Servings in this dataset:
7 in 10 first-time buyers never come back
How fast customers come back
Share of customers who have placed a second order, by days since their first order.
Monthly cohorts
Share of each month's new customers who ordered again in month 1, 2, 3… after joining.
Which products, offers and prices bring people back
90-day repeat rate by what happened on the first order. Red bars beat the average (dashed line), grey bars fall below it. Whiskers show the 95% confidence interval, so bars whose whiskers overlap may not really differ.
Price paid vs discount depth
The repeat rate climbs for mid-priced first orders and falls sharply when the first order was more than 40% off. Cheap trials and heavy discounts bring in buyers who don't stay.
Spot who won't come back before they disappear
A machine-learning model trained on first-order data (channel, product, offer, price, delivery time, day-7 rating and demographics) predicts each customer's chance of ordering again within 90 days. It scores recent buyers so the team can act while there is still time.
What the model relies on
Permutation importance: how much the model's AUC drops when one input is shuffled. A bigger drop means the input matters more.
Does the score match reality?
Customers in the test set ranked into 10 equal groups by predicted repeat chance. Bars show what actually happened.
At risk right now
Why they leave, in their own words
customers left a review; of them never ordered again. Each review from a one-time buyer is tagged with its main reason using keyword rules, then sized. The fix for each reason comes from the driver analysis and the nudge experiment.
Reasons one-time buyers give
Read the reviews
Which nudge wins the second order
On day 21, customers who hadn't reordered were randomly sent one of four nudges, or nothing. Because the groups were random, the difference from "nothing" is the nudge's real effect, and it differs a lot by segment.
The plan, segment by segment
For each segment, the nudge with the biggest measured uplift, plus the product, price and message to pair it with. "Proven" means the uplift is larger than its 95% interval; "Test first" means it could still be noise.
Uplift in 90-day repeat rate vs no nudge
Percentage points, ± the 95% interval. The best nudge per segment is outlined.
Segments
How AI was used
The AI in this tool is a machine-learning churn model that predicts which first-time buyers won't return. The tool itself (data simulation, analysis code and this dashboard) was built with Claude, Anthropic's AI assistant, through Claude Code.
5,000 customers, 8,700 orders and 1,650 reviews, with known causes built in so results can be checked.
Random forest vs logistic regression. The better one by test AUC scores every customer.
Keyword tagging of reviews, and uplift with confidence intervals from the randomised nudge test.
Claude wrote the simulation, the model pipeline and this dashboard from the project brief.
Planned but not built: live Claude features
The original design had four features that call Claude live from the page. Due to API issues they were not implemented in this version:
- Review theme discovery. Claude would read the reviews, find the themes on its own (without a keyword list) and suggest a fix for each. Its themes would be scored against the known causes.
- Win-back plan writer. Claude would turn the nudge experiment into a plan per segment, citing the numbers behind each recommendation.
- Ask the data. A chat where Claude answers plain-English questions by running filters and group-bys on the customer table.
- Personal win-back messages. A WhatsApp message drafted for each at-risk customer from their risk factors.
In their place, the review reasons use keyword rules and the win-back plan is written from the experiment results.
Model card
Checking the results
What is real and what is simulated
- The data is simulated. Products and prices are approximate, taken from public listings for illustration. None of it is SuperYou's real data.
- The simulation has known patterns built in: heavy discounts and trial packs lower repeat rates, variety packs and quick commerce raise them, and slow delivery and poor taste hurt. Because the true answer is known, the analysis can be checked against it.
- The model, the charts and the experiment results are real computations on that data.
Limits
- An AUC near 0.69 is a useful ranking but a poor yes/no verdict. Use the score to decide who to contact first, not to write anyone off.
- The reviews come from templates, so keyword rules work better here than they would on messy real reviews. That gap is what the planned Claude theme discovery was meant to close.
- Each nudge segment has about 90 to 250 customers per arm, so some uplifts are within noise. Those are marked "Test first".