Five small experiments, each stopped before it could say anything. Counting leads by channel with a total beside each one turns the choice into arithmetic.
Try five, keep none. A small budget goes to each channel in turn, and each one is judged by a feeling at the end of the month. A forty-person firm ends up paying for four channels and reading none of them.
By hand today
“I’ve tried Facebook and Twitter advertising (on a small scale - about 5 campaigns at $100 a pop)”
cyberferret, on Hacker News
That is the stage as most companies run it. Five small experiments, spread across whichever channels were being discussed that quarter, each stopped before it could say anything. The same founder describes spending over two thousand dollars on search adverts with limited success.
Above a certain size this becomes a hire. Demand generation adverts ask the person to decide the right mix between channels based on what is producing demonstrations. That is the correct question. The problem is that most companies cannot answer it, because nothing joins a channel to a booked meeting.
There is a good worksheet for the choosing part. The Bullseye framework by Gabriel Weinberg and Justin Mares lists nineteen channels, sorts them into three rings and tells you to test the inner three. It is a thinking tool, and it runs on your judgement rather than on your data. Once chapter 2 has given every company a band and a reason, you can do better than judgement.
What it costs you each week
Two thousand dollars, spent by one founder on search adverts with limited success, before the channel was judged. Five more campaigns at a hundred dollars each went the same way on two social networks.
Source: cyberferret, Hacker News, item 11937914
That is one founder describing one year. The money is not really the point. The point is that at the end of it, nothing was known, because no click carried anything into a record that could be counted later.
The recurring cost is the third and fourth channel: the retainer, the tool seat, the hour a week of somebody’s attention, all still being paid for a channel nobody has looked at since it was started. Channels are easy to open and socially awkward to close.
The build
Open Claude Code and count leads by channel. It reads the CRM export, groups every lead by where it came from, and puts a total beside each count. A channel with clicks and no scored companies is named. So is a channel with no data at all.
Work in the folder holding the scored companies from the previous chapter and an export of every lead you have, with whatever source field your CRM already records. Then ask for the count.
Read leads.csv and companies-scored.csv. Group every lead by its source
channel. For each channel report: leads, leads that match a scored
company, how many of those sit in the top band, meetings booked, and the
total read. List separately the leads whose source is blank or reads as
direct, and say what share of the total they are. Do not guess a channel
for a blank. Write channels.md with the table and one line per channel
saying what the data can and cannot say about it.
What comes back is a table and an honest list of gaps. Read the blank-source pile first. If most leads have no source, the answer is not a channel decision but capture, two chapters ahead.
Two rulings make the table usable. The first is the audience size. When ArkOne built its LinkedIn targeting on 6 September 2026 it recorded the reachable pool for the same title list in two markets: 670,000 or more in the United States, 22,000 or more in the United Arab Emirates. A pool a twenty-fifth of the size cannot be scaled by raising the daily budget, so the ruling was recorded in plans/linkedin-ads-build-day/LEARNINGS.md on the day it was observed, before any money moved. The second is what a document advert is judged on: click-through to the landing page, never the network’s own blended rate.
Then write the choice down where later work reads it. ArkOne holds its version as one line in docs/crm-module.md: one buyer, one gate, two channels. Beside it, in docs/content-pillars.md, sits a registry of destinations, so every piece of content points at one of a fixed set of places, and a destination is changed in one file rather than in forty. You build your own. The paths are the pattern to copy, not a product to install.
What it feeds
The content calendar. Two channels means a calendar that writes itself: one place to publish, one place to reply. The next chapter fills it, which is a small and cheap job when the answer is two channels, and an impossible one when it is six.
The number that proves this stage worked is leads by channel with a total beside them, plus a list of the channels you closed and the date and reason for each. ArkOne named two channels and marked two other audiences, journalists and investors, as a channel and never pipeline, so nobody would later count them as prospects.
What stays with a person
The channel you stop paying for. The table can show you that a channel produced nothing. It cannot take the decision to close it.
Closing a channel is uncomfortable in a way that opening one is not. Someone recommended it, someone set it up, and a small number of leads did come through. The table does not settle this. It only removes the argument about the facts.
A person decides, and writes two things: the reason, and the date. Without those, the channel is proposed again next quarter by someone who was not in the room, and the money starts flowing again. With them, the answer takes ten seconds.
The tools you already pay for
Nothing new to buy.
CRM already stores a source on the lead. This stage only asks that you read it honestly, totals included.
Ad accounts will tell you your reachable audience before you spend anything. Write that number down at build time, because it is the earliest signal that a budget will not spend.
Notes hold the ruling. Two channels kept, the rest closed, each with a reason and a date.
Two channels, chosen from counts rather than from feeling, and the rest closed with a date. Then the calendar has somewhere to publish.
Ask: Which channel are you still paying for that nobody has read a number from this quarter?
This is chapter 3 of twelve, after the ICP a machine reads; chapter 4 fills the calendar. The book, Building a real GTM engine with Claude Code and Codex, end to end, runs the whole engine in order, one chapter per stage. A free starter edition for firms under fifty people is coming through the 100k programme.