I checked who answers questions in six developer communities. In 94 threads out of 300, nobody did.

We say it on our own home page. If most replies come from your team, you are running a support queue with a nicer name. I believed that. Then I went to check it, because a sentence that confident should survive contact with data.

It survived. Something else did not.

What I pulled

Six projects: Next.js, Prisma, Supabase, Tailwind CSS, Vite, Astro. Last 60 discussion threads from each, through the GitHub API. Astro turned out to have discussions switched off at the repository level, which the API says plainly and which I should have checked before adding it to the list, so it contributes nothing at all to anything below. Five rooms left. Three hundred threads.

Then, per thread: did anyone besides the poster show up, who got there first, and what does GitHub say that person is. The last part is free. Every comment carries an authorAssociation, so I did not have to guess at anyone’s role.

Members get there first 77.2% of the time, staff 22.8%. So the sentence on our home page survives, and I could have stopped there and written a short smug post about being right.

The number that stopped me

Ninety-four threads got no reply at all. Not slow, not thin, not unhelpful: nothing, from anyone, ever, in rooms where thousands of people are standing around.

That is 31.3%, in five projects that all have large user bases and people paid to be there.

So change the number you report

Stop reporting reply share. Report silence rate instead. One number, two minutes to pull, and it is the only one that maps to a person who left.

Then set a floor instead of a target. Not “answer everything”, because no room on my list manages that and the two nearest to it are the staffed ones. A floor is smaller. Nothing sits 24 hours without a human saying something, and “nobody knows yet” counts as something. Prisma seems to run on a version of this already: three dropped threads out of sixty is not luck.

Mark things answered, too. Only 19.7% ever get marked. Four out of five solved problems leave no trace, which is why the same question comes back in six weeks and somebody types the same reply again.

Why the average was hiding it

Tailwind is the room our own sentence condemns. Staff got there first 28 times out of 45. A support queue, by our rule. It is also the place where the person answering can go and change the code that afternoon, which our rule never thinks to ask about.

Prisma dropped three threads out of sixty. Next.js dropped thirty-four. Same platform. Same year. Roughly the same kind of person typing the question. Whatever explains an eleven-fold gap in being ignored, it is not the software they are all using.

Time, which nobody reports either

Median wait for a first reply: 4.8 hours. Fine.

The ninetieth percentile is 114.7 hours. Five days, near enough. Nobody waits five days. They solve it somewhere else, or they drop it, and in both cases you never find out which one happened or what it cost you.

And yet 25.7% of replies land inside an hour. Quick. Glacial. Same room, same week, and from where the asker sits it looks like a coin toss, which is a worse thing to design around than plain slowness.

I spent an afternoon trying to correlate silence rate with repository size, star count and how many people had commented that month. Nothing lined up. It was probably the wrong afternoon to spend, and I mention it only so nobody repeats it.

The bit I could not settle

Supabase had zero staff-first replies in sixty threads. Zero.

Either that community genuinely runs itself, or the staff arrive later and my method never sees them, because I only ever recorded who was first.

I do not know which. It is a different measurement and I have not run it: not who speaks first, but who ends the thread.

Questions people ask about this