Customer segmentation is splitting your visitors or customers into groups that behave differently, such as new versus returning shoppers, mobile versus desktop users, or paid versus organic traffic, so you can show each group a better experience and read test results group by group.
It only pays off when the groups really respond differently to the same change, and when each group has enough traffic to test on its own.
Most segmentation advice stops at "know your customer."
This page uses tests from the LiftArchive database to show which segments have produced measurable differences, where segment results misled the people who ran them, and how to tell the two apart.
Ecommerce customer segmentation
An ecommerce store can segment on almost anything it records.
In practice, a small set of segments shows up again and again in tests that moved revenue, because these groups arrive with different questions.
| Segment | What differs about the group | What tends to get tested |
|---|---|---|
| New vs returning visitors | New visitors need trust and explanation; returning visitors already know the brand | Trust content and guarantees for new visitors; carts, recommendations and offers for returning ones |
| Device (mobile vs desktop) | Screen size, attention and how people browse | Layout, image galleries, navigation |
| Traffic source (paid, organic, email) | Paid visitors arrive from one ad with one promise | Ad-matched landing pages, auto-applied codes, removing distractions |
| Location | Currency, shipping speed and local relevance | Local currency, country-specific shipping lines |
| Purchase history (RFM: recency, frequency, monetary value) | Value and loyalty | Mostly used in email and retention, less often on-site |
New vs returning visitors
This is the most tested split in our database, and several results are clear.
Canopy moved its press and retailer logos higher on the homepage for new visitors only, and conversion rate rose 6.4% (Visually case study).
Underoutfit showed new visitors a homepage banner about its 180-day try-on and returns policy and got a 2.6% conversion lift, also reported by Visually.
Returning visitors respond to different things. Branch Furniture added a five-product recommendation slider for visitors from the last 60 days and saw +8.9% conversion rate and +21.2% revenue per session.
An Optimizely retailer swapped featured products for bestsellers for new visitors and got 23.01% more clicks to product pages.
A caveat applies to all of these: they are vendor case studies, and none of them states the sample size or significance.
Treat them as ideas worth testing, not as expected results.
When the segment split didn't hold up
Clean Commit's tests are a useful counterweight because the agency published the weak results too.
One Shopify store got an educational homepage for new visitors and a sales-focused one for returning visitors.
The headline was +241.1% conversion rate, but it never reached significance, and overlapping tests and holiday promotions got in the way.
The same idea on product pages produced +1.0% conversion rate with flat orders across 2,581 visitors over 33 days, while add-to-cart fell 10.7%.
The lesson is not that segmentation fails. Building a separate experience per segment cuts your traffic, and a huge percentage on a thin sample is more often noise than a finding.
Device segments
Device is where the same change most often goes in opposite directions.
Stone Creek Coffee tested two-image product cards on collection pages: mobile conversion rose 20% while desktop fell 12%, so Blend Commerce shipped the change on mobile only.
An apparel brand saw the reverse with a simplified homepage hero, which lifted desktop conversion 26% while mobile declined.
A gin subscription brand kept a press-quote change on desktop only after it added A$9,392 a month on desktop and lost on mobile (Convertibles).
If you only read the blended result, each of these would have looked like a modest win or a wash.
For device-level benchmarks, see mobile vs desktop conversion rates.
Traffic source and location
Paid visitors are an easy segment to define because the ad tells you what they came for.
Laura Geller Beauty auto-applied the advertised discount for paid-campaign visitors and saw +15.9% revenue from paid traffic.
Dave Diederen hid the "As Seen On" press bar on the homepage for paid traffic and measured +2.61% revenue per visitor.
Location tests in the database lean on currency and shipping.
Dave Diederen's geo-detection popup asking international visitors to confirm their country and currency returned +13.01% revenue per visitor, and Readers.com's country-localized checkout gained 30% in conversion rate (One Click).
@DaveDiederen
VisuallyVisually
Blend CommerceVisually
Clean CommitWebsite segmentation
Website segmentation is used for two different jobs, and mixing them up causes most of the trouble.
Targeting means choosing who sees a variant before the test starts.
Analysis means slicing a finished test's results by group afterwards.
Segments in your analytics tool
In Google Analytics 4, a segment is "a subset of your Analytics data that can be used in Explorations", and it comes in three scopes.
User segments group people across visits (for example, past purchasers).
Session segments group single visits (for example, sessions from one campaign).
Event segments group individual actions. Google's help page says you can create up to 10 segments per exploration and apply up to 4 at once.
Pick the scope that matches the question. "Do returning customers buy more after seeing recommendations?" is a user question.
"Does this landing page work for traffic from one campaign?" is a session question.
Targeting: decide the segment before the test
Targeted tests are the safer use of segmentation.
You state the hypothesis ("new visitors lack trust, so a returns guarantee above the fold will help them"), limit the test to that group, and judge it on one metric.
Canopy, Underoutfit and Laura Geller above all worked this way.
The cost is traffic. If returning visitors make up 25% of your sessions, a test limited to them takes about four times as long to reach the same sample as a site-wide test.
Run the numbers in the sample size calculator with the segment's own traffic and baseline conversion rate before you commit.
A segment you can't power is one to target with a site-wide change, not one to test on its own.
Analysis: slicing results after the fact
Post-test segmentation is where false wins come from.
Each extra segment is another chance for random noise to cross the significance line.
A worked example: you run a test where the variant truly has no effect, then check 20 segments at a 5% significance level.
The chance that at least one segment shows a "significant" result is 1 − 0.95^20, which is about 64%.
On average you will find one false winner (20 × 0.05 = 1), even though nothing changed.
| Segments checked | Chance of at least one false positive (α = 0.05, no real effect) |
|---|---|
| 1 | 5% |
| 5 | 23% |
| 10 | 40% |
| 20 | 64% |
This assumes the segments are independent, which real segments rarely are, but the direction holds.
Optimizely's documentation is explicit that its correction does not carry over: when you segment results, "the false discovery rate control is not maintained", and it recommends using segments for exploration rather than decisions.
Our page on multiple comparisons covers corrections such as Bonferroni and Benjamini-Hochberg.
Segment results can also reverse when traffic mix shifts between groups, which is Simpson's paradox.
A device split where mobile wins and desktop loses, as at Stone Creek Coffee, is something else: a real interaction, which you can only trust if each device has enough traffic on its own.
How to use a segment result you found afterwards
- Treat it as a new hypothesis, not a result. Rerun the test targeted at that segment.
- Ask whether there's a believable reason. "Mobile cards with two images help thumb browsing" is a reason; "it won for Tuesday Safari users" usually isn't.
- Check the segment's sample. A lift on a few hundred visitors, like many headline numbers above, can easily be noise.
- Watch the full funnel. The Global Leadership Summit's localized landing page raised clickthrough 32% but did not change registrations.
When segmentation is the wrong choice
Skip a separate segment experience when your site can't power the test for that group, when the change fixes something broken for everyone (a slow page or a confusing form), or when you'd have to maintain more versions than your team can keep updated.
In those cases, one better page for everyone usually beats three average ones.
Segmentation is the base layer for personalization: behavioral targeting, dynamic content and product recommendations all depend on defining the group first. Browse all tests for more segment-level results.