I was reading 44 call transcripts for a completely different reason when I noticed it.
Four of those founders had told me, unprompted, that they found us by asking an LLM. One named ChatGPT. Three named Claude. Not one of the four came from a campaign we ran.
Nothing in our reporting had recorded a single one of them.
That's fifteen working days, between 20 July and 11 August 2026. It sits inside a bigger sample, I trust. Between 1 June and 13 August, there were 71 T2D3 Growth Workshops booked and not canceled, across 67 different companies. Mostly founders and chief executives in the room.
Four out of 44 isn't a market study. I'm the only person who counted it, and I'll say that plainly. But it's a channel, and here's why I'm comfortable calling it one. It was producing qualified pipeline for weeks before it appeared anywhere in any report.
Nothing in our process asks the question, and there's no field for the answer.
Our attribution can tell you a contact came in through the workshop application form. That's true, and it's useless, because the form is the door. It isn't the road.
The road was a founder typing a question into a model and getting a shortlist back with us on it. Nothing captures that. There's no referrer, no link to trace, no click on anything we own. The buyer arrives already knowing who you are, and the only record of how is a sentence they might say out loud in the first three minutes.
Here's the part that should bother anyone building an AI-heavy revenue stack right now. A model is very good at finding a pattern in data you've already collected. It can't tell you about the thing nobody wrote down.
I didn't find this with a tool. I found it by reading transcripts, and I wasn't even looking for it. Every AI system you build sits on top of inputs that a human decided to capture, and the inputs nobody captured are invisible no matter how good the model is.
In the same few weeks, a prospect's head of sales told me he runs two or three AI filters over his own mailbox and doesn't see cold email at all anymore.
Not deletes it. Doesn't see it. The filtering happens before anything reaches him, and he described it the way you'd mention a spam folder. Completely ordinary to him.
Put those two facts side by side, and they're really one fact. Buyers are putting a model between themselves and the noise coming in, and they're asking a model for a shortlist when they go looking. The same technology is closing one path to your buyer and opening another one, at the same time, to the same person.
Almost every company I talk to is spending on the path that's closing and nothing on the one that's opening. That isn't stupidity. One of those channels has been measurable for fifteen years, and the other one doesn't have a dashboard yet, so of course the budget sits where the reporting is.
The gap between a channel working and a channel being visible is where the whole advantage sits. It's uncomfortable, because it means the best thing happening in your pipeline right now is probably the thing you can't see. There's no tool you can buy this quarter that will show it to you.
Three things, and the order matters, because the first one is free and the third one is where most people want to start.
Ask the question out loud, in the first few minutes. Not on a form, where the answer is a dropdown of options you thought of last year. In the conversation, where the answer can be something you weren't expecting. "How did you come across us?" is a fair question to ask a founder who already booked time with you. The answers worth having are the ones that don't fit any of your categories.
Then give the answer somewhere to live. One field, filled in by hand, is enough to turn four anecdotes into a count. It doesn't need to be elegant. What kills a finding like this isn't crude tracking; it's no tracking, because then every instance evaporates when the call ends and the pattern never assembles.
Only then work on the optimization. Answer engine work is real, and it's worth doing. It's the third step and not the first, because until you can count the thing, you have no way to know whether the work moved it. Doing it in the other order gets you a project with no scoreboard, and that's how most marketing budget quietly disappears.
Notice that none of those three is a purchase. The whole sequence costs one question and one field, and it runs on the conversations you're already having.
A four-count off one seller's calendar isn't a competitive advantage that survives being held quietly. It's an early reading, and early readings get more useful when other people check them against their own rooms.
So if you're running a B2B SaaS or AI company and you're having these conversations too, count yours. Ask your next ten prospects how they found you, and write the answers down somewhere they'll still be in a month.
If a model referred any of them, you now know something about your own demand that your reporting doesn't. Nobody is going to hand you that number. The founders who have it in six months will be the ones who started writing it down before there was a dashboard for it.
Four out of 44, in fifteen working days. That number will be different by the time you read this, and I doubt it'll be smaller. What's yours?
If you want help finding the parts of your pipeline your reporting can't see, that's the work in a T2D3 Growth Workshop. Bring your last ten deals, and we'll go through where they actually came from.
What I'd actually do
Three things, and the order matters, because the first one is free and the third one is where most people want to start.
Ask the question out loud, in the first few minutes. Not on a form, where the answer is a dropdown of options you thought of last year. In the conversation, where the answer can be something you weren't expecting. "How did you come across us?" is a fair question to ask a founder who already booked time with you, and the answers worth having are the ones that don't fit any of your categories.
Then answer somewhere to live. One field, filled in by hand, is enough to turn four anecdotes into a count. It doesn't need to be elegant. What kills a finding like this isn't crude tracking; it's no tracking, because then every instance evaporates when the call ends and the pattern never assembles.
Only then work on the optimization. Answer engine work is real, and it's worth doing. It's the third step and not the first, because until you can count the thing, you have no way to know whether the work moved it. Doing it in the other order gets you a project with no scoreboard, and that's how most marketing budget quietly disappears.
A four-count off one seller's calendar isn't a competitive advantage that survives being held quietly. It's an early reading, and early readings get more useful when other people check them against their own rooms.
So if you're running a B2B SaaS or AI company and you're having these conversations too, count yours. Ask your next ten prospects how they found you, and write the answers down somewhere they'll still be in a month.
If a model referred any of them, you now know something about your own demand that your reporting doesn't. Nobody is going to hand you that number. The founders who have it in six months will be the ones who started writing it down before there was a dashboard for it.
Four out of 44, in fifteen working days. That number will be different by the time you read this, and I doubt it'll be smaller. What's yours?
If you want help finding the parts of your pipeline your reporting can't see, that's the work in a T2D3 Growth Workshop. Bring your last ten deals, and we'll go through where they actually came from.
How do I track leads that come from ChatGPT or Claude?
You mostly can't, not through analytics. There's no referrer on a conversation with a model, so the traffic lands as direct or as a form fill. The practical answer is to ask on the call and record the answer in one field by hand. Crude tracking that exists beats good tracking that doesn't.
Is answer engine optimization worth doing yet?
Yes, and not first. Until you can count how many buyers arrive that way, you have no baseline, so you can't tell whether the work moved anything. Spend a month asking the question and recording answers, then start the optimization with a number to measure against.
Four out of 44 is a small sample. Why act on it?
Because the cost of acting is one question and one field, and the cost of being late is a quarter of pipeline you never knew existed. Small samples are a reason to measure properly rather than a reason to wait. If you run the same count and get zero, that's a real finding about your market too.
"Answer engine work is real" →https://www.kalungi.com/blog/b2b-saas-seo-in-the-age-of-ai and "T2D3 Growth Workshop" → https://www.kalungi.com/workshop