Why averages hide the real problems

Average order value: $75. Looks fine. But behind that average: half the orders are $150, half are $0 shipping-only orders that someone configured wrong. The average looks healthy while a significant problem lurks underneath. Averages are comfortable, convenient, and often misleading. They compress complex reality into single numbers, and in that compression, important information disappears.

Understanding how averages hide problems helps you look beyond them to see what’s actually happening in your business.

What averages actually show

The mathematical reality:

Averages show central tendency

The average is the sum divided by count. It indicates where the center of the data is. Central tendency is useful but limited information.

Averages don’t show distribution

A $75 average could be everyone ordering around $75. Or half at $50 and half at $100. Or a few at $500 pulling up many at $20. The average alone can’t distinguish these scenarios.

Averages don’t show variance

How spread out is the data? High variance and low variance can produce identical averages. The stability or instability underlying the average is invisible.

Averages are sensitive to outliers

One $10,000 order among 99 orders of $50 creates an average of $149. The average suggests typical orders are $149. In reality, 99% of orders are $50. Outliers distort averages severely.

What gets hidden

Problems that averages obscure:

Bimodal distributions

Two distinct groups that behave differently. High-value customers and low-value customers. Happy customers and frustrated ones. The average falls between two peaks that actually exist, suggesting a middle that’s actually empty.

Long tails

Most customers cluster low, with a few outliers pulling the average up. The average suggests typical behavior that most customers don’t actually exhibit.

Segment problems

Average conversion is 2.5%. But mobile is 1.2% and desktop is 3.8%. The average hides a significant mobile problem. Segment-level issues disappear into overall averages.

Trend reversals

New customers converting well, returning customers converting poorly. The average might stay stable while the mix shifts dangerously. Compositional changes hide in stable averages.

Quality issues

Average satisfaction is 3.5 out of 5. But maybe 70% are 4s and 5s, and 30% are 1s and 2s. The average suggests moderate satisfaction. Reality is polarized opinions.

Specific e-commerce average traps

Where this shows up:

Average order value

A healthy-looking AOV might hide: a few whale orders distorting typical behavior, a segment ordering unusually low, free shipping thresholds creating artificial clustering, or discount abuse creating low-value orders.

Average conversion rate

Overall conversion might hide: traffic source quality differences, device-specific problems, time-of-day patterns, new versus returning customer gaps, or campaign-specific issues.

Average customer lifetime value

Average LTV might hide: a small cohort of super-loyal customers pulling up the average, most customers never returning, or recent cohorts performing worse than older ones.

Average response time

Average support response time might hide: quick responses for easy tickets and terrible delays for complex ones. The average looks fine while difficult customers suffer.

Average page load time

Average load time might hide: fast loads for cached returning visitors and slow loads for new visitors. Or fast in some regions and slow in others.

Better alternatives to averages

What to look at instead:

Median

The middle value when data is sorted. Less sensitive to outliers than average. If average is $75 but median is $45, something is pulling the average up. The gap between average and median reveals distribution shape.

Percentiles

What’s the 25th percentile? The 75th? The 90th? Percentiles show distribution spread. The interquartile range (25th to 75th) shows where most data falls.

Histograms

Visual distribution of values. Where do customers cluster? Are there multiple peaks? What’s the shape? Histograms show what averages hide.

Segment breakdowns

Average by segment. By device, by source, by customer type, by product category. Segment-level averages reveal problems that overall averages obscure.

Distribution over time

How has the distribution changed, not just the average? Has variance increased? Have the tails shifted? Distribution change can be more important than average change.

The segmentation imperative

Why you must break averages down:

Businesses aren’t monoliths

Different customers behave differently. Different channels perform differently. Different products have different patterns. One average can’t represent heterogeneous reality.

Problems are usually localized

When something goes wrong, it usually goes wrong in a specific place. Mobile checkout. Facebook traffic. New customers. The overall average might not move much while a specific segment suffers badly.

Solutions require specificity

“Improve conversion” is vague. “Improve mobile checkout for new visitors from paid search” is actionable. Segmentation reveals where action should focus.

Simpson’s paradox lurks everywhere

It’s possible for every segment to improve while the overall average declines (or vice versa). Compositional changes create paradoxes. Only segment-level analysis reveals them.

When averages are appropriate

Fair use of averages:

Quick health checks

Is the average roughly where expected? Big deviations from historical average warrant investigation. Averages work for initial screening.

Communication simplification

“Our average customer spends $75” is easier to communicate than a full distribution analysis. For summary communication, averages have a role.

Trend monitoring

Is the average moving over time? Trend direction in averages can indicate overall trajectory, even if the average doesn’t tell the full story.

Comparison benchmarks

Industry averages provide rough comparison points. Your average versus industry average gives directional sense, even though both averages hide detail.

Building beyond-average habits

Practical implementation:

Default reports include distribution

Don’t just show the average. Show the median alongside. Show a simple histogram. Make distribution visibility automatic.

Segment view always available

One click from any average to segment breakdown. If looking at overall conversion, easy access to conversion by device, by source, by customer type.

Outlier flagging

Automatic identification of values far from typical. Outliers might be problems to fix or insights to understand. Either way, they should be visible.

Variance tracking

Track not just averages over time but variance over time. Increasing variance might indicate emerging problems or changing customer base.

Regular distribution review

Monthly or quarterly, look at full distributions of key metrics. See what the averages have been hiding. Make distribution review a scheduled practice.

Questions to ask about any average

Standard investigation:

What’s the median?

If average and median differ significantly, outliers are influencing the average. Investigate what’s creating the pull.

What does the distribution look like?

Is it normal, skewed, bimodal, long-tailed? Distribution shape changes interpretation of the average completely.

Are there meaningful segments?

Does the average differ significantly between important groups? If yes, the overall average is misleading about segment-level reality.

What are the extremes?

What are the highest and lowest values? Understanding the range provides context the average alone can’t give.

Has variance changed?

Even if the average is stable, has the spread of values changed? Stable average with increasing variance might indicate emerging instability.

Frequently asked questions

Isn’t analyzing distributions more complicated than just using averages?

Slightly, but not much. Median is one additional number. A histogram is one chart. The added complexity is minor; the insight gained is significant. Most analytics tools make distribution analysis easy.

What if my data is normally distributed?

If data is truly normally distributed (rare in e-commerce), averages are more reliable. But you can’t know it’s normally distributed without looking at the distribution. Assuming normality is the same error as relying only on averages.

When should I be suspicious of an average I see?

Always be at least mildly suspicious. But especially: when average seems too good or too bad, when average hasn’t changed but something feels different, when making important decisions based on the number, or when average represents diverse groups.

How do I communicate distribution complexity to stakeholders?

Visualizations help. Histograms communicate distribution intuitively. You can also use language like “while the average is $75, most orders are actually between $40-60, with a few large orders pulling the average up.” Narrative plus visual beats number alone.