The accept-rate breaker

Here is the failure, told as it happens. You write an ideal customer profile and it is a little too wide. Not wrong, just generous: Head of Sales at any B2B company, say, when the people who actually buy from you are Heads of Sales at thirty-person SaaS companies in their second year. The tool starts sending. The first week looks fine, because nobody has had time to not accept yet. The second week, acceptances come in slower than you expected and you tell yourself it is the season.

By the third week there are a hundred and forty invitations sitting pending on profiles that were never going to say yes. LinkedIn's help page lists exactly this as a reason to restrict: many of your invitations have been ignored, left pending, or marked as spam. The notice arrives. The restriction lasts a week, or a month if the pending pile is the reason, and withdrawing them does not help. You lost the account for a month because of a sentence in a text box three weeks earlier.

What makes this one dangerous is that nothing in it looks like a mistake while it is happening. Volume was fine. Timing was fine. The targeting was slightly off, and slightly off at twenty a day is a hundred pending by the time anyone notices.

So the product watches the number LinkedIn watches

Every day, for every account, the breaker looks at the invitations that have had a fair chance to be answered and asks what fraction were accepted. If it is below 15%, the daily cap halves. If it is below 8%, inviting stops entirely, the queued invitations are cancelled rather than paused, and you are told why, in the digest and on the dashboard, with the actual numbers: only so many of your last so many were accepted.

Messages keep going in both cases. Somebody who already accepted is waiting for the message we said we would send, and cutting them off would punish the people who did want to hear from you for the behaviour of the ones who did not.

The bug in the first version, and why a sample matters more than the thresholds

The first version had no minimum sample. On the morning an account started, with three invitations out and no acceptances yet, it read zero percent and stopped inviting before anyone had had time to accept anything. So now the settled check needs at least 20 invitations that are each 7 days old. An invitation sent this morning is not an ignored invitation; it is one nobody has opened yet, and counting it would make every healthy account look like it was failing, most severely on the days it sent the most.

The blind window, and the second tier

That fix created a different problem. Walk the ramp: 5 a day for three days, then 10. The twentieth invitation goes out around day three, matures a week later around day ten, and by then the ramp has sent about a hundred. For the first ten days of an account's life the settled breaker cannot fire at all, which is precisely the window in which a bad profile does its damage.

So there is a second, blunter check. From day four, it looks at 10 invitations that are 3 days old rather than seven, because most invitations that get accepted are accepted inside forty-eight hours, and the question on day four is not what this account's true rate is but whether anything is being accepted at all. One of ten halves the cap, which is cheap and reverses itself. Zero of ten stops inviting. Against a normal 25 to 30 percent accept rate, none of ten is an alarming observation rather than bad luck. The thresholds are 20% and 10% rather than 15% and 8%, and the settled tier takes over as soon as it has the sample, because more evidence beats faster evidence once both exist.

Why the percentages are mine, not LinkedIn's

LinkedIn publishes no accept-rate threshold. It says ignored invitations are a reason and gives no number, which is on the limits page with the source. So these thresholds are chosen, and chosen to fire early: throttling a healthy account for a day costs almost nothing, and not throttling a doomed one costs the customer the account. The asymmetry is the whole design.

How often it has fired, across every customer, is on the numbers page, as two rows: the times it halved a cap and the times it stopped an account. A tool that sells restraint should be able to show itself refusing to send, and those two rows are that.