What Is Conversion Rate Optimization?
Benchmark your rate against ten published sectors, see what checkouts leak in order, and find out how often a tested idea wins.
Conversion rate optimization is the work of getting more of the people already on your site to do the thing the site exists for: buy something, fill in a form, sign up, place an order. It is the half of marketing arithmetic that gets the least attention, because buying more visitors feels like progress and fixing the page they land on does not.
Microsoft's experimentation team published the number that should shape how you plan the work: only about "1/3 of ideas improve the metrics they were designed to improve" across its own experiments.
What Conversion Rate Optimization Is
Conversion rate optimization (CRO) is a loop, not a project: decide which action counts, measure how many visitors take it, find where the others leave, change one thing, and check whether the number moved. Then do it again.
A conversion is any action you decided in advance was worth counting, such as an order, a filled-in form, a tapped phone number or a sign-up. The word means nothing until you pick one, which is why an honest process starts there and not with a list of changes to try.
The work is diagnostic before it is creative. You are hunting for the step where the largest number of people who could have converted did not, and for the reason they stopped, which is more often a cost, a doubt or an obstacle than a color.
What CRO Is Not
Three things carry the name and are something else.
It is not traffic work. More visitors usually raise the top and the bottom of the fraction together, so the rate sits roughly where it was. Traffic and conversion are separate problems with separate fixes, and buying more of the first to solve the second is the expensive way to find that out.
It is not a redesign. A redesign changes everything at once, which makes the result unreadable: you know the number moved and you cannot say which of forty changes moved it. If it goes down, you have nothing to roll back to except the old site.
It is not a list of tips. A tip is somebody else's winning test, run on somebody else's audience, on a page you have not seen. The base rate for ideas like that has been measured and published, and it is lower than the confidence of the lists suggests.
What a Conversion Rate Is, and How to Calculate It
A conversion rate is the share of visits that ended in the action you decided to count. You calculate it by dividing conversions by visits and multiplying by a hundred.
conversion rate = conversions / sessions x 100
12 orders / 400 sessions x 100 = 3% conversion rate
The arithmetic is the easy part. The denominator is where published benchmarks stop agreeing with each other, and where two reports on one site stop agreeing too.
Sessions, users and unique visitors give three different answers. One person who visits on Monday, returns on Wednesday and buys on Friday is three sessions and one user. In that example, dividing by sessions gives a third of the rate you get dividing by people. Neither is wrong. Comparing one against the other is.
This is why a published figure is worth little without its definition beside it. IRP Commerce prints its own: conversion rate is transactions divided by sessions, multiplied by a hundred, and it labels the result a session conversion rate. That is a number you can compare against, because you know what it counted.
Conversion Rate vs Click-Through Rate
A click-through rate (CTR) is the share of people who were shown something and clicked it. A conversion rate (CVR) is the share of people who arrived and completed the action. Different denominators, and different stages of the same journey.
The pair is worth holding apart because they move independently, and a rise in one can cause a fall in the other. A sharper ad headline lifts the click-through rate and brings in people the page was never written for, and the conversion rate drops while the business gets better. Read either one on its own and the conclusion you draw about the other can be exactly backwards.
Is Your Conversion Rate Good?
Whether your conversion rate is good depends on what you sell, and the gap between categories is wider than the headline averages suggest. IRP Commerce publishes monthly conversion rates from live trading data across its platform for Great Britain, Northern Ireland and Ireland. In July 2026 the all-market average was 2.26%, and the ten published sectors ran from 0.55% to 5.23%.
| Sector | Conversion rate, July 2026 |
|---|---|
| Arts and Crafts | 5.23% |
| Health and Wellbeing | 3.57% |
| Kitchen & Home Appliances | 3.34% |
| Pet Care | 2.95% |
| Sports and Recreation | 2.12% |
| Cars and Motorcycling | 1.82% |
| Fashion Clothing & Accessories | 1.81% |
| Toys, Games & Collectables | 1.72% |
| Food & Drink | 1.47% |
| Baby & Child | 0.55% |
| All markets | 2.26% |
Source: IRP Commerce, eCommerce Market Data, July 2026. Session conversion rate, business-to-consumer trading on one platform.
Read the table the useful way. A store selling baby products at 1% is doing well against its own sector and badly against a headline average built mostly from other people's categories. A craft store at 3% is underperforming while looking respectable. The average is the least informative row in the table.
