'What the Market Is Telling Us': How to Build a Customer Truth System

Five people, five versions of the customer

Put your sales lead, support lead, founder and head of marketing in one room and ask what customers want right now. You will get four confident answers, each true for the slice of customers that person talks to. Add the agency and you have five.

Nobody is lying. Each of them heard something real, but nobody heard all of it, and nobody wrote it down in a form the others could check.

A customer truth system fixes that. It is the first of the six systems in what a marketing engineer builds, and Greg Isenberg puts it first in his episode on the role, with one output, a single markdown file he names what-the-market-is-telling-us.md. His test for a good one is that it makes the business "harder to lie to".

One file. Updated every week. Every line with a receipt.

What the system is, and what it is not

The system has four parts. Inputs land in one folder, and an agent reads the new ones each week. It writes a memo where every claim carries a quote, a link and a count, and a person picks one test from it before Friday.

It is not a dashboard of satisfaction scores, and it is not a summary. A summary that says customers want easier onboarding is true of every company on earth and tells you nothing to do on Tuesday.

The good version is specific enough to be wrong. For a made-up scheduling product, a good line reads: calendar sync came up in 7 of 12 discovery calls, and in all four that closed. You can check that line, so you can act on it.

The memo lives in the customer-truth folder of your growth repo. That way every other agent, the one drafting ads and the one writing outbound, reads it before it writes a word.

The inputs, and how to get each one out

Most companies already record more than they read. The calls are transcribed, the tickets are stored, and the cancel reasons sit in a form nobody exports.

Start with calls. Zoom can save a transcript of every cloud recording as a separate VTT file once audio transcription is on, and Google Meet saves meeting transcripts to the organizer's Drive. If you use Gong, its transcript endpoint returns every call in a date range, split by speaker with timestamps.

Then tickets. Intercom's help center is clear about one trap: its CSV export holds reporting data, not the text of conversations, which needs the Conversations API or an S3 export. Zendesk's incremental export returns tickets changed since a time you give it, which is exactly what a weekly job wants.

In HubSpot, notes come out of the notes API in a field called hs_note_body, and logged call transcripts have their own transcript endpoint. Keep the deal stage on each file, because the most useful line in the memo compares what came up in won deals with what came up in lost ones.

Churn notes are the thinnest input and the most valuable per line. Add one required question to your cancel flow, keep every answer word for word, and save each one as a file. The habit of capturing material the day it happens is covered in Collect, Don't Invent.

Last, the words buyers type when nobody is listening. The queries in Search Console's performance report show what people searched before they found you, in their own words, with no salesperson in the room.

How many conversations are enough

You do not need thousands of conversations. You need a steady handful, read properly.

In 1993, Abbie Griffin and John Hauser published The Voice of the Customer, a study of how many customers a team must hear to find their needs. In their category, interviews with 20 customers surfaced more than 90 percent of the needs that 30 customers produced.

Hauser's later teaching note puts it at roughly ten one-hour interviews, or seven focus groups, for about 90 percent of needs in a study of complex office equipment, and says the result has been repeated in other categories. He also says some markets need more. Treat it as a floor, not a law.

Rhythm matters more than volume. Teresa Torres defines continuous discovery as, at minimum, weekly touchpoints with customers by the team building the product, and the weekly memo applies the same habit to marketing.

Our rule of thumb for a company of 10 to 500 people: ten to twenty new conversations a week across all sources is enough for a memo worth reading. Fewer than five, and the memo should say so in its first line.

The Quote-or-Cut Rule

One rule makes the memo worth reading. Every line carries a receipt, a direct quote with its file path plus a count where there is one, or the line gets cut.

The model makers recommend the same discipline. Anthropic's guide to reducing hallucinations says to have the model pull word-for-word quotes before it does its task, cite a quote for each claim, and retract any claim it cannot back with one.

The same guide says to give the model permission to say it does not know. That is why the memo has a section called What we do not know: a thin week reads as thin instead of padded.

If you build on the API, Anthropic's citations feature returns the passages from your own documents that support each answer, so the receipts come back with the text.

No quote, no claim.

