Mean vs median: when the average lies about your users

Mean vs median in one line: the mean is the sum divided by the count, the median is the middle value once you sort the data. When data is roughly symmetric, they are close and either works. When data is skewed, as revenue per user, session length, order value and time-to-delivery almost always are, a few extreme values drag the mean far from what a typical user experiences, and the median gives a more honest picture of "normal". Neither is always right: use the median to describe the typical user and the mean when you need totals, such as revenue forecasts or budgets.

A ten-customer example

Here is 90-day spend for ten customers of CatChow, an online pet-food shop, sorted from smallest to largest:

Customer 10 is a cat shelter that stocked up for the whole winter.

• Mean: $(18 + 22 + \dots + 45 + 480) / 10 = 752 / 10 = 75.2$ dollars.
• Median: with ten values, it is the average of the 5th and 6th: $(30 + 32) / 2 = 31$ dollars.

Now look at the table again. Nine of the ten customers spent less than the mean. If a marketer builds a campaign for "our average customer who spends 75 dollars", it targets almost nobody. The median of 31 dollars is much closer to what a regular CatChow customer looks like.

To see how fragile the mean is, change the shelter's 480 to 48. The mean drops from 75.2 to 32 dollars. The median stays at 31. One customer moved the mean by more than 40 dollars and didn't touch the median at all.

Why this happens: skewed data

A distribution is skewed when one tail is much longer than the other. Product data is full of right-skewed metrics, with most values bunched near the low end and a long tail of large ones:

• Revenue and order value: most orders are ordinary, a few are huge (wholesale buyers, gift orders, annual subscriptions).
• Session length: most visits last a couple of minutes; a few tabs stay open for hours.
• Time to first purchase, time to deliver, support resolution time: you can't go below zero, but you can go very far above the typical case.
• Number of orders per customer: many one-time buyers, a small loyal core with dozens of orders.

In a right-skewed distribution, the mean is pulled toward the long tail, so mean > median. The bigger the gap between them, the more skewed the data. That gap itself is a quick diagnostic: if your dashboard shows an average session of 7.5 minutes but a median of 3, the "average" user doesn't exist, and a handful of very long sessions are carrying the number.

When the mean is the right choice

The median is not automatically better. The mean has one property the median lacks: mean × count = total.

That makes the mean the right tool whenever the decision is about totals:

• Revenue forecasts: if 10,000 customers spend a mean of 75.2 dollars, you can expect about 752,000 dollars. Multiplying the median by 10,000 would underestimate revenue by more than half.
• Budgets and unit economics: customer acquisition cost, lifetime value and cost per order are all means, because you pay for the total.
• A/B tests on revenue: the business cares whether total revenue goes up. A change that leaves the median customer untouched but makes big spenders spend more is still worth money, and only the mean will show it.

So the rule is not "median good, mean bad". It is: describe the typical user with the median, plan money with the mean, and always check both.

When the median can mislead too

The median has blind spots of its own.

Zero-heavy metrics. Suppose 1,000 people visit a CatChow landing page, 40 buy, and each order is 50 dollars. Revenue per visitor has a mean of $40 \times 50 / 1000 = 2$ dollars and a median of 0, because more than half of visitors spent nothing. Run an A/B test on that page, and both variants will almost certainly have a median of 0, even if one earns twice as much. For metrics like this, report the conversion rate and the mean order value among buyers separately.

Changes in the tail. A loyalty programme that makes top customers order more often may not move the median at all. If the tail is exactly what you're trying to change, the median is the wrong lens.

Middle-ground options

When you need something between the two:

• Trimmed mean: drop a fixed share of the smallest and largest values, then average the rest. A 10% trimmed mean of our ten customers drops the 18 and the 480: $254 / 8 = 31.75$ dollars. It keeps more information than the median while resisting outliers.
• Capping (winsorising): replace values above, say, the 99th percentile with the 99th percentile itself. Common in A/B tests on revenue, because it reduces noise without throwing users away.
• Percentiles: report the 25th, 50th and 90th percentiles together. "Median delivery time is 2 days, but 10% of orders take more than 6" says far more than any single number.
• Log scale: for very skewed data, analysing the logarithm of the values often makes the distribution close to symmetric.

How to compare groups with skewed data

A common question after "mean or median?" is "then which test do I use?". If you compare two groups on a skewed metric:

• With large samples, a t-test on means still works reasonably well, because the averages of big samples are close to normally distributed. Capping extreme values helps a lot.
• With small samples or heavy tails, consider a rank-based test such as Mann–Whitney, which compares whole distributions rather than means and isn't thrown off by a single huge value.
• Bootstrapping, resampling your data many times, gives a confidence interval for the mean or median without strong assumptions.

Whichever you choose, decide before looking at the results. Switching from mean to median because the median "looks better" is a quiet way to fool yourself.

A quick checklist for any dashboard metric

1. Plot the distribution (a histogram or box plot), not just the summary number.
2. Compare the mean and the median. A large gap means skew or outliers.
3. Ask what the decision is about: a typical user (median) or a total (mean)?
4. Check for zeros. If most values are zero, split the metric into "how many" and "how much".
5. Report a spread measure too: percentiles or an interquartile range.

Key takeaways

• The mean is the sum divided by the count; the median is the middle of the sorted data.
• In skewed data, a few extreme values pull the mean away from the typical user; the median stays put.
• Use the median to describe what's typical, and the mean when you need totals such as revenue or budgets.
• Watch for zero-heavy metrics, where the median is 0 and useless.
• Trimmed means, capping and percentiles are good compromises. Choose your summary before you look at the results.

FAQ

Is the median better than the mean?

Neither is better in general. The median describes a typical value and resists outliers; the mean connects directly to totals. For skewed product metrics, look at both, and pick the one that matches the question you are answering.

Why is the mean higher than the median?

Because the data is right-skewed: a few very large values pull the mean up while the median, which only depends on the middle of the sorted data, barely moves. Revenue, order value and session length usually behave this way.

Should I report average order value as a mean or a median?

Report the mean when you use AOV for revenue planning, because mean AOV times the number of orders gives total revenue. Add the median, or a few percentiles, when you want to show what a typical order looks like.

What is a trimmed mean?

The average after dropping a fixed share of the lowest and highest values, for example 5% or 10% from each end. It is a compromise between the mean and the median: less sensitive to outliers than the mean, but it uses more of the data than the median.

Learn it hands-on

The free lesson Typical value: mean, median, mode works through real-looking CatChow data and lets you practise choosing the right summary, with quizzes and an AI-checked task. When you're ready for the next step, Outliers and skewed data shows how to spot and handle the extreme values behind most misleading averages. Both are part of the free course Statistics for Product Managers and Marketers.