Your orders
Leave margin at 100 to value revenue; enter your gross margin (for example 45) to value profit. Amounts should be in one currency. Orders on the same day count as one purchase; refunds and zero amounts are ignored.
Upload or paste your orders, then click Predict.
Result
What a new customer is worth
Does it pay back?
Running total of the new-customer value above against the acquisition cost you entered, at your margin and discount rate. The 3 : 1 line is a common rule of thumb, not a law of nature.
Customer segments
| Segment | Customers | Predicted value | What to do |
|---|
Cohorts: who comes back, by month of first order
Share of each month's new customers who bought again within 30, 90, 180 and 365 days. A dash means some of that cohort has not been around that long yet, so the cell is left blank instead of understating it.
Back-test: predicted vs actual purchases in your last weeks
| Repeat orders before | Customers | Predicted | Actual |
|---|
Highest predicted value
| Customer | Repeat orders | Still active | Avg repeat order | Predicted value | Segment |
|---|
Fitted model
How it is calculated
Frequency and recency: for each customer, the number of repeat purchase days, the time between first and last purchase, and the time since first purchase, in weeks. BG/NBD (Fader, Hardie and Lee, 2005): while active a customer buys at a steady random rate that differs between customers (gamma distributed); after each purchase they may become inactive for good (beta distributed). The four parameters are fitted by maximum likelihood (Nelder-Mead). Expected future purchases and the probability a customer is still active follow from the published closed forms. Gamma-Gamma (Fader and Hardie, 2013): average repeat order value varies between customers around a gamma distributed mean; predicted order value blends each customer's own average with the population's, so a customer with one repeat order is pulled toward the average. Value: expected purchases in each future month times predicted order value times your margin, discounted monthly. Customers with no repeat order get the population's average repeat order. Back-test: the model is refit on data ending some weeks before your last order (8 to 26 weeks depending on history) and its predictions are compared with the purchases that followed.
Not modelled: subscriptions with fixed billing, one-off purchases, seasonality, promotions, price changes, returns, customers who pause and come back, and any link between how often someone buys and how much they spend. Unverified: checked against simulated data with known parameters and closed-form identities in unit tests, not against a published customer dataset or a commercial tool.
Common questions
What is customer lifetime value and how do I predict it?
Customer lifetime value (CLV) is the total margin a customer will bring you over their relationship with you. For a shop where customers come back at irregular times, the standard way to predict it is a probabilistic model: BG/NBD estimates how often each customer buys and how likely they have already stopped, and Gamma-Gamma estimates what they spend per order. This tool fits both to your own order list in your browser and multiplies them, month by month, with discounting.
What data do I need?
One row per order with a customer identifier (an id or an email), the order date and the order amount. A Shopify, WooCommerce, Stripe or spreadsheet export works if it has those three columns; the tool finds them by header name. Refunds and negative amounts are dropped. Six months of history is the practical minimum and a few hundred customers gives stable totals.
Why not just use average order value times average orders per customer?
Because that average hides who is still buying. A customer who ordered five times and then went quiet for a year looks valuable in a spreadsheet but is probably gone. The BG/NBD model uses recency as well as frequency, so it separates customers who are between orders from customers who have left, and it will not credit the lost ones with future revenue.
How accurate is the forecast?
The tool shows its own report card. It refits the model on your data up to some weeks before your last order, predicts the purchases in those final weeks, and compares with what really happened, in total and by number of past repeat orders. On data that follows the model's assumptions the total is typically within a few percent. If the back-test misses badly, a promotion, a price change, seasonality or a new channel is usually why, and you should trust the forecast less.
What is the model not good at?
It assumes each customer's buying rate is steady, that customers stop for good rather than pausing, and that order value is unrelated to how often they buy. Subscriptions with fixed billing dates, one-time purchases such as furniture, strong seasonality and businesses that change a lot during the window are poor fits. The tool warns when spend and frequency are correlated or when history is short.
What do I do with the segments?
Customers who spent a lot but look unlikely to still be active are the cheapest people to try to win back. Customers who look lost should come out of expensive retargeting audiences. The predicted value of a new customer is the ceiling for what you can afford to pay to acquire one. Download the per-customer CSV to build these lists in your own tools.
Is my customer data uploaded anywhere?
No. The file is read and analyzed in this browser tab. Nothing is sent to a server, there are no accounts or cookies, and the tool never stores your orders. Customer emails in your file stay on your device.