Predictive analytics for ecommerce: what it predicts, what it needs, and where it misleads
A good model tells you who's about to buy. A good marketer asks the next question: would they have bought anyway?
Key takeaways
- Predictive analytics in ecommerce uses each customer's order history to estimate what they'll do next: buy again, spend more, or stop.
- Shopify, Klaviyo and Google Analytics already offer some predictions, each with its own minimum data and time window.
- A prediction is only worth making if it changes an action: who gets an offer, when, and how much you spend.
- The customers most likely to buy are often the worst place to spend a discount, because many would buy anyway.
- Check any model by backtesting it on a past date, then measure the campaigns it drives against a holdout.
Predictive analytics for ecommerce uses your store's order history to estimate what each customer will do next: buy again soon, spend more, or quietly stop. Instead of reporting what happened last month, it scores every customer on what's likely to happen next, so you can change who gets which message, when, and at what cost. This guide covers what ecommerce predictive analytics can predict, what the tools you already use offer, how much data you need, how to tell whether a model works, and the mistake that turns good predictions into wasted discounts.
What is predictive analytics in ecommerce?
Most store reports look backward. Revenue last month, repeat rate last quarter, average order value since launch. Predictive analytics looks forward, one customer at a time.
The inputs are mostly things you already have: when each customer ordered, how often, and how much they spent. That's no accident. The best predictor of what a customer will do next is what they've already done, which is why the classic models in this field are built on order history.
The best-known example is the BG/NBD model, published by Peter Fader, Bruce Hardie and Ka Lok Lee in Marketing Science in 2005. Its abstract describes managers as "very interested in predicting the future purchasing patterns of their customers" and notes the model's parameters can be obtained quite easily in Microsoft Excel. Its core idea, predicting each customer's future purchases from their past ones, is the basis of every prediction in this guide.
Think of predictive analytics as a forecast for each customer rather than one for the whole store. The question to ask of any prediction is which decision it will change.
What can an online store predict?
Five predictions cover most of what stores use day to day:
| Prediction | The question it answers | The decision it changes |
|---|---|---|
| Customer lifetime value | How much will this customer spend from here? | How much to spend keeping or acquiring customers like them |
| Purchase probability | How likely are they to buy in the next few weeks? | Who goes in this week's campaign, and who to leave alone |
| Next order date | When is their next order due? | When to send the reminder |
| Churn risk | Are they drifting past their usual rhythm? | Who needs a win-back before they're gone |
| Product affinity | What are they likely to buy next? | What to recommend |
Each one has its own guide on this site. The customer lifetime value guide covers value, and the customer churn guide covers drift and early warning signs.
Don't start by asking which predictions you can get. Start with the decision you make most often, such as who gets the weekly campaign, and pick the one prediction that improves it.
What predictions do Shopify, Klaviyo and Google Analytics already give you?
You may already have predictive analytics switched on and not know it. The three tools most stores run each offer something, with different minimums and time windows:
| Tool | What it predicts | Minimum data | Source (checked) |
|---|---|---|---|
| Shopify | Predicted spend tier: high, medium or low, from how often a customer buys, their average order compared with the store's, their number of orders and how recently they ordered | More than 100 sales; only customers who have purchased are included | Shopify Help Center (Nov 18, 2026) |
| Klaviyo | Predicted lifetime value, expected date of next order, churn risk and more | For lifetime value: at least 500 customers who have ordered and at least 180 days of order history, among other conditions | Klaviyo Help Center (Nov 17, 2026) |
| Google Analytics 4 | Purchase probability (next 7 days), churn probability (next 7 days) and predicted revenue (next 28 days) | At least 1,000 returning users who triggered the condition, and 1,000 who didn't, over a seven-day period in the last 28 days | Google Analytics Help (Nov 18, 2026) |
Shopify's predicted spend tier arrived as a segment filter in its March 2022 changelog. Google says its predictive metrics are generated for each active user once per day, and that predictions stop updating if model quality falls below its threshold.
These tools answer different questions. Google's predictions are about website visitors over the next week. Shopify's sorts buyers into three broad tiers. Klaviyo's live on email profiles. None of them, out of the box, tells you whether the campaign you send on the back of a prediction caused anything. That gap is the subject of the next section.
Check which of these you already qualify for before buying anything. Then note what each prediction can't do.
Why the likeliest buyer is often the wrong target
Here's the trap. You get a purchase probability for every customer. You send your best offer to the top of the list, because they're the most likely to buy. The campaign report looks great, because they bought. And you lost money.