Four caveats, because a benchmark used carelessly is worse than none. It is one platform, so the merchant mix is not the whole market. It is one month, and the same table a year earlier put the average at 1.94%. It is consumer retail, so a site selling to businesses over a long cycle is not comparable to it. And it counts sessions, so a site measuring people will read high against it for reasons that have nothing to do with the site.
So is 20% good? On a whole-site sessions denominator, a 20% rate is almost never a great site. It sits outside the entire published range in the table, and the ordinary explanation is that the denominator is wrong. A filtered view, internal traffic, a goal firing on a page load instead of a purchase, or bot sessions stripped from one side of the fraction and not the other. On a narrow denominator the same number is unremarkable. One landing page, reached from one email to a list that already knows you, converting at 20% is a Tuesday.
Why Conversion Rate Optimization Is Worth Doing
Conversion rate optimization is worth doing because it changes the price of every customer you already pay for. Acquisition spend buys visits. Conversion decides how many of those visits become customers. Move the second and the same spend divides across more people, which improves the economics of a channel without spending more on it.
$1,000 spend, 1% conversion, 10 customers = $100 each
$1,000 spend, 2% conversion, 20 customers = $50 each
Nothing about the traffic changed in that pair. The doubling is the whole return, and it repeats every month you keep it, on every channel feeding the page. Which channel to buy in the first place is a different decision with different inputs, and it is worked through in the digital marketing guide.
The honest counterweight is that a sector's conversion rate does not settle what its customers cost. In the same IRP data for July 2026, Arts and Crafts had the highest conversion rate at 5.23% and the lowest marketing cost as a share of revenue at 1.56%. Baby & Child sat at the other end, lowest rate at 0.55% and highest cost at 12.29%. That pairing looks like a law until you read the middle of the table, where Health and Wellbeing converts at 3.57% and still spends 11.93%. The two numbers move together sometimes and not dependably, which is an argument for measuring your own pair instead of inheriting your sector's.
Why Visitors Leave, in Order
They leave over money more than over design. Baymard Institute keeps a running average of documented cart abandonment across 50 separate studies published between 2006 and 2025, and that average is 70.22%. The studies themselves range from 55.00% to 84.27%, which is worth knowing before anyone treats a single vendor's figure as the truth.
A large part of that is not a problem you can fix. Baymard's own survey found 42% of US online shoppers have abandoned a cart because they were browsing and not ready to buy. Set that group aside and what remains is ranked, and specific.
| Why the order was abandoned | Share of US online shoppers |
|---|---|
| Extra costs too high (shipping, tax, fees) | 40% |
| Delivery was too slow | 20% |
| Did not trust the site with card information | 19% |
| The site wanted an account created | 18% |
| Too long or complicated checkout process | 17% |
| Website had errors or crashed | 17% |
| Returns policy was not satisfactory | 13% |
| Could not see or calculate total order cost up front | 12% |
| The credit card was declined | 10% |
| There were not enough payment methods | 9% |
| Did not know | 7% |
Source: Baymard Institute, Cart Abandonment Rate Statistics, last updated September 22, 2025. Reasons excluding the browsing group. Shoppers could give more than one reason, so the shares do not total 100%.
The top reason is not a design problem. Extra costs, at 40%, is a pricing and shipping decision that somebody made in a spreadsheet, and no amount of button testing reaches it. The second, at 20%, is a logistics contract. Together with the 12% who could not see the total up front, three of the eleven ranked reasons are commercial decisions before they are interface ones.
The interface reasons underneath are real and they are cheaper to fix. Baymard's checkout testing found an ideal flow can be as short as 12 to 14 form elements, while the average US checkout shows 23.48 by default. Across the large sites it tests, it estimates the average one can gain a 35.26% increase in conversion rate through better checkout design.
Two things follow for anyone deciding where to start. Read your own drop-off before you copy the ranking, because your site is one site and the list is an average. And when the biggest leak turns out to be a shipping charge revealed on the final step, the fix belongs to whoever sets shipping charges, not to whoever owns the template.
The Same Read on a Site That Does Not Sell Online
A site with no cart has the same shape of problem in different clothes. The equivalent of an unexpected shipping cost is a price that appears only after a call. The equivalent of forced account creation is a form asking for a phone number before it has earned one. The equivalent of a slow delivery promise is a response time nobody states.
