Most stores segment on the wrong axis. They split a list by country or by gender, send two versions of the same email, and see no difference — because neither group was defined by anything that predicts buying.
Useful segmentation starts from a different question: what would I say differently to this group? If the answer is nothing, the segment should not exist. This guide covers the segments worth building, how to layer them from simplest to most sophisticated, and the practical traps — duplicates, stale definitions, over-segmentation — that make segmentation cost more than it returns.
A segment is only real if it changes the message
The test is simple. Write the campaign you would send to the segment, then write the campaign you would send to everyone else. If they are the same email with a different name in the greeting, you have not segmented — you have added maintenance work.
This is why "customers in California" is usually a bad segment and "customers who bought once, 45 days ago, and have not returned" is a good one. The second implies a message. The first does not.
Practically, that means starting from campaigns and working backwards to segments, rather than building a segmentation model and hoping campaigns emerge from it.
Layer 1: engagement tiers
The first split every store should have, because it protects deliverability as much as it improves relevance.
| Tier | Last engagement | What they get |
|---|---|---|
| Engaged | Within 30 days | Your regular sends |
| Cooling | 30–90 days | Your best offers |
| At risk | 90–180 days | A win-back attempt |
| Dormant | 180 days or more | One re-permission campaign, then suppression |
What you do with these matters more than where you draw the boundaries — and continuing to mail the dormant tier is the single most common cause of the problems described in our deliverability guide. Adjust the windows to your purchase cycle: a coffee subscription and a mattress store should not share a definition of dormant.
Layer 2: RFM scoring
RFM is the most useful segmentation model in ecommerce because it needs only order data — which you already have — and it maps cleanly onto actions.
How to build it without a data team
- Recency — days since last order. Sort customers, split into five equal groups, score 5 (most recent) down to 1.
- Frequency — total orders. Same process, same 1–5 scale.
- Monetary — total spend. Same again.
Scoring by quintile rather than by fixed thresholds matters: it makes the model relative to your own customer base, so it works whether your average order is $20 or $2,000, and it keeps working as the business grows.
The groups that fall out, and what to send them
| Group | RFM shape | What to send |
|---|---|---|
| Champions | High on all three | Early access, new arrivals, referral asks. Never discount — they were going to buy anyway |
| Loyal | High frequency, mid recency | Replenishment reminders and cross-sells based on what they already own |
| Potential loyalists | Recent, one or two orders | A second-purchase nudge. The highest-leverage group in most stores — worth more than almost any acquisition spend |
| At risk | Previously high frequency, recency falling | A genuine win-back with a real reason to return |
| Lost | Low on everything | One last attempt, then suppress. Sending forever costs reputation and returns nothing |
Notice how different those five messages are. That is what a working segmentation model looks like.
Layer 3: behavioural segments from on-site events
Order data tells you what people bought. Event data tells you what they were about to buy — which is where the timely campaigns live.
Useful behavioural conditions, all derived from ordinary storefront tracking:
- Has an abandoned cart — added to cart, no purchase since.
- Viewed a product more than once without adding to cart — interest without commitment, usually a question you have not answered.
- Started checkout and stopped — the highest-intent group on your whole site.
- Category affinity — repeated views or purchases in one category, which makes new-arrival campaigns actually relevant.
- Time since last product view — separates "went quiet as a customer" from "went quiet entirely".
Layer 4: predictive-style segments without a model
You do not need machine learning to act on the pattern. Most of what predictive segmentation delivers can be approximated with conditions you can write down:
- Likely to buy soon — an active cart, or several product views in the last few days.
- Churn risk — order frequency slowing relative to their own history, not to the store average.
- High-value trajectory — a second order inside the window where your best customers placed theirs.
- Due for replenishment — days since purchase approaching the typical reorder interval for that product.
Replenishment timing in particular is worth computing properly: take the median gap between repeat orders per product category and trigger a few days before it. That single segment often outperforms every clever idea in the account.
A starter set: seven segments most stores should build
| Segment | The campaign attached to it |
|---|---|
| 1. New subscribers, never purchased | Welcome flow |
| 2. First-time buyers, 14–45 days ago | The second-purchase push |
| 3. Repeat customers | Cross-sell and loyalty |
| 4. Champions | Early access, referral asks, review requests |
| 5. Cart or checkout abandoners | Recovery |
| 6. At risk | Win-back, before they are gone |
| 7. Engaged-only sending audience | The safe list for high-volume moments, and for warming a new sending domain |
Seven segments, each with an obvious campaign attached. That is a complete programme for most stores, and it is achievable in an afternoon.
Mistakes that make segmentation expensive
- Building segments nobody sends to. Every unused segment is a definition that will silently drift out of date and eventually mislead someone.
- Slicing too thin. Splitting a 400-person segment into four means you can never tell whether a result is real. Keep segments large enough that the numbers mean something.
- Static lists that never refresh. A CSV exported in March is wrong by April. Segments should be conditions that re-evaluate, not snapshots.
- Ignoring duplicates. The same person appears as a Shopify customer and as a manually imported contact under a different email or a differently formatted phone number. They then receive the campaign twice — the single fastest way to look careless.
- Segmenting on data you do not have. A condition on a field that is empty for 80% of contacts is not a segment; it is a way to accidentally exclude most of your list.
How CampGain handles this
CampGain's segment builder is designed around the two problems above — conditions that stay current, and people who exist in more than one system.
- One segment, multiple sources. A single segment can combine your own contact conditions with Shopify conditions, so "lapsed Shopify buyers who are also on my newsletter list" is one audience rather than two lists you reconcile by hand.
- AND/OR condition groups. Nest rules to express the actual definition instead of approximating it.
- Cross-source deduplication. Members are matched on normalised email or phone, so the same person arriving from two sources is counted once. The segment shows you the duplicate count rather than hiding it.
- Refresh on demand. Re-sync a segment and every source re-evaluates, with member counts and last-synced time visible per source.
- Drill into members. Open a segment and see the actual people in it before you send — the fastest sanity check there is on a condition you just wrote.
Segments then feed straight into campaigns, including multi-channel sequences that reach the same audience over email, SMS, and WhatsApp. Our multi-channel drip campaign guide covers how to sequence those steps, and the campaign analytics guide covers how to tell which segments are actually producing revenue.
Start with three
If this feels like a lot, build three segments this week: engaged contacts, first-time buyers from the last 45 days, and cart abandoners. Attach one campaign to each. Measure revenue per recipient against your usual send-to-everyone baseline.
That comparison — same store, same products, different audience definition — is the argument for segmentation, and you will have it in a fortnight. See how segments work in CampGain when you are ready to build them.