I’ve been the designer on Distribution.AI since day one, across the whole product. This one is about brand voice, and the problem sitting underneath it, which is that getting an AI to sound like you means asking you for a lot of your writing first.
The situation
Making content with AI is easy now. Anyone can do it. The hard part is making it sound like you.
That’s pretty much the whole bet behind Distribution.AI. You take something you’ve already written, whether that’s a blog post, a podcast episode or a YouTube video, and we turn it into posts for every platform you’re on. Create once, distribute forever. We also watch what’s happening on the internet right now and match it against the stuff you’ve already made, so you’re never starting from a blank page.
But none of that matters if what comes out doesn’t sound like you. That was the assumption we built everything on. People can tell, and once they can tell, the whole thing kind of loses its point.
The turn: the voice tax
Here’s the problem we kept running into. To write like someone, the system needs a lot of their writing, so the obvious thing is to just ask for it. Paste in your best posts. Upload some samples. Tell us about your tone. And that’s exactly where people quit.
The better we wanted the output to be, the more homework we had to hand out, and the more homework we handed out, the fewer people ever got to see the output at all.
We built it the obvious way first anyway. Everything you needed to make the AI sound like you went into onboarding, right at the front, before you could touch anything else. On paper it made sense. Get the voice right, and everything after it is good.
Three ways in
The fastest way to get someone’s voice is to go read their site, so that’s the whole first screen. Paste a URL and we’ll handle the rest.
If you don’t have a site, “Try another way” takes you to the other two options, which are writing or pasting some text, or uploading a file. Plenty of people have their best writing sitting in a doc somewhere rather than published anywhere we can reach.
The same flow shows up later too, over in Brand Voices, whenever you want to add another one.
The onboarding was the leak
Here’s the thing we got wrong. All of that sat at the very front, before you could use anything, and onboarding turned out to be where we lost the most people. They were leaving before they’d even tried the thing they signed up for.
So we made it closeable. There’s an X now, and you can shut the whole thing and go straight into the app. Everyone lands on a checklist on the home screen showing what’s left to do, with setting up your brand voice sitting on it alongside connecting your accounts and turning on autopilot. Anything you’ve already finished comes in ticked off, so someone who went through onboarding sees a head start rather than a fresh pile of work, and someone who skipped sees the voice step still waiting for them.
Two ideas from one number
We’d wired up events for every onboarding step in PostHog, so once things had been live a while I went and looked.
Of the first 98 signups, 81 saw the checklist and 15 of them completed a single step on it. That’s the worst drop in the whole funnel by a distance, about 81% at one transition.
So we tried two things at once, coming at the same problem from opposite sides.
The first was making the door optional, which is the X on onboarding and the checklist for anyone who closes it. The second was moving the door back entirely. We put a URL field straight on the landing page, so you can paste a link and get content without an account, and try the thing before we ask you for anything at all.
The one that asks for nothing works better
66 people generated something anonymously, and 8 of them came back and made an account within the week. That works out at 12.1%.
Put that next to the other one. Ask for nothing and roughly one in eight comes back to make an account. Ask for setup first and four out of five walk away from the checklist. Two experiments pointing in opposite directions, and they both land on the same answer, which is that everything you ask for before someone has seen the product costs you people.
Here’s the part that’s harder to sit with, though. The anonymous version doesn’t sound like anyone, and it can’t, because it’s one link with no writing samples and no voice behind it. It’s the most generic output the product can make, and it’s the version converting best.
So we’re trading the actual promise for the chance to be tried at all. I think that’s the right trade and those 8 people say it’s working, but 56 others generated something and never came back, and my honest guess is that the generic output is a big part of why.
I’m reading all of this as a direction rather than a verdict. 98 signups and 66 anonymous sessions isn’t much to go on, and both changes only shipped in May, so there’s no before-and-after to compare against yet.
Splitting rules from judgment
This is the decision the rest of the system sits on. There are two totally different things hiding inside the word “style,” and when you put them in the same place, both of them get worse.
Some things are just rules. Semicolons. Em-dash spacing. En-dashes for date ranges. Ampersands. AP format. US or UK English. There’s a right answer, the system can follow it exactly, and you can check it in a second.
Other things are judgment calls. “Authoritative, but not arrogant.” “Balanced informality.” Nobody can check those at a glance. You only notice when something feels a bit off.
Most tools throw both into one tone setting, and what you end up with is mechanical settings that feel vague and personality settings that feel way more precise than they really are. Keeping them apart meant you could fix something small right away, like turning off exclamation marks and seeing it in the next post, while the judgment side took longer and worked in a completely different way.
Showing the machine’s model of you
The Writing Guideline is just plain writing. It’s not a score or some hidden setting, just a paragraph describing how the system understands your voice, and you can edit it.
Excerpts work the same way. They’re the actual samples we’re learning from, so you can read them, change them, or delete the ones that don’t sound like you.
The rule I kept coming back to is that nothing about your voice should be a guess you can’t see. If we got you wrong, you should be able to find the exact spot and fix it yourself.
Voices are their own thing too, not a setting buried in preferences. You name them, you can keep a few, and you pick a default, which means an agency can hold one per client.
Making feedback visible
Here’s where all of it lands. One thing you wrote goes in, and what comes back is a stack of posts per platform: LinkedIn, X, Facebook, Threads, Instagram, a blog version. You pick the ones worth keeping and regenerate the rest.
Which means the voice question isn’t asked once. It’s asked every single time, across every option on the screen, and you’re answering it constantly just by choosing what to keep.
Every post has a like and a dislike button. One tap, that’s it. The idea being that you never have to sit down and describe your voice in the abstract, you just react to the thing in front of you.
But here’s the part I think matters most. A dislike button that looks like it does nothing is worse than no button at all. Our worry was that people would tap it a few times, see nothing change, and quietly stop bothering. So we made the learning visible.
- Recently Learned
- prefers concise LinkedIn hooks. Less promotional phrasing. More conversational openings. Shorter CTA wording.
- Frequently Refined
- LinkedIn post length. Emoji usage. Hook structure.
That’s the receipt. You tapped dislike a few times last week, and here’s what we picked up from it, written out so you can check it. It closes the loop the onboarding decision opened up. We didn’t want to hand you homework upfront, so the learning had to move into the moment you’re actually using the thing, and then it had to be visible, or you’d never believe it was happening.
There’s a Preview Voice button as well, so you can test a voice before committing to anything with it. Same idea, which is making it cheap to find out you’re wrong.
What I’d do differently
Everything is a toggle. Voice isn’t really on or off, though. “Concise” is more of a dial than a switch, and so is “authoritative.” Toggles were fast to build and easy to understand, and they were the right call for getting something shipped, but they don’t quite match the shape of the thing they’re describing. Weights would get closer, and that’s where I’d take it next.