The method transfers cleanly: list every moment the visitor is asked to give something up, in order, and count how many of them arrive before you have given anything. On a business site, that count usually explains the form abandonment without any tooling at all.
The Six Steps of a CRO Process
A CRO strategy is a written process: what you will look at, what you will change, and how you will know whether it worked. Six steps cover it, and skipping one of the first three is the usual reason an effort stalls at the fourth.
- Pick the conversion that counts. One primary action per page, decided before you open an analytics tool. Secondary actions get counted, never optimized against.
- Find where people drop off. Walk the path from arrival to conversion and measure the exits at each step. You are looking for the largest fall between two adjacent steps, not the lowest number overall.
- Gather the behavior data. Numbers say where. People say why. Session recordings, heatmaps, an on-page survey with one question, and the last twenty support tickets will each tell you something the funnel cannot.
- Write the hypothesis. One sentence naming the change, the metric it should move, and the reason you think it will.
- Run the test, if your traffic can carry one. One variable, a run length fixed in advance, and no peeking at the result to decide when to stop.
- Decide and iterate. Ship it, drop it, or run it again bigger. Write down what you learned either way, because a losing test that is recorded is cheaper than the same idea returning next quarter.
Writing a Hypothesis That Can Be Wrong
A hypothesis that cannot fail is a plan wearing a lab coat. "Improve the checkout" cannot be wrong, so it cannot teach you anything. "Showing the shipping cost on the cart page will reduce checkout abandonment, because our exit survey says people are surprised by it at the last step" can be wrong, and that is what makes running it worth the traffic.
The test is mechanical. Write the sentence, then ask what result would make you say the idea was mistaken. If no result would, rewrite it until one does.
Can You Even Run a Test?
Whether you can even run a valid test is a traffic question, and it gets answered before any tool is chosen. Microsoft's experimentation team put its general guidance in print. Teams working on products "with thousands to tens of thousands of users (our general guidance is at least thousands of active users) are typically looking for larger effects, which are easier to detect than the small effects that large sites worry about".
The reason is arithmetic and it is unforgiving. In the same paper: to increase the sensitivity of an experiment by a factor of 10, say from a 5% difference to a 0.5% difference, you need 100 times more users. Small sites can only detect large effects. Chasing a small improvement on a few hundred visits a month is not a slow test, it is an impossible one, and the result you eventually read is noise.
Can you even run a test?
Put in the traffic reaching the page you want to test and the rate it converts at now. The answer is the smallest improvement a four-week test could reliably tell apart from noise.
17,240
2.0%
About 3,978 visits a week, split evenly between the control and one variant.
Smallest improvement a four-week test could detect
31.1%
That is a rise of 0.62 percentage points, from 2.00% to 2.62%. Anything smaller than that sits inside the noise, and a dashboard showing it would be showing you nothing.
To detect a 10% relative improvement instead, the same test would need to run for 39 weeks.
Testable, but only for large changes. Save the tool for whole-page rewrites and pricing or offer changes, not button colors.
Microsoft's experimentation team puts the entry point plainly: “our general guidance is at least thousands of active users”. And on why small sites are stuck with big changes: to increase sensitivity by a factor of 10, “you need 102 = 100 times more users”.
Method, so you can check it. Two proportions, two-sided, 95% confidence, 80% power, one control and one variant at an even split, visitors counted once. Sample per variant is 15.698 × p(1 − p) ÷ d², where p is the current rate and d is the absolute difference; the figures above invert it for d. Real tests also lose power to repeat visitors, seasonality and weekday effects, so treat every number here as the optimistic end.
Below the threshold, the honest answer is to stop testing and start reading. Watch recordings, run an exit survey, read the support tickets, and make the obvious fixes without measuring them. Removing a broken form field does not need a control group. The cost of being wrong is small and the cost of waiting for statistical proof you cannot buy is a year.
The same paper is worth reading on how quiet the failure can be. Its authors looked at a published result from Etsy showing that a 200 millisecond delay did not matter, and said the more likely explanation was that the experiment lacked the statistical power to detect the difference. A study that finds nothing and a study that lacked the ability to find anything look identical from the outside.