The customers a model ranks most likely to buy are often the ones who need you least. A prediction of who will buy is not a prediction of who you can change. Researchers call the second question uplift modeling: modeling the incremental impact of an action on a customer outcome, as Gutierrez and Gérardy's 2017 review puts it. The survey by Zhang, Li and Liu lists online marketing among the fields where this matters.
A worked example makes it concrete. Take an illustrative store with 8,400 customers and an $86.80 average order, and assume a 45% gross margin. It sends a 15% off code to two groups of 840 customers (10% of its customers each). The buying rates are assumptions for the example, not benchmarks.
| Top group by purchase probability | Middle group | |
|---|---|---|
| Customers | 840 | 840 |
| Would buy in 30 days anyway | 30%, or 252 orders | 10%, or 84 orders |
| Buy with the code | 33%, or 277 orders | 17.5%, or 147 orders |
| Orders the code caused | 25 | 63 |
| Margin with no code | 252 × $39.06 = $9,843.12 | 84 × $39.06 = $3,281.04 |
| Margin with the code | 277 × $26.04 = $7,213.08 | 147 × $26.04 = $3,827.88 |
| Effect of the code | −$2,630.04 | +$546.84 |
The math: margin per full-price order is 45% of $86.80, or $39.06. The code takes 15% off, or $13.02, leaving $26.04 of margin per discounted order.
The top group's report would say 277 orders and about $20,437 in discounted sales. In fact the code cost $2,630 in margin, because it went mostly to 252 people who were buying anyway. The same code to the middle group made money.
A purchase probability tells you who will buy. It doesn't tell you who your offer will change, and those are often different people.
The discount addiction guide shows how this compounds when it becomes a habit.
Use purchase probability to decide who to leave alone, or who gets a reminder at full price. Spend discounts where a holdout shows they change behavior.
How much data does predictive analytics need?
The minimums in the table above are a good reality check. Every tool asks for some mix of enough customers, enough repeat buyers, and enough history to see a full buying cycle. LiftSignals, for comparison, needs about 500 customers and 90 days of orders, from Shopify, another store platform or a CSV file.
Volume matters because models learn from contrast. To learn what a lapsing customer looks like, a model needs plenty of customers who lapsed and plenty who didn't. That's why Google asks for 1,000 of each. A store with 300 customers and 40 repeat buyers simply hasn't produced enough examples yet.
History matters because of rhythm. If customers usually reorder every 60 days, a model trained on 45 days of orders has never seen a typical reorder. For products bought once or twice a year, you need well over a year of history before predictions settle down.
Clean data matters more than either. Test orders, staff orders, fully refunded orders and duplicate customer records all teach a model things that aren't true. Cleaning them out is dull work that pays off more than any model choice.
Below the minimums, use simple rules: days since last order compared with each customer's usual gap. They'll take you most of the way until you have enough customers to beat them.
How to check whether a prediction is any good
Vendors love the word "accuracy". It's the least useful word in this guide, because a model that predicts "nobody buys next week" is accurate for most customers at most stores and useless for all of them.
Two checks matter instead.
1. Backtest it. Score customers as of a past date, say 90 days ago, using only data from before then. Then look at what actually happened. Sort customers into ten equal groups by score and count real buyers in each. A useful model puts many more of the actual buyers in its top group than its bottom one.
For example, in the illustrative store, suppose 1,000 of its 8,400 customers ordered in the 90 days after the backtest date. If ranking were random, each tenth of customers would hold about 100 of them. A model that puts 300 of the 1,000 in its top tenth has found three times as many buyers as chance. That's a useful model. One that finds 120 is barely better than guessing.
2. Test the action, not the score. A model can rank customers well and still drive a campaign that causes nothing, as the discount example showed. Run each campaign built on a prediction against a random holdout. The email holdout test guide shows how to set one up in Klaviyo and how big it needs to be.
Also check calibration where it matters. If a model says a group has a 30% chance of buying, roughly 30% of them should buy. A model that ranks well but overstates every probability will inflate any forecast built on it.
Ask any vendor, including us, for a backtest on your own data. A model you can't check is a horoscope with better formatting.
Where predictive analytics breaks down
Predictions are estimates from patterns. When the patterns change, so do the errors.
- New stores and new customers. One-time buyers have no rhythm of their own. Their predictions lean on store-wide averages and are rougher.
- Seasonal products. If customers only buy before winter or before school starts, a model needs several seasons of history.
- Big changes in the business. A new product line, a price change or a large promotion can make last year's patterns a poor guide for months.
- Feedback loops. If you only ever email high scorers, they buy more because they got the email, and the model learns that high scorers buy. Holdouts break the loop.
