Why your conversion rate can fall while every segment improves

Your conversion rate was 3.1% last month. This month it is 2.4%. Nobody touched the homepage. Same product, same prices, same checkout. So you open the dashboard on Monday, see the drop, and start hunting for what broke.

Here is the uncomfortable part. It is possible that nothing broke. It is possible that every kind of visitor you get converted better this month than last, and the single number on your dashboard still fell. That is not a glitch in the analytics. It is arithmetic doing exactly what it does.

A blended conversion rate is an average. Averages hide things — and the thing they hide most often is that your traffic changed while your site stayed still.

TL;DR: One site-wide conversion rate is a weighted average of very different visitors. When your traffic mix shifts — more mobile, more cold ad clicks, more first-timers — the blended number can fall even if every individual segment improved. Before you redesign anything, split the rate by device, by new-versus-returning, and by source. Read the segments, not the blend.

An average is one number pretending to be the whole story

Your site does not have "a" conversion rate. It has dozens, stacked on top of each other and flattened into one figure for the dashboard.

Desktop buyers convert at one rate. Phone browsers convert at another. People who already know you convert far better than strangers who clicked an ad thirty seconds ago. Someone arriving on a branded search is a different animal from someone caught by a cold display banner. Each of these groups has its own rate, and each group is a different size every single week.

The number you stare at is a weighted average of all of them. "Weighted" is the word that matters. Each segment pulls the blended rate towards its own value in proportion to how much traffic it sent. Change the proportions — send more of the low-converting kind, less of the high-converting kind — and the average moves on its own. Your page never got a vote.

This is why the Monday-morning panic is usually aimed at the wrong target. The founder blames the new hero section. The real culprit is a campaign that started dumping cheap, curious, low-intent clicks into the top of the funnel. Same site. More tyre-kickers. Lower average.

Berkeley watched a rate reverse in 1973

The cleanest proof that an average can betray you did not come from marketing. It came from a university.

In the autumn of 1973, UC Berkeley looked like it was rejecting women. Across the graduate school, 44% of the 8,442 men who applied were admitted, against 35% of the 4,321 women. That is a wide gap across nearly 13,000 applications — far too wide to write off as noise.

Then three researchers split the data by department. Bickel, Hammel and O'Connell published the result in Science in 1975, and the correction was blunt: the departments showed "a small but statistically significant bias in favor of women." Not against. In favour.

Both facts were true at once. Women were admitted at a higher rate in most individual departments, yet at a lower rate overall. The trick was where people applied. Women applied in larger numbers to the most competitive departments — the ones that rejected almost everyone, men included. Men clustered in departments that waved most applicants through. The aggregate number reversed the truth of every department underneath it.

Statisticians call this Simpson's paradox: a pattern that holds inside every group can flip when you mash the groups together. Berkeley is the textbook case. Your conversion dashboard is the same maths wearing a marketing costume.

Watch it happen with your own traffic

Numbers make this concrete faster than any explanation. Take a shop with two sources of traffic and follow it across two months.

Read the segment rows first. Branded search improved, 5.0% to 6.0%. Cold display improved, 1.0% to 1.5%. Every visitor type got better at converting. You could not have asked for a cleaner month.

Now read the blended row. It fell from 3.0% to 2.1%.

Nothing on the site got worse. A display campaign in Month 2 poured 3,000 cold clicks on top of a much smaller warm pool, and cold clicks convert low by nature. The mix moved. The average followed the mix, not the page. If you had spent that week rewriting the homepage to "fix conversion", you would have been treating a healthy patient — and you would have congratulated yourself when the campaign eventually ended and the number floated back up.

Three splits move your blended rate more than your page does

You do not need a statistics degree to catch this. You need three cuts, and they take about ten minutes in any analytics tool.

Device. Desktop and mobile do not convert alike. In 2025 ecommerce benchmarks (Smart Insights), desktop tends to convert at roughly double the mobile rate — think 3.9% against 1.8%, though the gap has been closing. A seasonal swing towards phone traffic, or an Instagram push that lands mostly on mobile, drags the blended rate down while both device rates hold perfectly steady.

New versus returning. Returning visitors convert at two to three times the rate of first-timers — Shopify's benchmarks put returning around 4.5% against roughly 2.1% for new. So a burst of top-of-funnel reach, a viral post, or a fresh awareness campaign floods you with new visitors and pulls the average down. That is growth showing up as a "problem" on the dashboard.

Source intent. A branded search click and a cold retargeting impression are not the same buyer at different speeds. They want different things and convert an order of magnitude apart. Shift budget from search to display and the blended rate drops on contact — before a single visitor has judged your offer.

Device, intent, recency. When the site-wide number moves and you have not shipped a change, one of these three moved first almost every time.

What to do before you touch the page

The fix is a habit, not a tool. When the blended rate jumps or drops, run the diagnosis before the redesign.

Segment the same rate three ways — device, new versus returning, and source or channel. Compare each segment to its own past self. Comparing segments against each other tells you nothing; a segment is only interesting when it moves. If every segment is flat or up and only the blend fell, stop. The mix moved, and the site is fine. Go look at what changed in your traffic: a campaign that launched, a channel that spiked, a season that turned.

If a specific segment dropped — mobile cratered, or returning visitors fell off — now you have something real to chase, and you know exactly where to point. That is a page problem, or a checkout problem, or a broken flow on one device. It is fixable because it is located.

The average didn't lie to you. You asked one number to do the work of four.

When the blended number is still worth watching

None of this makes the site-wide rate useless. It makes it a smoke alarm, not a diagnosis.

A moving blended rate is a fine trigger to go and look. It tells you something shifted. It never tells you what, because it cannot separate a worse page from a different crowd. The alarm has one job: get you to open the segment view. Everything useful happens after that.

The trade is real, and worth naming. Segments mean smaller samples, and smaller samples are noisier — a 40% mobile conversion rate off ten sessions is a rounding error, not a finding. You give up the comfort of one clean figure for three or four honest, wobblier ones. That is the correct trade. A single confident number that points the wrong way is more expensive than four modest numbers that point where the problem actually is.

At SharpHaw we treat conversion as the only brief worth having — but the first move is always to read the number properly before reaching for the page. Diagnose, then build. If your conversion rate just moved and nobody can tell you which segment moved it, that is the conversation to have before anyone opens the site editor. That is the kind of read we do in the open.

Common questions

Why did my conversion rate drop right after I got more traffic?

More traffic is usually more new and more cold traffic, and both convert well below your site average. The extra visitors drag the blended rate down even when your existing audiences convert exactly as before. Check the new-versus-returning split first — a growth spike often looks like a conversion fall.

Is a falling conversion rate always a bad sign?

No. If every segment is holding or improving and only the blended number fell, the mix changed, not your site. That can even mean a successful awareness push. A falling rate is only a real problem when a specific segment — a device, a channel, a returning audience — falls on its own.

How do I segment my conversion rate?

Any analytics tool will do it. In GA4, compare conversion rate across device category, then across new versus returning users, then across session source or channel. Look at each segment against its own history, not against the others. The goal is to find the one segment that moved, or to confirm that none did.