CRO is not a marketing function, and putting it there caps what it can ever return
2026-09-21
Conversion rate optimisation is four things held together at once, and only one of them belongs to marketing. That placement decides the ceiling of the work before any of the results.
- Ecommerce
- CRO
- Leadership
I have lost count of the CRO job specs I have read in this market that sit under a marketing director. The title varies. Conversion Rate Optimisation Manager, Digital Experience Manager, sometimes Growth Manager, occasionally the work is not a job at all and has simply been handed to whoever runs paid media or whoever runs trading, on top of everything else they already do.
The placement looks administrative. It is the single decision that sets the ceiling on what the function can return, and it sets it long before anyone runs a test.
So it is worth being precise about what conversion rate optimisation actually consists of, because the reason marketing cannot own it becomes obvious once the parts are laid out.
Part one: what CRO is made of
There are four layers, and they function only when held together.
The smell from the advertisement to the landing page. According to information foraging theory, which was developed by Peter Pirolli and Stuart Card at Xerox PARC, people navigate information in a way similar to how a forager works through a patch — by following cues and deciding whether the reward is worth staying, and then leaving when the cue becomes weaker. The term 'information scent' refers to the cues that a person uses to decide if a particular path will lead them to what they are looking for. An advertisement promises a category and a price position. The landing page either fulfils that promise on its first screen or it doesn't. If it doesn't, the visit comes to an end before any of the interface elements have got a chance to function.
This is the part that sits closest to marketing, and even here the ownership is split. The copy is marketing's. The page the copy lands on belongs to whoever runs the site. In most organisations, those are different people with different targets and different release cycles, which is why the mismatch survives for months.
The interaction layer. This is the part that gets waved through as technical, and it is where the money is.
Start with the obvious measures, since they are at least well defined. Google's Interaction to Next Paint threshold is 200 milliseconds at the 75th percentile of real user interactions, meaning three-quarters of every tap and click on the page must resolve inside that window. INP replaced First Input Delay in March 2024, and it is now the Core Web Vital that sites fail most frequently, because it measures every interaction across the life of the page rather than the first one. A single third-party script added to a template can push an origin from passing to failing without a line of the retailer's own code changing.
Below the measurable sits everything that decides how much work the customer has to do, and this is the higher cost.
Take the number of actions. Baymard Institute's checkout benchmark finds the average US checkout displays 23.48 form elements by default, against an ideal flow they put at 12 to 14. Counting only fields rather than checkboxes and drop-downs, the average is 14.88. Roughly one in five shoppers abandons a checkout because it is too long or too complicated, which is a rational response to being asked for twice what the order requires.
Take where information sits. Baymard's product page testing found participants repeatedly overlooking core product content when it could only be reached through horizontal tabs, a pattern still used by around 28% of sites. On mobile, 26% of sites push some product page content onto subpages, and users miss it entirely. Truncating the additional images in a gallery causes 50% to 80% of users never to look at them. Their testing on free shipping display found 32% of sites present the offer in a way that is prone to being overlooked, which is a remarkable thing to do with the single most persuasive fact on the page.
The accordion should have its own line since it is the pattern most frequently cited in design reviews as a neat method of handling a long page. Progressive disclosure does help to reduce visual clutter, but it also completely depends on the trigger indicator carrying out its role, and in cases where it fails to do so, the user believes that the content is not available. The way Baymard puts this is the point I return to again and again: if the content isn't present in the user's mind, then the user assumes that the site doesn't have it and leaves to find somewhere that does. It is the delivery information, the return terms, the materials information, the sizing guidance and the specification tables that are most often collapsed, and it is precisely this kind of content that determines whether a purchase takes place.
Then we come to the cases where decisions are forced. Mandatory account creation is responsible for about 19% of abandonments; the customer is at the checkout with a card in hand and is asked to establish a relationship first. The phenomenon is accurately described by the concept of psychological reactance: when people feel that their freedom of action has been restricted, they are motivated to restore that freedom, and the easiest way of doing so is to leave.
None of this is a design taste question. Each of these is a lever with a measurable effect on revenue per visitor, and the person who holds them needs to understand the commercial consequence of each one.
Experiment discipline. The third layer is whether the numbers a team reports are real.
Underpowered tests do not produce smaller effects. They produce exaggerated ones. Andrew Gelman and John Carlin set this out as Type M, or magnitude, error in Perspectives on Psychological Science in 2014. I flag the age of that paper deliberately, because the standing rule in my own writing is that anything older than four years gets checked against newer work rather than quoted for authority. In this case the statistics have not moved, and the problem has got worse as testing tools have made it easier to run experiments without anyone calculating the sample required.
Include the sample ratio mismatch, a circumstance that renders a test invalid when traffic is split unevenly because of a bot filter or a caching rule, and also include the winner's curse, in which the tests that clear the significance threshold do so systematically because they were just lucky. As a result, a team might release a number of winning variations and end the year with no change in conversion rate. I have seen this exact situation occur, and the uncomfortable thing is that the team is not lying; they have carefully measured the noise and have reported it honestly.
Post-purchase communication. The fourth layer is the one most often left out of the definition entirely.
Expectancy disconfirmation is the mechanism. Satisfaction is a function of the gap between what was expected and what arrived, rather than of the quality of what arrived in isolation. The delivery estimate shown at checkout becomes the reference point against which every subsequent event is judged. A parcel arriving on Thursday is a good outcome if Friday was promised and a failure if Tuesday was. The confirmation email, the dispatch notice, the tracking link and the returns instruction all sit inside the conversion function, because they set the probability of the second order.
Part two: where the experience actually comes apart
The failures I see most often are not exotic. They are the same handful, and they all live on the seams between teams.
