Conversion rate — what the number actually tells you, and what it does not

The conversion rate has the simplest formula in all of marketing and is the most misread number on a company website. You divide the number of enquiries by the number of visits, multiply by a hundred, and you have a result that sounds like a grade for your work.

It is not a grade. It is a fraction in which the numerator and the denominator come from two different systems, and both measure something other than what they appear to. This text shows that on our own data, including the parts of it that do not add up.

The formula takes one sentence; the whole problem sits in the denominator

The conversion rate is the number of conversions divided by the number of visits. The trouble starts with each of those two words.

What counts as a "conversion" for you? A submitted form, a phone call, a newsletter sign-up, a downloaded proposal, a booked appointment? Each of these produces a different number, and the difference between them can be tenfold. A site with one hard goal and a site counting five micro-conversions do not have comparable results, even though both report a "conversion rate".

Who is in "visits"? That question settles more. The denominator holds bots, your own staff, your agency checking a fix, people who arrived via an irrelevant search, and everyone who declined cookies and was therefore never seen by the analytics at all.

Until both sides of the fraction are defined, the percentage is a number, not information.

It helps to separate two things that usually share one row in a report.

A hard conversion is the moment you learn enough about somebody to call them back: a submitted form, a phone call, a booked appointment. That is the only one fit to be the number you give the board, because it is the only one that turns into revenue.

A micro-conversion is a step along the way: finishing a calculator, downloading a specification, looking at a price list. It is fit for diagnosis, not for a report. Our own best example is the calculator: 167 of 715 people finished one, which is 23.4% — valuable information that the tool works, but it would be an abuse to call it the site's conversion rate.

The difference is practical. Micro-conversions rise easily, because all it takes is placing them further down the path. If you enter them in the same row as form submissions, you get a rising chart with no change in the number of customers — and that is the most common way a conversion report stops describing reality.

The rule we apply to ourselves: one hard conversion as the headline number, micro-conversions kept separately, and never added into the same percentage.

Our 180-day funnel — and the step that came out above one hundred per cent

Below is our own data from our Polish domain for the period from 19 March to 14 September 2026, counted in users, not events.

Four steps: 715 people viewed a page, 167 finished one of the calculators, 20 touched the form, 21 sent an enquiry.

The last step is 105% — more submissions than touches. This is neither something to be proud of nor a spreadsheet error. It is proof that the two steps do not measure the same population, and we checked why rather than guessing: the "touched the form" event fires only from the contact form block, on the first click into a field. The "sent an enquiry" event fires from two places — from there, and from the form that emails calculator results, which never fires the first event at all.

Somebody who worked out a cost in the calculator and asked for the results by email is in the numerator and absent from the denominator. This is not a funnel; it is two separate measurements stacked one under the other.

A second finding from the same set: there were 92 "touched the form" events but only 20 people — an average of 4.6 each. On one day, 22 events came from a single person, and that person was our own test session. Had we reported events instead of people, we would have announced ninety-two form starts and drawn a rising chart from it.

Both systems lie — in opposite directions

The natural reflex is: if the analytics loses people, count the enquiries in the database, because a database does not lie. We tried that on ourselves and the result was instructive.

Analytics understates. It sees only people who accepted cookies. Every figure from it is a lower bound, and the true denominator cannot be reconstructed from it.

The database overstates. Our enquiries table holds 79 rows for the same window. Going through the sender addresses, it turned out that 54 of them were our own submissions — tests sent from the company domain, including one signed literally "test". Two more were our other addresses, two pairs were the same person submitting twice, and several more were foreign agency pitches, which is to say sales spam through the contact form.

What remains is about a dozen genuine enquiries in 180 days.

The conclusion is sharper than it looks, and it is the main reason this text exists: had we set "21 according to analytics" against "79 according to the database", we would have had a striking paragraph about how badly analytics loses leads. It would have been untrue. The truth is that neither of those numbers is a conversion rate until somebody goes through the numerator and the denominator by hand. We had to do that on our own data in order to write this paragraph.

What we changed as a result: our own addresses and our own traffic are excluded from the count, and the lead figure comes from the database after a manual review, not from an analytics dashboard. Analytics is kept for what it is good at — the shape of the path, which is to say where people stop, not how many of them there were.

What a "good" conversion rate is

This is the question asked most often, and the answer is not the one you would like.

Nielsen Norman Group gives a range of 1–10% as typical, and about 3% for online shops in 2013. More important, though, is the sentence Jakob Nielsen adds in the same text: a good conversion rate is one that is higher than your previous one. The author of the range advises against using it as a reference point, because the result is affected by brand recognition, price, the complexity of the sale and the size of the commitment you ask of a visitor.

