Today, I'm launching Provv.ai
Three parts. What it is, why it exists, and what it actually looks like. Part two is an argument you can use even if you never touch the product. Part three includes the screenshots where it fails.
Part one: what this is
Right now, on a review site or in a thread somewhere, a person who almost bought your product is explaining exactly why they didn’t. In detail. In their own words. To a stranger who can’t do anything about it.
It’s been sitting there for months. Nobody on your team has read it, because nobody has time to read ten thousand threads.
Provv reads them, and builds you a panel of buyers out of what it finds.
Not personas. Not a model doing an impression of your ICP. Buyers reconstructed from the record: the threads, the reviews, the forum replies, the community posts where your market explained itself to itself while you weren’t in the room.
Then you run your work past them. Copy. Creative. Video. Briefs. Whatever’s about to ship.
Every buyer comes back with a verdict and the exact words behind it. Who bought it, who didn’t, and the specific sentence that lost them. Click any objection and it traces back to something a human being actually wrote. Not the software’s opinion. A position that exists in the record, held by a person.
It drops into Claude through our MCP. Paste the draft, watch the panel react in the thread, rewrite, run it again. Sixty seconds instead of six weeks.
Plays runs it the other way. Instead of judging what you made, our agents read what your market said this week and hand you back a campaign. Channels, copy, a schedule you can steal. Market signal to campaign in minutes.
That’s the product.
It runs on one idea. Your buyers already told you. They just didn’t tell you to your face.
What it costs
I’m opening ten invites. Free. Not a trial, not a discount. Free.
Not for the money. Because I want ten people building it with me while it’s still being decided, and ten is the number of people I can actually be useful to.
Want in? Reply, or DM me on LinkedIn, with the one thing you’re trying to figure out about your buyer right now. The more specific the sentence, the better the panel. I’ll close it when it’s full.
Not one of the ten? Provv Beta is open to everyone.
What it isn’t
It won’t predict your conversion rate. It doesn’t replace talking to customers. It’s useless in markets that don’t talk online, and it’ll tell you when yours is one of them.
Full list of reasons not to buy it is in part three. I’d rather you read that than the feature list.
Part two: why I built it, and how AEO research led me here
The meeting
You know the meeting.
Six weeks of work on the screen. Positioning doc, message house, the headline three people fought about, the proof points, the calendar. Twelve people on the call. It’s good work.
And somebody, usually whoever’s paying, asks the only question in the room.
How do we know this is going to land?
Everyone does the thing. The research supports it. We socialized it with sales. The advisory board loved it. Let’s ship and see.
Nobody says the true sentence: we don’t know, we think so, we’ll find out in three weeks, and even then we won’t really know.
Fifteen years. B2B and B2C growth at Affirm, through the IPO. Self-serve and enterprise growth at Webflow. A global digital marketing team at FIS across a dozen markets. Hundreds of millions in spend.
Different companies. Different motions. Different decades.
Same hole in the middle of every one of them.
Every other function tests. Marketing guesses.
Not an insult. Structural…and the vagueness about it is why nobody fixes it.
Engineering ships to staging first. Tests, canary deploys, feature flags, rollback in ninety seconds. An engineer can be wrong at 2:00 and correct by 2:04.
Finance models the downside before the money moves. Base case, bear case, sensitivity. The whole discipline exists to know what happens when you’re wrong.
Sales walks in with a strategy. Discovery, qualification, a champion, a mutual action plan, a known objection list with a rehearsed answer for each one.
Product puts a prototype in front of five humans before anyone writes production code.
Marketing writes a doc, argues about it, and picks a date.
Then we ship, and three weeks later a number shows up.
The problem isn’t “no data.” It’s the shape of the signal.
Marketing isn’t short on data. Marketing is drowning in it. We have more dashboards than headcount.
The problem is that the signal coming back is slow and flat, and the decision it’s meant to inform is fast and enormous.
Look at what you actually decide when you approve a campaign:
Which problem to lead with
Which of six ways to say it
Which objection to preempt, which to ignore
Which proof goes first
Which channel
Which sequence
What to leave out
Dozens of live variables. And the feedback is one number. Pipeline. MQLs. CTR. Whatever the board looks at.
You can’t take one number apart. When it’s bad, you have no way to know whether the message was wrong, the channel was wrong, the timing was wrong, the offer was wrong, or a competitor announced something on Tuesday. One bit of information about a forty-variable decision, three weeks late, by which point you’ve already committed the next six weeks.
So you do the only rational thing left. You guess, confidently, and then you defend the guess.
That’s not a character flaw. That’s what any sane person does when the loop is that broken. The confidence is a coping mechanism for the delay.
