Why Industry Averages Mislead (and How to Use Them Anyway)

Industry averages are the most widely used and least reliable number in eCommerce reporting. They get quoted in board decks, used to set targets, and treated as a standard a store is passing or failing, and almost none of that survives contact with how the figure was built. The useful move is not to discard them. It is to know precisely which distortion is operating, because each one has a different size and a different fix. Last updated: September 2026.

Omniconvert has measured conversion behaviour across the CROBenchmark dataset of 7,000+ websites in 15+ industries, against 248+ audit criteria, over 13 years in eCommerce, and has run Omniconvert Explore across 70,000+ experiments. The finding that matters here is the same one that makes benchmark publishing awkward: the spread within a category is consistently wider than the distance between categories, which means the category label explains far less about a store than the label implies.

This article is about the statistics rather than the decision. If what you actually need is which signal to act on when a benchmark and your own history disagree, our companion piece on benchmarks versus your own baseline answers that directly, and for the category figures themselves, start with a good conversion rate.

Further detail sits in two stores with the same CVR aren't equal.

What an industry average actually is

Nothing about an average is dishonest. The arithmetic is correct, and a competently published benchmark is measuring real stores. The problem is one of transfer: the figure describes a population, and it gets used to make a claim about an individual.

That transfer is only valid when members of the population resemble each other. For height in a national population it works well. For conversion rate in an eCommerce category it works badly, because two stores filed under the same label can differ by a factor of several while both being run competently, and the reasons have little to do with site quality.

So the question is not whether the average is right. It is how much information it carries about you, and the answer depends on four distortions that are worth naming separately, because they have different magnitudes and different remedies.

Distortion 1: mix, or the average that moves when nothing moves

Mix effects are counterintuitive enough that they routinely survive review by careful people. The mechanism is simple: if a benchmark's sample grows to include many small, young, paid-traffic-heavy stores, the average falls. Every existing store may have improved. The published figure still went down.

This matters in both directions. A benchmark that rose does not mean the field got better, and your position relative to it can shift by a meaningful margin without anything happening on your site at all. Year-on-year benchmark comparisons are therefore close to meaningless unless the source states that the sample is held constant, which it almost never does.

The same effect operates inside your own store. A sitewide conversion rate is itself an average across device, source and category, so it moves whenever your traffic composition moves. A successful campaign that brings in cold traffic reliably lowers a sitewide rate while raising revenue, and teams reporting on the rate rather than on the revenue have been known to switch off the campaign.

Distortion 2: shape, or the mean that describes nobody

When a distribution is symmetric, the mean is a fair summary. Conversion rates are not symmetric. The floor is zero and the ceiling is soft, and a handful of stores with unusual traffic, subscription mechanics or extremely narrow catalogues sit far above the bulk of the field. Those stores drag the mean.

The practical consequence is that the typical store is below average, which sounds like a paradox and is simply what skew does. A team benchmarking against a published mean and finding itself below it has learned very little, because so is most of the sample.

The fix is cheap when it is available: use the median. Where a source publishes only a mean, treat the figure as optimistic and do not build a target on it. Where a source publishes percentiles, use those instead of either, because a range tells you about the spread and a single number never can.

Distortion 3: definition, or two sources measuring different things

This is the distortion most easily fixed and most often skipped, because checking it is tedious. Two benchmarks for the same category, published in the same year, can differ substantially while both are correct, purely on measurement choices.

The usual suspects are few and worth memorising. Session definition and timeout. Bot and internal traffic exclusion, or the absence of it. Whether a conversion counts at order placement or at payment confirmation. Whether mobile app traffic is in scope. Whether returns are netted off. Each is defensible, and each moves the number.

This is also why comparing your figure to a published one requires reconciling your own definitions first. A store measuring conversions at payment confirmation against a benchmark measuring them at order placement will read low by a margin that has nothing to do with performance.

Distortion 4: time, or the average that already expired

Seasonality alone moves category conversion rates enough to swamp most of what a store could achieve through optimisation in the same period. A benchmark that spans a peak looks strong; one that avoids it looks weak. When the window is not disclosed, the reader has no way to correct for this.

There is a slower version of the same problem. Published benchmarks lag their measurement period, sometimes by many months, and the market conditions they describe may no longer hold. Using a figure without knowing its vintage is a common way to set a target against a market that has moved.

The bottom row is the honest summary. Because the four distortions compound in directions you cannot determine from a published figure, the responsible reading of any single benchmark number is as a rough region rather than a point.

