavi, Deng, Longbotham and Xu, in the KDD 2014 proceedings, dismantle page load time as a metric by citing Steve Souders: an Amazon page can render its above-the-fold content in 2.0 seconds while the window.onload event fires at 5.2 seconds. Gmail is the exact inverse, with onload at 3.3 seconds and above-the-fold content at 4.8. Two tools, one page, a three second disagreement, and both are reporting correctly. Bing abandoned the metric internally and measures Time-To-Success instead.

Google's own current speed metrics carry the same lesson in their definitions. Largest Contentful Paint is good at 2.5 seconds or less, Interaction to Next Paint at 200 milliseconds or less, Cumulative Layout Shift at 0.1 or less, and all three are assessed at the 75th percentile of visits, not at the average. A tool reporting the average will disagree with a tool reporting the 75th percentile every single time, on the same site, on the same day.

Conversion counting works identically. Before you accept any explanation involving attribution windows, open the conversion actions list in the ad platform and the key events list in analytics and read them side by side. Most gaps that look mysterious are two lists that do not match, and that takes ten minutes to establish.

Which number should I actually plan from?

The one closest to money, always.

Rank your sources by distance from a transaction. Your point of sale or booking system is closest. Your own call log and form database come next. Analytics is third. The ad platform is fourth, because it is both the furthest from the sale and the only one with an interest in the answer.

Then assign one source per metric, permanently. Revenue comes from the register. Leads come from your form database and call log. Traffic comes from analytics. Cost comes from the ad platform, which is the one number it reports without any incentive problem.

Write those assignments down and put them on the dashboard. Half of all reporting arguments end the moment everyone agrees which system owns which number.

Use the ad platform's conversion count for one thing only: comparing campaigns inside that platform against each other. It is internally consistent, so relative comparisons hold even when the absolute number does not.

How do I turn the gap into something useful?

Stop trying to close it and start tracking it. Here is the whole method with arithmetic.

Take one completed month, same start and end dates in both systems. Platform reports 42 conversions. Analytics reports 19. The register shows 11 new customers.

Compute two ratios. Platform to analytics is 42 divided by 19, which is 2.21 derived. Analytics to register is 19 divided by 11, which is 1.73 derived. Neither number means anything on its own.

Now repeat it for four months. The platform-to-analytics ratio comes in at 2.18, 2.24, 2.21 and 2.19. That stability is the finding. You now have a working relationship between two systems: whatever the platform reports, roughly 45%derived of it shows up in analytics, and it has done so consistently enough to plan against. 19 divided by 42 is where that comes from.

Month five breaks it. The ratio jumps to 3.60. Something changed, and the arithmetic tells you where to look before anyone theorises. Open the conversion actions list. In this case a phone tap was added as a platform conversion at the start of the month. Remove it from the count and the ratio returns to 2.20.

That is the alarm you actually wanted. Not the gap, which is permanent and normal, but a change in the gap, which is almost always either a definition edit or a broken tag.

The discipline behind that is worth naming. The KDD 2014 authors call it Twyman's law: "Any figure that looks interesting or different is usually wrong!" At Bing, with hundreds of thousands of users per experiment, the striking result was usually an instrumentation problem. On your account, with dozens of conversions a month, that is even more likely. When a number moves dramatically, suspect the measurement before you suspect the market.

How large a gap is normal?

There is no published normal, and anyone who quotes you one is guessing.

The size depends on your consent rate, your customer's decision length, your device mix, and how each system is configured. A business selling a considered purchase over three weeks will see a larger gap than a restaurant taking same day bookings.

It is worth seeing how far this goes even among careful people measuring the whole web. SparkToro, using a Similarweb panel of US Google searches from January to April 2026, put the share of searches ending without a click at 68.01%. A separate 2026 measurement using a different definition of the same phenomenon reported 27.6%, and Google has publicly disputed SparkToro's methodology. Most of that spread is one definitional question: does a click to another Google property count as a click. Two competent measurements of one thing, four decades of internet research behind them, and a gap of forty points because the definitions differ. Your platform and your analytics are a small version of the same problem.

There is also a newer contributor that no configuration fixes. Clicks arriving from Google's AI Overviews and AI Mode are included in ordinary organic traffic, with no referral parameter and no way to separate them. Some assistant traffic arrives with the referrer stripped entirely and lands in direct. Part of your unexplained variance is now structurally unattributable, and the right response is to label it rather than model it, which is the same discipline as naming the boundary between counted and modelled in AI reporting instead of buying a score that hides it.

Why does this keep costing businesses money?

Because the gap gets used as an argument rather than a fact.

A vendor reporting from the platform will always show a stronger month than one reporting from your register. That is not necessarily dishonesty. It is the natural result of reporting from the system that flatters the work, and nobody volunteers the harder number.

The second cost is decision paralysis. Owners who cannot reconcile two dashboards often stop looking at both, which is worse than trusting the flattering one. A monthly review habit that survives busy season depends on the numbers being settled in advance, because a review that reopens the same argument every month gets cancelled.

The third cost is delegation. If your manager cannot tell which number is the real one, they cannot act without you. Owner operators who cannot be everywhere need a single agreed number per decision more than they need a better dashboard.

What to do this week

Open both systems for the same completed month. Same start date, same end date. Write the two conversion numbers down.

Check definition drift first. Open the conversion actions list in the ad platform and the key events list in analytics. If they are counting different actions, fix that and the gap shrinks immediately. Then write both lists into your documentation so an edit is visible next time.

Check the attribution window setting in the ad platform and write it down. It is usually adjustable, and whatever it is set to is now part of your documentation.

Assign one source per metric and put it on the dashboard in writing. If you are considering live chat or a chatbot, decide how its conversations will be counted before you install it, because it adds a sixth source to reconcile.

Then set up the ratio as a monthly check. Two numbers, one division, one note when it moves.

If you are starting a new engagement, do all of this in the first two weeks. It belongs in what you measure in the first ninety days, because agreeing on sources after a disappointing quarter is a conversation nobody wins.

Questions about your numbers not matching?

Email me at eric@seod.com with two numbers from the same month: what your ad platform reported as conversions, and what your analytics reported. Tell me your industry. I will tell you which of the five causes is most likely driving your gap and what to check first, in one email.

I answer these myself. If the gap looks like a genuine tracking break rather than normal disagreement, I will say that plainly, because it changes what you should do next.

More on measurement sits in the analytics library.