110 People Started Our Trial. One Converted.

One hundred and ten people started our free trial. One of them paid.

That is 0.91%, and the single conversion makes it worse rather than better: he subscribed two minutes and fourteen seconds after starting the trial. He never evaluated anything. He hit a limit, paid to remove it, used the product for one day, and cancelled.

There is a standard playbook for a number like that. Rewrite the paywall. Add urgency. Extend the trial. Test a discount. We did none of it, because we could not answer the only question that mattered: we had no idea why 109 people said no.

So we emailed them.

Why the usual instinct is wrong here

A 0.91% trial conversion is not a copy problem in disguise. Copy problems produce mediocre conversion, not near-total absence. When essentially nobody converts, the likely explanations are more fundamental: the product did not do the thing, the thing was not worth paying for, or the moment of payment arrived when the user had already stopped caring.

You cannot A/B-test your way to that distinction. Every variant you ship is a guess about a mechanism you have not identified, and the sample sizes at our scale would never settle it anyway. Eleven conversations are worth more than eleven experiments when you do not yet know what you are testing.

The other reason we did not touch the paywall first: our data said the paywall was working. Of the people we wrote to, six had consciously tapped "upgrade" and then not bought. They found the offer. They reached the decision. They said no. That is not a discovery problem.

Who we wrote to, and why exactly them

We did not blast the list. We picked the people whose opinion was worth the most.

Of 915 registered users, 18 had run agent tasks on four or more separate days. Removing our own team's accounts left 11 real people, holding 14 accounts between them. Those eleven were, by usage, the entire top of our funnel.

All eleven had started the trial. None had converted.

Two details in that group were worth the email on their own. One person had three separate accounts; another had two. Almost certainly the same human working around a per-account limit. If your users are creating accounts instead of upgrading, that is a packaging finding, not a conversion-copy finding, and it will never show up in a funnel chart.

The two rules we set before writing a word

We did not offer the paid plan as an incentive. It is the obvious move and it would have destroyed the data. We were trying to learn why these specific people had not paid; handing them the paid product removes both their answer and their future purchase. If you need an incentive at all, make it something unrelated to the thing you are asking about.

We did not ask "would you pay $X?" Stated willingness to pay is close to worthless. People are bad at predicting their own purchasing, generous in surveys, and anchored by whatever number you say first. Their behaviour was already recorded. The interview was for the reasoning behind it, not a re-vote.

The email

The instinct is to ask for fifteen minutes. That is the mistake. A call turns the decision into "can I find time this week?", which resolves to "later" and then to never.

We traded the meeting for a single question answerable with one letter:

Hi <name>,

Edgar here — I build CodeAgent Mobile. 1,125 people use it; you're in the top eleven by actual usage. You tried Pro and stayed on free.

I'd like to know why, and I've made it a one-letter reply:

A — didn't really need it once the trial ended B — something broke or got annoying C — not worth the price D — other (one line is plenty)

Just reply with the letter. No call, no pitch, no follow-up sequence.

Edgar

P.S. You ran 2,776 tasks over 7 days — more than almost anyone who has ever used this. Whatever stopped you short of paying is the most useful thing I could learn this month.

Four things in there are doing the work.

The ask is five seconds. Multiple choice removes the blank-page problem. Anybody can hit reply and type B.

The harsh option is named out loud. People do not reply because they do not want to be rude. Writing "not worth the price" yourself makes it not rude.

The postscript carries the personal detail. It is the most-read line in any email, and the real usage numbers are what prove this is not a blast.

The subject line is a question. Questions get answered. Announcements get archived.

What we deliberately left out: fake urgency, a Re: on a conversation that never happened, account-consequence framing, guilt. Every one of those would lift the reply rate. Every one would also spend the goodwill of the eleven people whose goodwill is the scarcest thing we have — and a manipulated reply is not honest reasoning, which is the entire product of the exercise.

What came back

The first two replies pointed in completely different directions, and neither one was about price.

A paying customer wrote:

"I stopped using the service because I no longer needed it. I have already canceled my subscription."

That is project-shaped usage. He had a job to do, the product did it, and he was done. His whole trace is six seconds from trial start to purchase, 79 tasks, and silence the next day. He is the paying version of the dominant pattern in our data: a third of everyone who ever used the product used it on exactly one day.

If that is what the product is for, a monthly subscription is fighting the natural shape of the work. That is a packaging question. No paywall variant addresses it.

The second reply came from our heaviest user by volume:

"I 'ran out of credits' whenever I began a task, and eventually I could not even connect to the workspace. When it functioned, I had to guess the results because I could view only my own text."

We checked her account against our logs before drawing any conclusion. Two of her three complaints matched exactly: nine failed workspace deploys across two days, and a daily task limit that must have been maddening for someone working at her pace. Her rage-clicks ended at 01:05:07; her last failed deploy was at 01:05:34. Same minute.

Her third complaint we could not corroborate — our records show output was being delivered to her app. We wrote that down as unconfirmed rather than dressing it up, because the most likely explanation connects to the first two: when the workspace never starts, your messages still send and there is simply nothing on the other end to answer.

What we did with it

We stopped working on the paywall and spent the week on deploy reliability.

That was not a hunch. Our own funnel said 57.8% of users who chose a deploy target never completed a deploy — and the reason that number had gone unnoticed for so long is that the failures were silent. The wizard swallowed its own errors and rendered a blank page. No error, no event, nothing to see in analytics. Just people quietly leaving.

Two customers had now told us in plain language that the product had not worked for them, and the data agreed. Against that, tuning upgrade copy would have been theatre.

If you are sitting on a number like ours

Write the email. Today, not after the next release.

Pick the ten or twenty people whose opinion is worth the most — usually your heaviest users, not your newest. Make the ask small enough that replying is cheaper than ignoring. Name the uncomfortable answer yourself. Do not bribe them with the thing you are asking about. Send one follow-up, four days later, and then stop; the silence is data too.

Then read the replies for the words they use, not the letters they pick. Ours told us the problem was reliability and shape-of-use, in a business where every instinct and every growth article pointed at pricing. That correction was worth more than a quarter of experiments, and it cost eleven emails.