Conversion Rate Optimization Audit: Find the Leak First

The redesigned page shipped in March and the rate came back flat. That is a diagnosis failure, not a design failure, because three of the four leak classes were never design problems. Which one is your biggest drop hiding in?
Somewhere in your funnel there is a step where more people leave than the step before or the step after can explain. Nobody flagged it. Nobody owns it. It has been sitting there since a form field was added in 2023, and it takes a slice of every campaign you run, including the ones you are about to scale. Your analytics will show you the drop and will not tell you why. Your team has opinions, the opinions disagree, and the loudest one wins by default. Which step is it?
The conversion rate optimization audit below gives the loudest opinion in the room something to answer to. Before you touch a page, four counts and one query report will tell you whether the fix belongs on that page at all.
A leak hunt, not a checklist
A 28-point audit checklist is genuinely thorough. It covers landing pages, checkout, mobile usability, analytics, and the rest, and if you work through one honestly you will find real defects. Here is the problem: a checklist is unordered and unweighted by construction. Every point weighs the same, so a mobile tap-target finding outranks an extra-costs finding that sits behind 40 percent of your abandonments.
That is not a small flaw. Baymard Institute's checkout research, averaged across 50 studies and reported in a 2026 compilation, puts cart abandonment at 70.22 percent, and its breakdown of why people leave is the most useful number set you will find on this question. Excluding people who were only browsing: 40 percent said extra costs were too high, 20 percent said delivery was too slow, 19 percent did not trust the site with their card details, and 18 percent were forced to create an account. Four reasons, and they do not belong to the same kind of fix.
Put three facts next to each other and the case stops being an opinion. Most experiments do not return a usable answer. The average US checkout asks for roughly twice the input it needs. The largest abandonment reason is a total that came in higher than expected, which is not a design flaw at all. Add them up and the money in a first audit sits in the findings that are already known and fixable. A checklist cannot make that distinction, because it has no way to weigh one point against another. A classification can, because it attaches a fix type to each finding before anybody schedules a test.
The framing of an audit as a leak hunt is not mine, and one of the better pages in this field already makes it. What that page does not give you is a decision. It tells you where to look, then leaves you holding a drop at the shipping step and a room full of people arguing about whether it is a copy problem, a form problem, a trust problem, or a pricing problem.
Seven of the seventeen organic results I read for this query in November 2026 were agency service pages, gated downloads, or tools that hand back a score. Four were procedures you could run. A CRO audit you can run yourself is hard to sell by the hour, so read most of those pages as sales material.
Step 1: find the biggest drop
Write your conversion path as the steps a person actually takes, not as the pages you wish they visited. Four to seven steps. A website CRO audit that starts with a list of pages has already decided the answer, so start with the path instead. An ecommerce path: session, product view, add to cart, checkout start, order. A B2B path: session, pricing or feature page, form start, form submit, qualified call. If you cannot write the path in one line, the audit will not work, and that is the first finding.
Then count people at each step over one window, 28 days is enough, and keep the source consistent. You do not need a data team for this. A funnel exploration in your analytics tool will produce it, and a spreadsheet of exported events will too. If you are choosing tools rather than using them, the analytics stack options are worth reading first.
Now compute the step rate, which is simply the people at step two divided by the people at step one. Then, and this is the part most teams skip, compute the absolute number of people lost at each step. Rank by people lost, not by the worst percentage.
Take a made-up but realistic path, so the arithmetic is visible: 100,000 sessions, 42,000 product views, 9,800 add-to-cart events, 3,100 checkout starts, 1,850 orders. The step rates are 42, 23.3, 31.6, and 59.7 percent. The entry step loses the most people, 58,000, and ranking it would be a mistake, because that is where every visitor who was never in the market leaves. Start at the first step inside the funnel. Product to cart loses 32,200 people at 23.3 percent, both the worst rate and the biggest inner loss, so here the two measures agree. They often do not. A step losing 60 percent of 200 people is a rounding error next to one losing 25 percent of 40,000, and the small one gets discussed because it has a screenshot.