- Customers you can't see. Shoppers who also buy from you on marketplaces or in shops can look lapsed online when they aren't.
None of these are reasons to skip predictions. They're reasons to check them against a holdout instead of trusting them on faith.
Rerun your backtest after any big change to the business, and keep a holdout on every campaign that uses a score.
In LiftSignals
LiftSignals connects read-only to your store, scores every customer every night for lifetime value, churn risk, purchase propensity and segment, and sends those audiences to Klaviyo, Meta, Google and other tools. Each send keeps a random holdout, so a campaign built on a prediction shows the revenue it caused, not just the revenue it was near.
What to do next
You now know what predictive analytics for ecommerce can predict, how much data it needs, how to check a model, and why the likeliest buyer isn't always the best target. The practical next step is a score you can act on every week. Purchase prediction estimates every customer's chance of buying soon, so you can send reminders to those who are due, keep discounts away from those who'd pay full price, and measure each campaign against a holdout. Connect your store and see your own customers ranked during the 30-day free trial. No card, and no crystal ball required.
Sources
- Fader, P. S., Hardie, B. G. S. and Lee, K. L., "'Counting Your Customers' the Easy Way: An Alternative to the Pareto/NBD Model," Marketing Science 24(2), 2005. https://econpapers.repec.org/RePEc:inm:ormksc:v:24:y:2005:i:2:p:275-284
- "Predicted spend tier," Shopify Help Center (checked Nov 18, 2026). https://help.shopify.com/en/manual/customers/customer-segmentation/predicted-spend-tier
- "Predicted spend tier," Shopify Changelog, Mar 2022. https://changelog.shopify.com/posts/predicted-spend-tier
- "Understanding Klaviyo's predictive analytics," Klaviyo Help Center, updated Oct 2026 (checked Nov 17, 2026). https://help.klaviyo.com/hc/en-us/articles/360020919731
- "[GA4] Predictive metrics," Google Analytics Help (checked Nov 18, 2026). https://support.google.com/analytics/answer/9846734
- Gutierrez, P. and Gérardy, J.-Y., "Causal Inference and Uplift Modelling: A Review of the Literature," Proceedings of Machine Learning Research 67, 2017. https://proceedings.mlr.press/v67/gutierrez17a.html
- Zhang, W., Li, J. and Liu, L., "A unified survey of treatment effect heterogeneity modeling and uplift modeling," arXiv, 2020 (revised 2021). https://arxiv.org/abs/2007.12769
Predictive analytics ecommerce questions
What can predictive analytics tell an online store?
It estimates what each customer is likely to do next from their order history: how much they'll spend, how likely they are to buy soon, when their next order is due, and whether they're drifting away. The useful part is acting on it, such as timing a reminder or keeping discounts away from customers who would pay full price.
Is my store too small for predictive analytics?
Possibly. Built-in tools set minimums: Klaviyo's lifetime value predictions need at least 500 customers who have ordered and 180 days of order history, and Shopify's predicted spend tier needs more than 100 sales. Below that, simple rules based on time since last order usually work as well.
What's the difference between a purchase probability and uplift?
Purchase probability estimates how likely a customer is to buy. Uplift estimates how much a specific action, such as an email or a discount, changes that likelihood. A customer can have a high purchase probability and almost no uplift, because they were going to buy anyway.
How do I check whether a customer prediction is accurate?
Backtest it. Score customers as of a date in the past, then see who actually bought in the following weeks. A useful model puts far more of the real buyers in its top group than in its bottom group. Then test the campaigns built on it against a random holdout.
Put it to work: Purchase prediction → · Predictive customer lifetime value → · Sales forecasting →
Keep reading
All predictive analytics guides →Start here · Customer lifetime value
Customer lifetime value: what it is, how to calculate it, and how to raise it
The number that tells you how much a customer is worth, and therefore how much you can afford to spend winning and keeping them.
5 min read · Updated Nov 15, 2026
Start here · Churn & retention
Customer churn in ecommerce: how to measure it, spot it early and reduce it
Online shoppers don't cancel. They just stop coming back. Here's how to tell who's leaving before they're gone.
4 min read · Updated Nov 15, 2026
Start here · Email & Klaviyo
Email holdout test: how to prove what a Klaviyo flow actually earns
Your email platform says the flow made $9,000. A coin flip and a little patience will tell you whether it made $2,600.
11 min read · Updated Nov 18, 2026
Know who'll buy, and who needs a nudge
Score every customer's chance of buying soon, then send the right ones to your email and ad tools with a holdout on each send.