
Win back the regulars who went quiet.
Mend finds the regulars who are drifting away, has Claude draft a small cashback thank-you, pays it through PayPal, and holds a random few back so you can measure whether it worked.
Now showing in the PayPal sandbox
PayPal Payouts · Log in with PayPal · Planned with Claude
How it works
01 · Opt in
They say yes.
Your regulars join your thank-you list with Log in with PayPal. Mend only pays people who opted in, and anyone can leave with one click.

Join with PayPal
02 · Forecast
Mend sees who is slipping.
A BG/NBD forecast, the standard model for repeat buying, reads each regular's order history and estimates how many orders they will place in the next 13 weeks. Mend leaves the steady and the long gone alone, and nudges the ones still slipping, biggest spenders first.

steady · slipping · long gone
03 · Pay and hold back
Some get cashback. Some don't.
You approve one exact plan. Mend sends the cashback in a single PayPal Payouts batch with your note, and keeps a random group unpaid as the yardstick.

paid · held back
04 · Measure
You learn if it worked.
When customers buy again, Mend compares the paid group with the held-back group and gives you a straight answer, including too early to tell when there are not enough returns yet.

yes · no · too early to tell
The proof
Tested on a real shop's history.
Mend's forecast was scored on the same two years of orders the demo shop replays. Every number on this screen is read live from Mend's database, and it says so when the forecast does not win.
Quiet regulars who came back on their own
Who came back, by forecast group
Better than 'most recent order'?
How much better Mend ranks who comes back than simply sorting by each customer's last order. Each line is the measured difference with its 95 percent range; right of zero means Mend ranks better.
Reproduces the published model
The guardrails
Claude plans. Only you pay.
Mend's AI is Claude, running on Mend's own server. It reads the shop's numbers and PayPal orders, drafts the plan and writes each customer's note. It has no tool that sends money. A payout happens only when a person approves one exact plan.
Claude can
- Read the shop's order history and see who went quiet
- Look up PayPal orders, disputes and products, with keys PayPal refuses for payouts
- Draft a campaign and write each customer's note
- Answer your questions from live numbers
Only you can
- Approve one exact plan, checked by its fingerprint before anything is sent
- Send the cashback as one PayPal Payouts batch
- Cancel rewards nobody claimed and get the money back
- Keep a random group unpaid, so the results stay honest
In code, not in the prompt
Every reward has a cap per customer, per campaign and per day, set by Mend's engine, never by the AI.
Checked before it goes
Every note is checked before PayPal delivers it: no links, no contact details, no claims that PayPal is paying.
Sandbox only
Mend runs on PayPal's sandbox. No real money moves, and the engine refuses to start against live PayPal.
Every rule above can be tested from your browser, against the live app.

No sign-up. A sandbox shop with two years of real order history, real PayPal sandbox payouts and Claude. Plan a campaign, approve it, and watch the money land.
Mend. A win-back engine for shops. Built on PayPal Payouts, Log in with PayPal and PayPal's MCP server. Forecast by a BG/NBD model. Planned and written with Claude. Every payout is held to caps set in code and measured against a random holdout. Sandbox only.