Keeping this honest takes two habits. First, do not conclude anything from a drop until you have checked that the drop is real. A missing event, a duplicated event, or a redirect that swallows a parameter will draw a cliff on a chart where nothing happened to your visitors. Measurement discipline before conclusions saves a wasted quarter. Second, hold the payment step aside. It is usually the loudest number on the chart and often not the biggest loss, because everyone who reaches payment already decided.
The whole step takes about 30 minutes if your tracking is clean and your access is granted. That time box matters: an audit allowed to expand into a three-week project does not happen at all in most teams. Then pick one step and finish it. Fixing four things at once is how the next measurement becomes unreadable, because if the rate moves you cannot tell which change did it.
Step 2: classify the leak
This is the step that most audits never take. Once you know where the drop sits, sort it into exactly one of four classes. The classes are mutually exclusive on purpose, because each one has a different owner, a different cost, and a different kind of fix.

Class | What the evidence looks like | The test that confirms it | The fix type |
|---|---|---|---|
Intent mismatch | The drop sits at the entry step, or bounce is high on one page while the rest of the site behaves | Entry-step intent check | Change the promise or the destination |
Friction | The drop sits inside a form, a checkout, or a multi-step flow | Form-field audit | Remove effort, not persuasion |
Trust | The drop lands at payment or at account creation, and support tickets mention security, returns, or legitimacy | Read what is visible at that exact step | Put proof where the doubt happens |
Price shock | The drop lands where shipping, tax, or a fee first appears, or where plans get compared | Cost-reveal check | Move the cost signal earlier |
Three of these four are not design problems, which is the reason a redesign so often changes nothing. Fix the layout on a page whose leak is price shock and you have improved a page that was already working.
Intent mismatch. A page ranks for a question, or an ad promises something specific, and the first screen asks for a sales call. The visitor wanted to learn and the page wants a meeting. The class is real and common, and no percentage in circulation measures it honestly, so I will not quote one. The test is structural and five minutes is enough: if the query implies learning and the call to action implies buying, you have identified it. The fix is content work rather than layout work, and it is the page-level job described in website content strategy, which moves slower than a button color and moves more people.
Friction. Effort per step. Too many required inputs, a password rule announced before the value is clear, a coupon field that makes people leave the page to hunt for a code. The test is the form-field audit below, and the fix is subtraction.
Trust. The visitor is willing, then hesitates at the moment money or identity changes hands. In the same Baymard data, 19 percent of shoppers who abandoned a checkout for a reason other than browsing said they did not trust the site with their card details. Proof belongs at that step: a returns policy next to the payment button, a real address, a phone number that answers, security marks somebody can verify. Not on the homepage, where nobody is worried yet.
Price shock. In that same survey, 40 percent of non-browsing abandonments cited extra costs, the largest single bucket on this list, and 12 percent said they could not see the total cost up front. Discounting is the reflex here, and free delivery is the expensive version of it. The cheaper fix is moving the cost signal earlier, to the product page or the cart, so the total arrives while the person is still deciding.
The tie-break rule. Some findings sit in two classes, and the honest example is forced account creation: 18 percent of abandoners named it, and it reads as both distrust and friction. Decide by asking which fix would move the number. Guest checkout is a friction fix, so the class is friction. Applied out loud, that rule ends most arguments about labels before they start.
None of this step needs a CRO platform. It runs on funnel counts, a phone, and a query report. A sane CRO audit buys tooling from the class you found rather than from a feature list. If you are at that point, the CRO tooling landscape is worth a read before you pay for anything.
The two checks that find most of the money
Two checks, in this order. Between them they account for most of what a first audit turns up on a normal commercial site, and neither needs a tool you do not already have. Sequenced: five minutes writing the path, ten pulling counts and ranking the inner steps, ten on the entry-step check, five counting fields. That is the whole first pass, and it ends with one class and one fix rather than a report.
Start with the entry-step intent check. Take your three biggest entry pages by sessions in the same 28 days. For each one, pull the query that most often lands there from your search console, or the ad headline if the traffic is paid. Open the page on a phone, read only the first screen, then read the call to action. If the promise on that screen and the action in the button both match what the query asked for, the page passes. If the query implies research and the button asks for a call, you have an intent mismatch, identified without a single number. Ten minutes for three pages, and on most sites it produces at least one hit.