What Statistical Significance Actually Buys You
Statistical significance answers one narrow question: if the change did nothing at all, how surprising would this result be? It does not tell you the change is important, that it will hold next month, or that the effect is the size the dashboard printed.
It also gets weaker every time you look. Microsoft's team set out the mechanism: a team trying five treatments sees its 2.5% false positive rate grow to 12%, and six iterations of five-treatment experiments give more than a 50% chance of a positive significant result by chance alone. Their protection is to demand smaller p-values for those projects and to re-run the winner once before believing it. One variable and one final run is the small-site version of the same discipline.
Most Tests Lose, and That Is the Normal Result
Most changes you are sure about will do nothing. Microsoft's experimentation paper reports that among well-designed and executed experiments aimed at a key metric, only about one third were successful at improving it. A team that launches 10 ideas without measuring them, the same paper estimates, may have about 1/3 good, 1/3 flat and 1/3 negative. Features that fail to improve their metric or hurt it account for 66% of experiments there.
It is not a Microsoft problem. The same paper records that at Amazon, where evaluating every new feature is common practice, the success rate is below 50%. It reaches outside software too. QualPro, a consultancy the paper cites, tested 150,000 business improvement ideas over 22 years and reported that 75 percent of important business decisions and improvement ideas either had no impact on performance or hurt it.
The most useful thing in the paper is not a rate at all. Its authors ran a survey with eight A/B tests and offered a shirt to anyone who could pick six correctly. Over 200 people entered. Six of them got five right, the average score was 2.3, and not one shirt was handed out.
Hold that against how a tip list is written. Every item on one is presented as a thing that works, while the only published base rate for tested ideas puts about two in three of them at flat or negative. That is the argument for testing, and it is also the argument for testing the ideas you believe in instead of the ones you collected.
The Pages Worth Optimizing First
The pages worth optimizing first are the ones where a decision is being made, not the ones with the most traffic. Every page below has exactly one job, and each fails in a way you will recognize.
| Page | Its one job | How it usually fails |
|---|---|---|
| Homepage | Send each kind of visitor to the right next page | Explains everything to everyone, so it explains nothing to anyone |
| Landing page | Deliver on the promise made by the ad or email that sent the visitor | Says something different from the ad, so the visitor re-reads instead of acting |
| Product page | Answer the last three questions before a purchase | Answers the first three and hides shipping, returns and total cost |
| Pricing page | Let a buyer place themselves in a tier without help | Withholds the number, turning a ready buyer into a form to chase |
| Form | Collect the fewest fields that let the work begin | Asks for everything the database can store |
| Checkout | Move a decided buyer through with no new information | Adds costs late and demands an account first |
| Call to action | Say what happens next, in the reader's words | Says "Submit" or "Learn more", which promise nothing |
| Navigation | Let someone find the thing they came for | Built around the org chart instead of the questions |
Named practices fall out of the table. Show the total cost early. Cut form fields to what the next step needs. Offer a guest path through checkout. Write button labels as the outcome, not the mechanism. Match the landing page headline to the ad that earned the click. Put the answers to the last three objections on the page where the decision happens.
What CRO Tools Do
CRO tools do four separate jobs, and each category answers a question the others cannot. Knowing which question you have is what stops a trial turning into a subscription nobody opens.
| Category | The question it answers |
|---|---|
| Web analytics | Where do people go, and at which step do they stop? |
| Behavior tools (heatmaps, session recordings) | What happened on the page itself? |
| Voice of customer (surveys, exit polls, support tickets) | Why did they stop? |
| Testing platforms | Did the change I made do anything real? |
Order matters more than brand. Analytics tells you the room, behavior tools tell you the moment, customer voice tells you the reason, and a testing platform is worth paying for only once the first three have produced a hypothesis and your traffic can carry a valid test. Buying the fourth first is how a year of budget turns into an unusable result.
CRO and SEO
Search engine optimization (SEO) decides how many people arrive. CRO decides how many of them do the thing. They are measured on different denominators and they fail independently, which is why a page can rank first and earn nothing, or convert brilliantly for the four people who find it.
They meet in one place: intent. A page that matches what the searcher meant ranks better and converts better, because both outcomes reward the same thing. That is the overlap, and it is smaller than the amount written about it. How ranking itself works belongs to search engine optimization; what happens after the click is the subject here.