Every marketing org I’ve run or advised has this hole. Most have decorated it with a persona doc.
The detour
I did not set out to build this. I set out to do something else and walked into it sideways.
I spent the last two years deep in AEO. Reverse engineering what the models cite, why they cite it, what you can do about it. That work turned into the Answer Ownership System, Answer Capture Rate, most of what I write and most of what I do for clients.
The core mechanic is easy to say and miserable to execute: if you want to know what the model cites, go where the model learned it.
So that’s what I did. Two years tracing citations backward, into the substrate. Which forums. Which review sites. Which threads. Which communities. Which specific posts shape how a model talks about a category. It’s unglamorous work. It’s mostly reading.
And that’s the thing. It’s mostly reading.
I read for two years. Not summaries. Not dashboards. The actual threads, the actual reviews, the actual replies, because that’s the only way to see the shape of what a model learned.
Somewhere in there, I stopped caring about the citations.
Because I finally looked at what I’d been staring at all day. It wasn’t a citation graph.
It was the most honest record of buyer opinion that has ever existed, sitting in public, unread.
Everything I’d been treating as a means to an end was the end. I’d built a two year habit of reading the places where buyers tell the truth, in order to win a ranking. The ranking was the small prize.
Why that record is honest
Every method we call “customer research” is distorted in proportion to how present the vendor is in the room.
Rank them by that one variable and the whole discipline falls apart.
Sales calls. The vendor is physically there, wanting something. Your buyer is performing and negotiating. They will not say “your pricing page confused me and I felt stupid,” because that’s a status admission to a stranger trying to sell them something.
Win/loss interviews. The vendor is on the phone asking someone to explain a decision that affects the vendor. You get the polite reason, which is never the real reason. “It wasn’t the right time” is what people say instead of “your product annoyed me and your CSM made me feel like a ticket.”
Surveys. The vendor is implied. People pick whichever answer ends the survey fastest.
Customer advisory boards. Selection bias so severe it’s funny. You’ve assembled the people least able to tell you what’s wrong and asked them what’s wrong.
Personas. The vendor wrote them.
Now: your buyer, on Reddit, at 11pm.
No vendor in the room. No relationship to protect. No procurement leverage to preserve. No status to manage, because they’re anonymous. They’re talking to a peer, and the peer is the only audience that gets the truth, because the peer is the only person who can’t do anything with it.
The audience changes the answer.
That’s the insight. Not “there’s data on Reddit,” everyone knows there’s data on Reddit. The insight is that anonymous peer conversation removes exactly the distortions every formal method introduces. It’s the only place your buyer has no reason to manage you.
Your buyers already told you. They just didn’t tell you to your face.
Personas are a compression of what you wish were true
Where does a persona actually come from?
A workshop. Notes from sales calls, which are performances. Four customer interviews, which are your happiest customers, because those are the ones who take your calls. The PMM’s intuition, which is often good and is still one person. A stock photo of a woman in glasses looking thoughtful. A name that alliterates.
A persona isn’t a model of your buyer. It’s a compression of what your company wishes your buyer were. Aspirational by construction, because the people who built it needed it to be true. Written once. Updated never.
That’s not the fatal part. This is:
A persona cannot disagree with you.
It’s a document. Documents don’t push back. You can’t run a headline past a PDF. You can hold your copy up next to Marketing Mary and nod, and Mary nods back, forever, because she’s a slide.
The persona is a mirror you built and then asked for advice.
“Just ask ChatGPT to be my ICP”
I’ll get this objection in every conversation for the next two years.
You can open a chat window right now and say “you are a VP of Engineering at a Series B fintech, react to this landing page.” It’ll produce something articulate and plausible.
It’s close to worthless. Four reasons, none of which require you to touch my product.
1. You get the average, and the average never bought anything.
Ask for a VP of Engineering and you get a blend of everything ever written about VPs of Engineering. Markets aren’t decided at the average. They’re decided at the edge: the specific objection held by the specific fifteen percent who almost bought. The average VP of Engineering does not exist and has never signed a contract.
2. You get your own stereotype handed back to you.
Think about what’s actually been written, at volume, about VPs of Engineering. Job descriptions. LinkedIn thought leadership. And an enormous pile of vendor content about VPs of Engineering, written by marketers, describing the VP of Engineering that marketers wish existed.
So when you ask a model to play your ICP, you’re asking it to blend a pile substantially made of persona docs. You’ve automated the mirror. Same thing as before, faster, and now it talks back in complete sentences.