Contentsquare, in its 2026 benchmark, reports that conversion fell by 5.1% year on year — across a sample of 6,500 sites and 99 billion sessions, comparing the fourth quarter of 2024 with the fourth quarter of 2025. That is a figure about the direction of the market, not about your site, and it has to be read that way.

Baymard Institute reports average cart abandonment at 70.22%, calculated from 50 studies between 2006 and 2025. And here is the most important detail of their methodology, left out of most citations: 42% of abandonments happen because somebody was "just browsing" and never intended to buy. That share is mathematically unrecoverable by any redesign.

The practical conclusion from all three: a benchmark is fit for checking whether you are in the right order of magnitude. It is not fit to be a target.

Why comparing yourself to your industry almost never works

Benchmarks have one flaw that invalidates most of their uses: they compare numbers produced from different definitions.

Nielsen names four things that move the result regardless of the quality of the site, and they are worth knowing before somebody shows you an "industry average":

Brand recognition. The same form on the site of a company the visitor has heard of, and on one they are seeing for the first time, produces two different results. What you are measuring then is reputation, not design.

Price and the size of the commitment. A newsletter sign-up and an enquiry about a project worth tens of thousands are two different decisions. A site selling at a higher price will show a lower rate on better business.

The complexity of the sale. In business services the first contact is rarely the last. Conversion there measures entry into a conversation, not a close — and comparing it with a shop's conversion rate makes no arithmetic sense.

The traffic source. The same page converts differently on search traffic and on campaign traffic, because those are different people at different points in a decision.

Add to that what Contentsquare shows: the market as a whole moves over time, and at the moment it is moving down. If your result has not changed year on year while the market fell, that is not stagnation but a relative gain — and no single benchmark will show you that.

The only comparison that yields information is the comparison with yourself: the same definition of conversion, the same denominator, two consecutive periods of the same length.

Why other people's results do not transfer to your site

The internet is full of case studies reporting a two- or three-hundred per cent lift after one element was changed. The three documents this text was built from contained such figures themselves — which is why they were deleted rather than corrected.

There is an answer to this from research rather than opinion. Ron Kohavi and co-authors, who ran the experimentation platform at Microsoft, published a paper in 2014 called "Seven Rules of Thumb for Web Site Experimenters". What follows from it should end most conversations about other people's case studies:

• The success rate of ideas at Bing is 10–20%. Quoting the paper: "If our success rate on ideas at Bing is about 10-20%, in line with other search engines…". One idea in five works.
• The average across Microsoft is one third. The authors describe it as "the average across multiple experiments at Microsoft".
• Rule #3 is literally called "Your Mileage WILL Vary" and says outright that attempts to reproduce the spectacular results others describe usually do not end the same way.

There is one more calculation there, worth remembering for anyone planning to run tests. At the standard significance threshold, and assuming one third of ideas work, a statistically significant result is true in 89% of cases. But if you are hunting for a breakthrough — a change on the order of one in five hundred — that same statistically significant result is true in only 3.1% of cases. In other words: the more spectacular the effect you are looking for, the greater the chance that what you found is chance.

That is the mechanism by which 300% lifts come into being. Somebody ran a test, got a large result, wrote it up — and never repeated it.

What actually changes conversion on a company site

Below are the things that recur in our own projects regularly enough that we treat them as a starting point. Deliberately without percentages: had we attached a number to each, we would be reproducing exactly the genre the section above has just taken apart.

One path to action instead of five. Four equally weighted buttons on a home page mean the visitor picks none of them. Choose the action that should happen most often.

A form shorter than you would like. A first contact needs only enough to call somebody back. Fields for a tax number, a budget and "how did you hear about us", added just in case, cost you enquiries and rarely change what you do with the first reply.

A form that actually arrives. A "thank you" message proves nothing. The proof is a message in an inbox somebody reads every day. Forms can quietly go to spam, or to a former employee's address, for months — and an absence of enquiries looks exactly like an absence of interest.

Contact details visible without searching. In service businesses the phone number is often the most used element on the site and the hardest to find.

Proof that the company exists. Photographs of your own work rather than stock images, registration details in the footer, a described scope of services rather than the phrase "comprehensive solutions". This is the element that cannot be bought along with the site, and its absence stops people just short of the form.

A price, or a range. A site silent about money filters traffic the wrong way: the people who had a budget and did not want to ask drop out, and the people who ask about everything remain. How to give a range without tying your hands, we describe in the cost section.