How to use industry averages anyway

Bounding. Knowing roughly where a category tops out stops a team committing to a target that nothing in the data supports. This works because it needs only the order of magnitude, which survives every distortion above. A target set at four times the category range is wrong regardless of how the benchmark was built.

Triage. A benchmark is a reasonable trigger for whether to spend a day looking. If a figure is wildly outside the plausible region, something is worth checking, and frequently what is worth checking is your own measurement rather than your performance. Implausible numbers are more often tracking faults than business events.

Communication. An external reference is genuinely valuable when explaining a result to a board, an investor or a new stakeholder who has no context for what normal looks like. Used honestly, with the caveats attached, it makes a conversation possible that would otherwise be an argument about vibes.

What none of the three do is tell you what to change. That requires information about your store, which an average does not contain by construction. The widely quoted Baymard Institute figure of roughly 70% average cart abandonment illustrates the point neatly: it is a real, carefully produced average, it is useful for knowing that abandonment is normal and large, and it cannot tell any individual store which of its own checkout steps is losing people.

What to do this week

• Delete the gap-to-average line from your reporting. It is the most misread number in a CRO deck, because it looks like available upside and is mostly composition.
• Write down your own definitions. Session, conversion, exclusions, device scope. One paragraph. Without it no external comparison you make is valid.
• Lead the weekly report with your own trend. Movement you cannot explain is the real alarm; distance from an average is not.
• Use the benchmark for bounding only. Set the plausible range at the start of planning, then put it away.
• Check the vintage of any figure you quote. If the window is not stated, say so when you quote it, or do not quote it.

Where the underlying difficulty is that order, traffic and cost data never sit together long enough to build a defensible baseline of your own, that is a data problem rather than a benchmarking one. Nexus by Omniconvert is an AI for eCommerce growth engine that unifies commerce data, prioritises experiments by True Profit, and generates campaigns and creative you approve before they go live. Bain and Company's retention work with Fred Reichheld makes the adjacent case that the compounding advantages are the ones worth measuring properly, and those are precisely the ones a category average cannot see.

Once you know which distortion is biting, the next question is what to fix, and that is an audit rather than a comparison. Our neighbours at croaudit.marketing cover how to audit it step by step, and Omniconvert Explore is the CRO platform for validating the change afterwards, with A/B and multivariate testing, on-site surveys, heatmaps and session recordings.

FAQ: industry averages and benchmark data

Why do industry averages vary so much between sources?

Because each source aggregates a different set of stores and defines the metric differently, and both differences are usually undisclosed. One dataset may be weighted toward large paid-traffic retailers and another toward small email-led brands, which produces genuinely different averages from equally honest measurement. Before comparing two published figures, check what counted as a session and what counted as a conversion in each, because that single definitional difference explains more of the gap than anything else.

What is a mix effect, and why does it matter for benchmarks?

A mix effect is when an aggregate number moves because the composition of the group changed, not because any member of the group changed. A category average can fall while every single store in it improves, simply because more low-converting stores joined the sample. This is why a benchmark moving year on year tells you almost nothing about whether stores are getting better, and why your own position relative to it can shift without you doing anything at all.

Should I use the median instead of the average?

Yes, wherever it is offered, because conversion data is skewed rather than symmetric. A small number of very high performers pull an arithmetic mean upward, so the average describes a position that most stores in the group are below. The median is the figure that answers the question people think they are asking, which is what a typical store in this category looks like. If a published benchmark does not say which it used, assume the mean and treat it as optimistic.

Are industry averages useful for anything at all?

Three things, and they are genuinely useful. Bounding tells you the plausible range so you do not set an impossible target. Triage tells you whether a number is unusual enough to justify investigation time. Communication gives a shared external reference when you need to explain to somebody outside the team why a figure is or is not alarming. What an average cannot do is tell you what to change, because it contains no information about your store.

How far from the average should I be before I worry?

Distance from an average is the wrong trigger, because the spread inside any category is wide enough that being well below it can be entirely normal for your traffic mix. A better trigger is movement in your own trend that you cannot explain. A store sitting far below a category average with stable, profitable unit economics has no problem. A store sitting comfortably on the average while its own rate declines month over month has a real one.

What is the most common mistake people make with benchmark data?

Treating the gap between their number and the average as a quantity of available improvement. It is not a gap that can be closed, because much of it is explained by traffic mix, price point and category composition rather than by anything on the site. Teams that plan against that gap set targets they cannot reach, and then read a normal result as a failure, which is a reliable way to abandon a programme that was working.