Conversion Rate Optimization Best Practices
The best practices worth keeping are decisions, not tips, and each one below traces back to a number or a mechanism already set out.
- Name the conversion before opening a tool. One primary action per page. Everything else is context.
- Check the denominator before comparing to anyone. Sessions and people give different answers, and published benchmarks rarely say which they used.
- Compare against your sector, not against a headline average. The published range for a single month runs from 0.55% to 5.23%.
- Fix the money surprises before the interface. Extra costs is the largest ranked reason for abandonment, at 40%.
- Show the total cost early, and allow a guest checkout. Those two answer 12% and 18% of the ranked reasons on their own.
- Change one thing per test. Multiple treatments inflate the false positive rate, from 2.5% to 12% at five of them.
- Fix the run length before starting, and do not stop early. A test stopped when it looks good is a test that measured your patience.
- Expect two ideas in three to do nothing or make things worse. Plan the queue around that rate instead of around the winner you hope for.
- Below a few thousand active users, stop testing and start watching. Recordings and surveys beat an underpowered experiment.
- Write down the losers. An undocumented failed test is an idea that comes back next quarter with a new sponsor.
Common Questions About Conversion Rate Optimization
Four questions about conversion rate optimization come up more than any others. Three of them are definitional, which is a fair sign of who is asking: people who were handed the term recently and now have to explain it to somebody else in one sentence.
What Is the Difference Between SEO and CRO?
SEO works on getting found. CRO works on what happens next. One is measured against searches and impressions, the other against visits that turned into an action, so a change can improve one and leave the other where it was. They also fail independently, which is why a page can hold the top result and earn nothing.
What Is the Difference Between CVR and CTR?
CTR counts clicks against the people who saw something. CVR counts completed actions against the people who arrived. The first belongs to whatever sent the visitor, an ad, an email or a search result. The second belongs to the page that received them. Raising one can lower the other, because a bolder promise brings in people the page was never written for.
Is a 20% Conversion Rate Good?
For a whole site measured by sessions, no published sector average in July 2026 came anywhere near 20%, so the number is a reason to audit the measurement, not to celebrate it. For one landing page reached from one warm email, 20% is ordinary. The question to settle first is what sits underneath the line: which visits were counted, and which action was counted as a conversion.
What Are Some Examples of Conversion Rate Optimization Practices?
Deciding which single action a page is for. Reading a drop-off report before changing anything. Watching session recordings to find where people hesitate. Removing a cost surprise from the final step. Testing one variable at a time, with the run length fixed in advance. Recording the tests that lost. Every one of them is a hypothesis until your own data supports it, which is the difference between a practice and a tip.
The Short Version
Conversion rate optimization is the discipline of moving the second number in the marketing equation instead of the first. It starts with picking one action, understanding your denominator, and reading your own drop-off before copying anyone's list.
The published evidence points the same way twice. Where people leave, the largest single fixable reason is what the site charges them and when it tells them, not how it looks. And when you finally test a change, the odds are close to one in three that it improves what you aimed it at.
Two questions settle what to do on Monday. Do you have enough traffic for a result to mean anything, and if not, what would you change today if you had to decide without a test? Both answers are useful, and only one of them costs anything.
Sources
- IRP Commerce, eCommerce Market Data, July 2026, the all-market and ten sector session conversion rates, the year-earlier average, and the cost-per-acquisition figures. First-party trading data from merchants on the IRP platform in Great Britain, Northern Ireland and Ireland.
- Baymard Institute, Cart Abandonment Rate Statistics, last updated September 22, 2025, the 70.22% average across 50 documented studies from 2006 to 2025, the ranked reasons for abandonment, the checkout form element counts, and the 35.26% figure.
- Kohavi, Crook, Longbotham, Frasca, Henne, Lavista Ferres and Melamed, Online Experimentation at Microsoft, Microsoft ThinkWeek 2009, the one third estimate, the 66% figure, the Amazon success rate, the QualPro survey of 150,000 ideas, and the eight-test prediction challenge.
- Kohavi, Deng, Frasca, Walker, Xu and Pohlmann, Online Controlled Experiments at Large Scale, KDD 2013, the guidance on the number of active users an experiment needs, the sensitivity arithmetic, the Etsy statistical power diagnosis, and the multiple-treatment false positive rates.