3. It wants you to like it.
These tools are built to be helpful and agreeable. Correct for an assistant. Catastrophic for research. Ask “would you buy this?” and the rate of enthusiastic yes is absurd. You’ll get “this is compelling, and here are three thoughtful suggestions” for copy that would die on contact with a real market.
The tool is optimized to make you feel good about your work. That is the precise opposite of the job.
4. It has no idea who your buyer is.
“VP of Engineering at a Series B fintech” isn’t a market. It’s a demographic. Your market is the four hundred people with this specific problem, this quarter, at this price, against this competitive set.
No prompt fixes any of this. You can’t write your way out of an average. The only fix is anchoring the thing to a real, current, traceable record of what your actual market actually said.
Which is the thing I’d been building for two years without noticing.
Don’t take my word for it. Take ten minutes.
Stop reading. Do this. It costs nothing and it doesn’t involve me.
1. Open your persona doc. The real one, not the one on the wiki. Find the section called pain points, or challenges, or “what keeps them up at night.”
2. Open a browser. Search site:reddit.com plus your category plus the problem you think you solve. Sort by new, not by relevance. Relevance sorts for what’s popular. New sorts for what’s true this week.
Fair warning: your first page of results will mostly be people selling services in your category. Scroll past it. That’s the disease, and I’ll come back to it in part three.
3. Read ten threads. Not skim. Read. Including the replies, because the replies are where people correct each other, and correction is where the honesty is.
4. Write down the objection that keeps coming up. Not the one you think matters. The one that won’t go away.
5. Go back to your persona doc and find it.
That gap is the company. That gap is the six weeks. That gap is every number that came back bad and couldn’t tell you why.
You just ran the manual version of what I built. Sample size of ten. Ten minutes. Free.
Now try it at ten thousand.
Part three: what it actually looks like
Here’s the thing itself, including the parts I’m not proud of.
Building a panel
Every buyer on that roster is a cluster of real positions. Not a job title with a personality bolted on. The problems they described, the objections they raised, the tradeoffs they made, the words they used to describe the category to each other.
A verdict, and the trace behind it
This is the part I’d judge any tool in this space on, mine included.
When a buyer objects to your headline, you can click the objection and land on the thing a human being actually wrote. It isn’t the software’s opinion. It’s a position that exists in the record, held by a person, in their words.
If you can’t click through to a human, you don’t have research. You have a vibe with a nice interface.
The loop, inside Claude
This is where I actually use it. Not the web app. Claude.
Paste the draft. Watch the panel react in the thread. Rewrite. Run it again. The whole loop lives in one conversation, which means the cost of testing an idea drops to roughly zero, which means you test ideas you’d never have bothered to test.
That’s the part I didn’t predict. When the loop is sixty seconds, you stop defending your first draft, because defending it costs more than checking it.
Plays
Same record, other direction. Instead of judging what you made, the agents read what your market said this week and hand you back a campaign. Channels, copy, a schedule you can steal.
Market signal to campaign in minutes.
What’s underneath: the part nobody wants to talk about
Everything above is worthless if this part gets hand-waved. Take this one and use it against me.
The data is not clean. Anyone who tells you their market data is clean has either not looked at it or is hoping you won’t.
Astroturf. Review sites are gamed. Vendors run campaigns. Leave a review, get a twenty-five dollar gift card. Not fraud exactly, but a brutal selection effect: incentivized reviews skew positive and, worse, skew generic.
What we do: weight by specificity. “Easy to use” carries almost no signal. “The SAML setup took our IT team three weeks and we almost bailed” carries an enormous amount. Specificity is the best available proxy for honesty, because it’s expensive to fake and nobody bothers for twenty-five dollars.
Power-user skew. Reddit is not your market. Reddit is the three percent of your market that posts on Reddit, and that three percent is more technical, more price-sensitive, more contrarian and considerably more negative than your actual buyers.
What we do: say so, loudly. A panel is not a representative sample and I’ll never claim it is. It’s an honest sample. Different instruments, different jobs, and confusing the two is how this whole category discredits itself.
Staleness. A thread from 2021 describes a market that no longer exists. Your category moved. The competitive set moved. Price expectations moved. Old data produces confident buyers with obsolete objections, which is worse than no buyers at all, because it’s wrong with a straight face.
What we do: recency weighting on panel construction. Plays windows to the last week, on purpose.
Contamination. This is the one that keeps me up, and the thing you should interrogate hardest in any tool like this, mine included.
If your competitor’s content marketing gets into the pile, your buyers start sounding like their blog. Vendor content is written to be found, which means it’s overrepresented in exactly the places you’re looking. Scrape naively and you don’t build a panel of your market. You build a panel of your category’s marketing departments, agreeing with each other, in a nice voice.