The speed of the first screen. Not because "every second costs X per cent of conversion" — we have not measured that figure and we do not quote it. Because a user who never saw the content had nothing to convert on. How to measure it on your own site, we describe in the text on building a website.

What to do when the result is low

A low rate is not a diagnosis but a symptom, and there are three causes requiring three different responses. Telling them apart takes a quarter of an hour and saves you from redesigning a site that happens to be working.

Bad traffic. The most common cause, and the only one where improving the site will not help. If the people arriving are looking for something other than what you sell, no button will reverse that. How to check: look at the queries that bring people to the site. If informational phrases or a different product dominate, you have a traffic problem, not a conversion problem. Our own denominator — 715 people in 180 days — is made up largely of readers of articles who never intended to order anything, and it would be a mistake to treat them as customers who got away.

A bad offer. The site works, the traffic is right, and still nobody asks — because what you are proposing is more expensive, slower or less clear than the competitor the visitor opened in the next tab. How to check: ask the last three customers what convinced them, and three who did not buy what stopped them. That is cheaper than any analytics tool and more often decisive.

A bad site. Only third in order, though usually checked first. How to check: sit five people from outside the company down and give them the task "find the price list and send an enquiry". We describe the method in the wireframe text — it works the same way on a finished site.

And a fourth case, which looks like all of the above: the result is low because the measurement is broken. Before changing anything, check that the form reaches the inbox, that your own traffic is excluded, and that the funnel steps measure the same population. On our site both of those were broken.

How to measure so that the result means something

Five things, all of them following directly from what went wrong here.

Count people, not events. Our 92 events were 20 people, and on one day 22 events were one person — us.

Cut yourself out. Two thirds of the rows in our enquiries database were our own tests. If you do not exclude your own traffic and your own addresses, you are mostly measuring yourself.

Define one hard conversion and hold to it for the whole measurement period. Adding a micro-conversion halfway through a quarter produces a rise that did not happen.

Check that the funnel steps measure the same population. Ours came out above one hundred per cent, and only reading the code explained that the two events came from two different forms.

Wait for a sample. At about a dozen enquiries in six months — which is what we have — the difference between one month and the next is noise, not a trend. Our own funnel is not fit for drawing conclusions about the effect of changes, and we do not pretend otherwise.

What this text does not settle

How to run A/B tests. Methodology, sample size and when to stop a test are a separate subject with a text of their own — together with Kohavi's calculation set out in full.

How to write calls to action. Button copy, placement and what distinguishes a CTA from a button also get their own text.

Where to get traffic. Conversion is a fraction; this article is about the numerator and the denominator, not about how to enlarge the denominator. That is the marketing section.

How to design a layout that leads to action. Hierarchy, the order of sections and the path to conversion are settled on the sketch — we describe that in the wireframe text.

Where these figures come from

• Our funnel: 715 / 167 / 20 / 21 users — our own measurement of our Polish domain over 180 days, from 19 March to 14 September 2026, using pnpm ga4:funnel. Counted in users. The caveat travels with the figure: the analytics sees only people who accepted cookies, so each of these values is a lower bound.
• 79 rows in the enquiries database, 54 of them ours — our form-submissions table for the same window, reviewed by hand against sender addresses on 15 September 2026. The remaining 25 still contain two more of our addresses, two double submissions and several foreign agency pitches.
• 70.22% cart abandonment, 42% "just browsing" — Baymard Institute, an average calculated from 50 studies between 2006 and 2025. It concerns online shops rather than company websites, and that is how we present it here.
• Conversion of 1–10%, e-commerce around 3% — Jakob Nielsen, "Conversion Rates", Nielsen Norman Group, 24 November 2013. The publication does not state how the average was calculated and itself advises against treating the range as a target.
• A 5.1% year-on-year fall — Contentsquare Digital Experience Benchmark 2026, 6,500 sites and 99 billion sessions, comparing the fourth quarter of 2024 with the fourth quarter of 2025.
• 10–20% success at Bing, one third across Microsoft, 89% against 3.1% — Ron Kohavi, Alex Deng, Roger Longbotham, Ya Xu, "Seven Rules of Thumb for Web Site Experimenters", KDD 2014. The quotations come from the downloaded text of the paper.
• What is not here: the three documents this text was built from gave a return on UX investment of 9,400%, conversion lifts of 200% and 300%, a case study with a 180% lift and no client name or before-and-after data, and the story about Barack Obama changing a button from red to green. None of those figures had a source, and the last is a distorted version of a real 2008 test in which an image and the button's wording were changed, not its colour. All were removed, not corrected.