Counts do the other half of the work. "Price came up a lot" becomes "price came up in 8 of 12 calls and 1 of 4 wins," and the loudest topic turns out not to be the one that sells.

The memo format

The memo is one page with seven parts, in the same order every week, so a reader can find last week's answer in the same spot. The template is in the figure, with example lines for a made-up product.

Two parts matter most. "Talked about vs bought" splits topics by won and lost deals, which is where a memo stops describing the market and starts pointing at money. "One test this week" turns the memo into a decision, with an owner and a date it gets read.

Keep last week's memo. Move it into a dated file in memos/ before the new one is written, so in three months you can watch a phrase rise in buyer words and then show up in your best headline.

Read it in the Monday meeting instead of a round of opinions. Five minutes of reading, then one question: do we run the test in part seven, or a better one? If nobody can name a better one with a receipt, you run it.

That last step is the payoff. Buyer words are the raw material for headlines, first lines and page copy, and the pages that use them sound like the reader wrote them.

The agent, the prompt and your first week

The agent's job spec is short. It runs Monday morning, reads only files added since the last memo, writes the new memo in the template, and stops for a person to approve it before any other agent reads it. The format for writing that spec is in How to Write a Job Spec for an AI Agent.

The prompt in the figure does the Quote-or-Cut work in order: extract quotes first, count, compare, write, then check every line against its quote and delete what does not hold. It never runs the test it proposes.

Run it in whatever you already use: a scheduled Claude Code task, an n8n workflow, or a person pressing a button on Monday. The schedule matters more than the tool.

Your first memo can be done by hand this week. Monday, export twenty recent conversations from two sources, say ten calls and ten tickets. Tuesday, strip personal details and give each file its five header lines.

Wednesday, run the prompt and read what comes back. Thursday, check every line against its quote yourself, and delete any line that stretches its receipt; the lines you delete show you what to add to the prompt. Friday, pick the one test and put it in the experiment log.

Do that twice by hand before you schedule anything. An agent that runs every Monday on a prompt nobody has checked will be wrong every Monday, on time.

Strip names, emails and phone numbers before raw files go in, and keep the raw folder out of version control. Europe's GDPR calls this data minimisation: keep only what the purpose needs, and the memo needs what was said and by what kind of buyer, not who said it. Check the call recording consent rules where your buyers are before you record anything.

"We already have a tool that does this"

You might. Your call recorder or help desk may already summarize its own data, and if it produces a weekly, quote-backed account of what changed with a test attached, keep it and skip this build.

The gap is structural. The call tool knows the calls, the help desk knows the tickets, the CRM knows the notes, and none of them puts a sales objection next to a support complaint about the same thing.

The memo is the one place where they meet. It also sits in a plain file that your ad, content and outbound agents read before they write, which no single vendor's summary can do, because each one only sees itself.

Our customer-language agent

We run a version of this on ourselves for the social input. A script scans 121 practitioner accounts on X for a topic, capped at $0.50 a run and typically about $0.02, and saves the raw posts for the writer before a word is drafted.

The writer reads what people in the market said that week, in their words. Anything we use from it is cited as that person's claim, never as a fact, which is the Quote-or-Cut Rule applied to ourselves.

If your calls, tickets and CRM notes live in three tools that do not share data, connecting them is CRM and revenue operations work. Building the memo on top, and the first test that comes out of it, is the kind of system our marketing engineering engagements build: one system, measured before the next.

What to read next, and why

The Growth Cockpit [system]. The Monday memo for the whole team, fed by this one. Customer truth says what buyers want; the cockpit says what your tests did about it.

The Founder Content Engine [system]. Buyer words are the best raw material a founder has for posts. This is how one recorded call becomes twenty pieces.

Argue Against the Story Your Market Believes [copy]. Once the memo shows you what buyers believe, this shows you how to argue with it on the page instead of agreeing with every competitor.

How We Run a Marketing Agency as Marketing Engineers [proof]. Our own systems and checks, with numbers. Read it if you want to see the receipts habit run across a whole business.

So the simple takeaway is this: a customer insight without a quote is an opinion, and a memo full of opinions is the meeting you were trying to replace.