The second check is the form-field audit. Count the required inputs on your main conversion form, then count them for the whole checkout if you have one. Baymard's benchmark work puts the ideal at 12 to 14 form elements, which is 7 to 8 fields, against an average US checkout of 23.48 elements. Your checkout is probably asking for roughly twice the input it needs.

Resist the obvious move. Every checklist in this field says cut fields, and the most useful data point I have found says the opposite happens when you cut the wrong ones. In Venture Harbour's 2026 compilation of five studies, Marketo reduced fields on its sign-up form and conversions rose progressively. Michael Aagaard removed fields and conversions dropped 14 percent, because he removed the ones people wanted to interact with; restoring them and fixing their labels produced a 19.21 percent increase. Does this field carry the decision, or does it carry your reporting? Price, delivery date, fit, and quantity carry the decision. Company size, referral source, and a phone number nobody will call carry your CRM.
One warning about the numbers you will read elsewhere. Baymard's older figure of 11.3 form fields has been superseded by 14.88, so a 2026 checklist still quoting 11.3 is a copied list rather than a current audit.
The check produces two changes almost every time. Cut the fields that serve your reporting, and move the form promise above the fields, so the person knows what happens after they click. The one clean before and after I can point to is a B2B SaaS page whose demo requests moved from 2.1 percent to 3.0 percent over six weeks after the form was cut to five required inputs and the promise moved above the fields. Its author calls it not a clean lab result, which is the right label: one reported case, not a benchmark.
The other half of the problem is that most tests do not return a usable answer. Two 2026 compilations of A/B testing results disagree on the exact share of experiments that produce a significant winner, and they measure different things, so the decimal is not worth quoting. The direction is consistent in both, and it matches SearchPilot's benchmark of 75 percent inconclusive SEO tests across 2022 and 2023. Spending a quarter to prove a fix you already know is right is a poor trade. Fix what the evidence already identifies, and save the tests for changes where two reasonable people would genuinely disagree.
The output of the afternoon is one page: biggest drop, the class, the evidence behind it, and the fix you are about to make. Write it down even if nobody asked, because a finding without a recorded number rarely survives its first budget conversation, and six weeks later that note is what stops the fix from being quietly reverted. If you report to a client rather than a manager, reporting what the audit found covers the version that survives a review.
What an audit cannot fix
An audit cannot repair two of the leaks you will find, and most audits leave them out. Both remove the fix from your hands, which is precisely why you rarely see them named.
The first is traffic quality. If the people arriving were never going to buy, no page change converts them. The audit will show a flat, unremarkable rate across every step, you will keep redesigning, and the number will not move. The tell is in the segmentation: when one source converts at a fraction of another over a long window and both land on the same pages, the leak is upstream in the targeting. An audit measures what happens after arrival. It has no opinion about who arrived.
The second is the offer. If the price, the promise, or the market is wrong, a better page only makes the wrong offer clearer, faster. Baymard's own modeling, reported in a 2026 CRO statistics compilation, puts a 35.26 percent upside on checkout design alone, from scoring 60 top-grossing sites against 134 usability guidelines, and that is modeled potential rather than an observed test result. Read it as a ceiling on execution quality, not a rescue for a product nobody wants at a price nobody will pay.
Knowing which of the two you are looking at changes the recommendation. If the answer is traffic or offer, the honest output is a short memo saying so, and the next conversation is with whoever owns acquisition or pricing.
When to bring someone in
Run this conversion rate optimization audit yourself first, because it costs an afternoon and tells you whether you even have a page-level problem. Bring in outside help for the cases where the loop cannot close alone: the leak sits inside a checkout or a platform you cannot instrument, the fix needs engineering capacity you do not have this quarter, or the number at stake is large enough that being wrong costs more than being advised. There is also the case of running the same loop three times and watching the rate stay flat, which usually means the classification was wrong. Bring the funnel numbers and the class you assigned. Anyone worth bringing in will start by questioning that classification, and it is the conversation you want.

Run the first pass before you change a page
Take one funnel, one 28-day window, and one afternoon, then write down the biggest drop and the class you assigned to it before you edit anything.