Here’s how bad it is. While writing this post I picked a well known B2B product and went looking for real buyer conversation about it. Plain searches, the kind anyone would run.
Here’s what came back on page one: a Medium post about the category, four Fiverr gigs, three Dribbble service listings, a Capterra comparison page, two job ads, and a Substack selling a consulting course.
Almost no buyers. Almost entirely people selling services in that category, ranking for the exact words buyers use.
Try it on your own category right now and you’ll get the same thing. That’s not an edge case. That’s the normal condition, and it’s what every naive scrape eats first.
What we do: classify sources, aggressively. Vendor-authored content can inform the category map. It is not allowed to inform a buyer. A buyer can only be built from someone talking to a peer, not to a funnel. This has eaten more engineering time than anything else in the product and it is not finished.
Attribution. Is this person actually your buyer, or a student writing a paper, or a competitor’s SDR, or a bot?
What we do: cluster on problem language, not identity claims. Someone who describes your problem in lived detail is far more likely to be your buyer than someone who opens with “as a CMO.” Stated identity is cheap. Demonstrated fluency in the problem is expensive.
Volume versus honesty. Ten thousand threads sounds impressive. It’s a vanity number. What matters is whether the threads contain a decision: someone weighing something, choosing something, regretting something. A million threads of “interesting, thanks for sharing” is a million threads of nothing.
What we do: grade every panel. Source density, recency, specificity, decision content.
Steal this: five questions to ask any synthetic research tool
Including mine. Especially mine.
1. Where did the data come from, specifically? Not “public sources.” Which ones. If they won’t name them, they’re embarrassed by the answer or there isn’t one.
2. Can you trace one verdict back to one source? Click the objection. Does it lead to a human who said something, or does it lead nowhere? This is the whole ballgame. Everything else is downstream.
3. How old is the median source? Ask for the number. If they don’t have it, they aren’t measuring it, which means it’s whatever the scrape happened to catch.
4. What’s excluded, and why? If nothing’s excluded, everything’s included, and everything contains your competitor’s blog, three SEO farms, and a Medium post a machine wrote in 2023. An honest tool has a deny list and will show it to you.
5. What does it do when it doesn’t know? A tool that always has a confident answer is a tool that’s always guessing. Ask to see it fail. If it can’t fail, it can’t be right either.
That rubric is yours whether you ever touch Provv or not. Use it on me first.
What it can’t do
Plainly, because the limits are what make the rest of it credible.
It won’t predict your conversion rate. Directional, not quantitative. Anyone selling you a synthetic panel that outputs a forecast is selling you a number they made up.
It doesn’t replace talking to customers. It replaces the six weeks where you were talking to nobody. Go talk to your customers. Then use this to work out what to ask them.
It’s useless in markets that don’t talk online. Some enterprise categories are dark. Some regulated industries are silent. Some buyers only speak in rooms with no internet. The grade will tell you, and if the grade is bad, don’t buy it.
It skews negative. People post when they’re annoyed. Complaints are overrepresented in the record and therefore in the panel. Read every verdict with that in your head.
It tells you what people say, not what people will do. Nobody can do the second one. Anyone who claims they can is selling you something, and I’d like to be the person in this category who says that out loud on launch day.
I ran this launch through it
I’m not going to write four thousand words about evidence and then ask you to take my word for it.
I ran my own launch post through a Provv panel before I published anything.
Then I rewrote it. Six times.
The product told me I was wrong. I was.
Is the product perfect yet?
No. That’s the point of building in the wild.
I could have spent another six months polishing behind a curtain until it felt safe, and shipped something extremely good at satisfying me and possibly nobody else. I’ve watched a hundred companies walk into that. You build in private, you launch to applause, and then you find out.
So I’m doing it the other way. It’s in beta, in public, while it’s still being decided, and the people using it get to decide it with me.
I don’t want you to try my product
I want ten people to build it with me.
Ten invites. Free. Not a trial, not a discount. Free.
A cohort. Ten operators chasing something real about their market. You’ll use it. You’ll break it. You’ll tell me the thing I got wrong, and I’ll fix it that week. You get it free, you get the room, you get me directly, the whole way through.
Because the goal here isn’t to blow this up. The goal is to learn.
The unfair advantage isn’t the tool. It’s that your market already told you, and nobody else read it.
Want in? Reply, or DM me on LinkedIn, with the one thing you’re trying to figure out about your buyer right now. The more specific the sentence, the better the panel. I’ll close it when it’s full.
Not one of the ten? Provv Beta is open to everyone.