how a product link turns into customers: the gtm engine i'm building

most founders i talk to have a product and no plan for getting it in front of people.

not no effort. they post, they reply, they try a launch. but it's a pile of tactics, not a system. nothing tells them who to talk to today, what to say, or whether last week's thing worked.

so i've spent the last few weeks mapping the whole thing out as one pipeline, page by page, and building it into vibedraft. this post is the map.

the short version: you paste a product link. the agent works out who buys it, finds where demand is showing up, drafts the next move, and learns from what paid. you approve every message and every page.

push and pull are one pipeline, not two

customers reach you in two directions.

• push: you go to them. someone posts the problem in their own words right now, and you show up with something useful. fast. it fades the day you stop.
• pull: they find you. people search for it, or ask an AI assistant, and a page you wrote is the answer. slow. it compounds.

most advice treats these as different jobs with different tools. i don't think they are. it's one pipeline that works on two kinds of thing: a person, or a page or search. a sign-up from a page becomes a person, and from then on it's the same flow.

the first stage decides which direction fits your product, from evidence: do buyers talk about this problem in public? do they search for it? if the answer is neither, that's worth knowing before you spend a month posting.

the 12 stages, in 4 phases

understand: link → brief → ICP
listen:     moment → detector → signal → opportunity
act:        play → campaign → next move
learn:      funnel → loop

each phase hands exactly one thing to the next.

1. understand the product

link. one url. that's the whole intake.

brief. the agent reads the site and the market and writes down what you sell, who buys it, and the problem in the buyers' own words. every line carries an evidence level, from "a buyer said this in public" down to "the site claims it." then the founder corrects it. a brief built only from your landing page is a brief built from your guesses.

ICP. the brief becomes a filter: lines for who fits, and lines for who never fits. the not-if lines matter as much as the fit lines. mine include a rule i set on oct 9: a lead's product only counts as hot if buyers describe the problem in public, one customer is worth roughly $50 or more, and it's founder-led. a $5.99 one-time gift fails that, however well it matches everything else.

hands off: who buys, and which directions to run.

2. listen for who's ready

this is the part i find most interesting, and the part most tools skip.

moments. fit tells you who could buy. a moment tells you when. it's a change in someone's situation that opens a buying window: they just launched, they just complained about a competitor's pricing, they asked "what do you use for X." i mapped 24 of them, split into cold (strangers showing the moment in public) and warm (people already in your funnel), and the agent picks the three that matter most for your buyer as the place to start.

detectors. a moment is an idea. a detector is the rule that watches for it: where to look, what words to match, how to judge a match. a new detector runs in shadow beside the live one on the same posts before it's switched on, so you can see what it would have caught.

signals. one real person, or one page or search, caught in one moment. a judge checks it's really the moment and not a keyword collision, then routes it to one of four queues: pain, engage, listen, launch.

opportunities. signals get merged into a person, enriched, scored and tiered: hot, warm or watch. watch means help them in public, never chase them for a sale.

your own reading feeds this too. newsletters, rss, youtube, hacker news, your x bookmarks. a save is a strong signal of what you find valuable, and any source can feed a detector.

hands off: who's ready now, and what people are searching for.

3. act, one play at a time

plays. when this moment fires, do this: the channel, the offer, the steps, the stop rule. there are 26 in the registry today, things like listen-and-reply, comment-a-keyword and signal-based outbound.

campaign. one run of one play, on specific people, in a window, with a stop rule set before it starts.

next move. the agent ranks one move per lead and drafts it in your voice. you send. a contact policy sits over all of it: caps, quiet hours, and a never-contact list that wins every time.

that last part is non-negotiable for me. vibedraft drafts. it doesn't auto-follow, auto-like or auto-DM strangers. the agent does the finding and the writing. the founder does the sending.

hands off: the drafts you approved, sent, each with an id.

4. learn what worked

funnel. reached → visited → signed up → activated → paid. the weakest rung is the constraint. fix that one first, and leave the rest alone.

loop. when a number crosses a line, something specific happens: a runbook fires, a hypothesis gets written, and a test opens on one knob.

hands off: what paid, fed back into which moments to watch and which plays to run. then it loops.

the experimentation engine under all of it

every stage has a knob. hero copy. which ICP lines are in. the judge's prompt. which play runs. the opener of a draft. its send time.

the engine picks which knob to test, assigns at random with the odds logged, reads the result against a win rule set before launch, and writes the learning down, losses included.

it doesn't test everything the same way, because it can't:

• offline replay for the brief, the ICP and the judge. re-run the new version over stored posts and labels. no live risk.
• shadow for detectors. new and old run side by side on the same items.
• randomized A/B for drafts, openers, send times and pages.
• holdout lift for campaigns. a slice of eligible leads gets nothing, so the effect is measured instead of assumed.
• off-policy estimates for the play recommender. judge a new policy on old decisions before it touches a real lead.

three rules keep it honest. one live test per knob. size it before you run it (a small account can only detect big effects, so test big changes). readouts happen on schedule, win or lose.

the reason this works at all is that every stage records its decisions in a shape a model can learn from. i wrote that half up separately: every decision in a go-to-market pipeline is training data.

what's built, and what isn't

checked against the code on oct 17.

built:

• the brief, with an evidence level and a status on every fact
• the ICP rules, with stated and evidenced ICP kept apart
• detectors as per-account lanes that run in shadow before going live
• signals routed into the four queues
• the 26-play registry, campaigns with stop rules, and a drafts table that logs every keep, edit and discard
• post metrics at 10 minutes, 30 minutes, 1 hour, 1 day and 7 days
• the funnel ladder on the engine page

partly:

• opportunities (sign-ups as leads is next)
• the next-best-action engine (phase 1 shipped)
• tracking (the events table exists; touch ids and logged odds don't yet)

not yet:

• a holdout on any campaign, which means every result so far is a correlation, not a measured effect
• one experiment registry inside vibedraft. my tests still live in a separate registry i run by hand
• sources beyond reddit and x. and reddit rss ends nov 13, so that one has a deadline

the goal the code is pointed at right now is small and specific: 50 waitlist joins by dec 31.

why build it as one map

because the tactics aren't the hard part. anyone can write a reply.

the hard part is knowing which reply, to whom, today, and whether the last ten did anything. that only works if understand, listen, act and learn are one loop that passes data forward, not four tools that each know a quarter of the story.

i'll post the numbers as they come in, including the ones that don't move.

FAQ

what is a gtm engine?

a go-to-market engine is a repeatable system for turning a product into customers. here it's 12 stages in 4 phases: understand who buys, listen for who's ready, act with one play at a time, and learn from what paid. each phase hands one thing to the next, and the last one loops back to the second.

what's the difference between push and pull distribution?

push is going to people who are describing the problem right now, with a reply or a message. it's fast and it stops when you stop. pull is being found when people search or ask an AI assistant, through pages that answer them. it's slow and it compounds. the engine runs both in one pipeline, and a page sign-up becomes a person in the same flow.

what is a buying moment?

a change in someone's situation that opens a window where they're likely to buy: a launch, a complaint about a competitor, a direct "what do you use for this" question. fit says who could buy. the moment says when.

does the agent send messages on its own?

no. it finds the people and drafts the move in your voice. you review it and send it. there's no auto-following, auto-liking or auto-DMing strangers.