SEO ROI & Measurement · 4 min read
Why Averages Mislead: Reading Conversion and UX Numbers
Short answer
A single average conversion rate or task time blends very different visitors, pages and intents, so it can rise while every segment gets worse; read numbers by segment, check sample size, and prefer medians for skewed data like time on page or order value.
Key takeaways
- Always split conversion rate by intent, landing page type and device before drawing a conclusion.
- A change in traffic mix can move the site-wide average even when no page improved.
- Ask any benchmark for its sample, definition and spread before comparing your site to it.
A dashboard says your conversion rate fell from 3.1% to 2.6%, and someone wants to redesign the checkout. Before anyone touches it, find out what the average is made of. This page shows how blended numbers mislead and gives a short routine for reading conversion and usability figures without fooling yourself.
Why is a site-wide conversion rate a weak signal?
A site-wide rate mixes visitors with different intents, so it mostly reflects your traffic mix rather than how well any page performs.
A brand searcher landing on pricing is close to buying; someone asking a general question and landing on a guide is months away. Blend them and the rate mostly reflects the ratio between them.
Typical segments worth separating:
- Search intent: branded, commercial, informational, local.
- Landing page type: product, category, service, article, tool.
- Device and browser.
- New versus returning visitors.
- Channel: organic, paid, email, referral, AI assistant referrals.
Our guide to search intent mapping explains how to label pages by intent so these splits are easy to build.
What is Simpson's paradox, and does it really happen in marketing data?
Simpson's paradox is when a trend that holds in every segment reverses in the combined total, and it shows up whenever traffic mix shifts.
Hypothetical example, with round numbers:
| Segment | Month 1 visits | Month 1 rate | Month 2 visits | Month 2 rate |
|---|---|---|---|---|
| Product pages | 8,000 | 5.0% | 8,000 | 5.5% |
| Blog articles | 2,000 | 0.5% | 12,000 | 0.6% |
| Blended | 10,000 | 4.1% | 20,000 | 2.6% |
Both segments improved, yet the blended rate fell by more than a third because a content push added low-intent visits. Total conversions rose from 410 to 512.
How small is too small for a sample?
If the confidence interval around a difference is wider than the difference itself, you do not have a result yet.
Conversion rates are proportions, and proportions from small samples swing a lot. Hypothetical example: page A gets 12 conversions from 400 visits (3.0%) and page B gets 18 from 400 (4.5%). That looks like a 50% lift. Run a basic two-proportion test and the difference is well within what chance produces at that volume.
A few habits help:
- Decide the minimum change worth acting on before the test starts.
- Estimate the sample needed for that change, then run the full length.
- Do not stop early because the chart looks exciting on day three.
- Report the interval, not only the point estimate.
Small segments also hit reporting limits. Google Analytics 4 may withhold rows through data thresholds when user counts are low, which is one more reason to widen date ranges before slicing finely.
When does the median tell the truth better than the mean?
For skewed data such as order value, time on page or task completion time, the median describes the typical case and the mean describes the tail.
The NIST statistics handbook notes that extreme values in the tails distort the mean, while the median is based on ranks and resists them. Hypothetical example: nine shoppers spend about $60 and one buys $2,400 of equipment. The mean order is roughly $294; the median is $60. Plan inventory or bids on the mean and you will be surprised every month.
Usability task times have the same shape: most people finish quickly and a few get stuck for ages.
How should you read benchmark claims?
Treat any benchmark as a claim to check: ask how it defines the metric, who was sampled and how widely results varied.
Questions to put to a "the average site converts at X%" statistic:
- What counted as a conversion: purchase, lead, signup, any goal?
- Which industries, traffic sources and devices were in the sample?
- How many sites, and over what dates?
- Is the figure a mean or a median, and what was the spread?
- Who published it, and do they sell something that benefits from the number?
Without those answers, compare against your own segmented history instead. Plug segment-level rates, not the blended rate, into the SEO ROI calculator when you project returns from new organic traffic.
What does a better conversion report look like?
A useful report shows counts and rates by segment, with sample sizes and a note on what changed in the traffic mix.
Keep the blended rate if executives expect it, with the segment table and a one-line note on mix changes right under it. Pair it with why analytics tools disagree so readers know the measurement gaps, and see our UX for conversion page for design fixes that hold up once the numbers are read correctly. If you want this reporting set up properly, our SEO ROI measurement service builds it from your own data.
Frequently asked questions
What is a good conversion rate?
There is no universal figure, because the number depends on what counts as a conversion and who arrives. A product page reached by a brand search and a blog post reached by an informational query should never share one target.
How many conversions do I need before trusting a test?
Enough that the confidence interval around the difference is narrower than the change you care about. With only a few dozen conversions per variant, most observed differences are noise, so plan the test length before you start.
When should I use the median instead of the average?
Use the median for skewed measures such as order value, session duration or task completion time, where a few extreme values pull the mean. Report both if stakeholders are used to averages.
Why did our conversion rate drop after an SEO win?
New organic traffic often arrives on informational pages with lower purchase intent. Total conversions can rise while the blended rate falls, so track both the count and the rate by segment.