Costs that appear late. Extra costs at checkout, meaning delivery, taxes and handling charges that were not visible earlier, are the largest documented reason for abandonment in Baymard's survey work, at around 39% to 40% of shoppers who abandoned for a reason other than browsing. The mechanism is anchoring. The basket total the customer saw becomes the reference price, and every pound added afterwards is evaluated as a loss against it rather than as part of the price. A delivery charge of £4.95 disclosed on the product page is a fact. The same £4.95 revealed on the payment step is a betrayal, and the customer prices it accordingly.
A promise regarding delivery that becomes increasingly vague. The product page states three to five working days, the basket page mentions standard delivery, and it is only at checkout that a specific date is displayed. By this stage, the customer has been given three different descriptions to keep in mind, each one less precise than the one that came before, exactly when certainty is what they are looking for. The reason why the more vague version fails is that information which is easy to process is regarded as more likely to be true, and a specific date is easier to process than a range of days which the customer has to convert into a calendar on their own.
Urgency language that only exists in one place. A campaign runs on scarcity and time pressure in paid social. The site says nothing of the kind. Persuasion knowledge, the model set out by Friestad and Wright, describes how consumers develop and deploy a working theory of how they are being sold to. Once that theory activates, it applies to everything on the page rather than to the one element that triggered it. Stock counters, review counts and the delivery promise all start being read as sales techniques. The cost lands on the elements that were telling the truth.
Product content which fails to survive the category page consists of a price and an image listed. The product page, on the other hand, displays a price, a promotion, a bundle offer and a member price. The customer establishes a reference point from the list and then has to check three different prices against it. Each of these comparisons involves cognitive effort, which competes with the decision to buy.
A returns policy that is generous on the landing page and conditional at the bottom of the product page. This is the one that does the most damage to lifetime value, because the customer discovers the gap after they have paid.
Look at what the five cases have in common: each one involves a handover between two teams, each team is carrying out something that can be justified within its own area of responsibility, and no single party is responsible for the entire sequence that the customer actually follows.
Part three: why marketing cannot own this
There are two reasons, and both of them are of a structural nature and not a criticism of anyone's abilities.
It's a matter of competence: if you want to optimise a funnel, you have to look at the stock positions, understand how the delivery promise is arrived at and what it costs to move it, know which payment methods have which authorisation rates and what effect the fraud rules are having on genuine orders, and also understand the returns economics for each category on the site. That isn't something that can be covered in a marketer's training; it's a completely different profession. Asking a paid media specialist to take on those variables is equivalent to asking them to act as a merchant on the side, and the more experienced ones will tell you so.
The span of control is such that marketing has control over the start of the session and an increasingly smaller portion of what takes place afterwards. In the case where the limiting factor is stock availability, the part of the CRO programme that is under marketing's control will keep testing different headline options and various button placements until it has used up all its ideas, and the conversion rate will remain unchanged since the customers who abandoned the process weren't going to have been convinced by the headline; they had abandoned it because the size they wanted was not available, or because the delivery date didn't suit them, or because the payment method they normally use wasn't offered.
It isn't a debate concerning effort; it's a matter of arithmetic. A function whose authority ends at the landing page is capable of identifying only local optima, and in a funnel of such length the value of these local optima is only a fraction of the value of the constraint.
And then the target problem. Goodhart's law is usually quoted as a joke about metrics, and it describes something real here. A team measured on cost per acquisition and traffic volume will optimise cost per acquisition and traffic volume. Both can improve while contribution margin per session falls, because the cheapest acquired traffic is frequently the least valuable, and because discount led conversion improves the conversion rate and destroys the margin it was supposed to defend. I have watched a team deliver a genuinely excellent quarter against its own targets while the business made less money than it had the quarter before. Nobody did anything wrong. The targets were wrong.
Part four: why the ecommerce leader is the only seat that fits
In most organisational structures, the position of Director or Head of Ecommerce is the only one which has a span equal to that of the funnel. That point is the main one, and it is worth going into rather than just stating it.
To identify the true constraint, it is necessary to consider the entire sequence. In the case where 8 per cent of the sessions which reach the product page result in a bounce because the size is out of stock, the solution should involve a discussion about buying and allocation rather than carrying out a test. If customers drop away at the delivery stage during the checkout process, then the solution might be to enter into a carrier contract. When authorisation rates fell by two points following a change to a fraud rule, the remedy would lie with the payments team, and the marketing team would never come across it since, from their point of view, the session just ended. The ecommerce leader is the individual who can see all three of these signals within the same week and be able to rank them by value.
Ranking is the skill that matters. A CRO function that runs a queue of tests in the order they were requested is a service desk. A CRO function that puts its testing capacity on the one part of the funnel where the constraint currently sits is a commercial instrument, and deciding where that is requires authority over merchandising, operations and payments as well as over the site.
That is also the reason why the metric must be altered. Conversion rate alone is a weak indicator. Revenue per visitor shows the balance between conversion and the size of the order. Contribution margin per session shows the effect of the discount and the return cost. A manager who is responsible for the entire process can be judged by those figures in an honest way, whereas a team that is in charge of the landing page cannot be, and it would be unfair to ask them to be.
If you're carrying out a recruitment in this case, the question I'd put to the candidate isn't one about the tools they use or their test velocity. Instead, I'd ask them to describe a test which they decided not to carry out and the reason why. The way they answer reveals whether or not they realise that the scarce resource is actually traffic and attention, not ideas.
And if you are a Director or Head of Ecommerce reading a CRO plan that only contains page-level experiments, the plan is telling you where your organisation's authority stops. That is useful information, and it is usually not the information the plan was meant to convey.