What is a GTM brain?
A GTM brain is the shared state a company's go-to-market work runs on — customers, market, positioning, results — recorded so people and AI agents read the same thing before they act, with every claim carrying who decided it and on what evidence.
The category has not settled this term. Definitions differ between vendors, and this one is a position rather than a consensus. Where a common reading is being argued with, the entry says so.
A GTM brain is the shared state a company's go-to-market work runs on: what is true about its customers, market, positioning and results, recorded so that both people and AI agents read the same thing before they act.
What separates it from a knowledge base is that every claim carries its status. Who decided this, when, and on what evidence. A brain that cannot answer those three questions about a given sentence is a folder of documents with a more ambitious name.
The phrase also has an older colloquial sense, describing a person rather than a system: she has a GTM brain, meaning someone who thinks naturally about how a product reaches a market. This entry covers the system sense, which entered vendor language around late 2025 and is not yet standardised. Definitions currently differ between vendors, and this one differs from most in treating provenance as part of what makes a brain a brain.
What a GTM brain contains
Four kinds of state. The division is not filing convenience: each has different commit rights and a different decay rate, and mixing them is how a brain rots.
| Store | Holds | Written by | Decays |
|---|---|---|---|
| Decisions | Positioning, voice, ICP, offer, value propositions, what may be published | Humans only | Rarely, by intent |
| Market map | Competitors, segments, category vocabulary, keyword and citation landscape | Agents and humans, reviewed | On a clock — needs review dates |
| Entities | Companies, contacts, deals, customer stories, lifecycle stages | Agents, reviewed | Mirrors a system of record |
| Learnings | What converted, what a reviewer corrected, what an experiment settled | Agents, from measurement | Only accumulates |
A new brain is nearly empty in three of the four, and that is correct. Decisions can be filled in a day and market map in a week. Learnings cannot be filled at all until the system has run. A brain that arrives full of learnings on day one is fiction.
Why provenance is part of the definition
State without provenance is a cache. The distinguishing property of a GTM brain is that each assertion records where it came from, so a reader can tell a ruling from a guess.
A workable minimum is four statuses: decided by a named human, inferred by an agent and plausible but unratified, observed from data, and open — a known unknown left unresolved on purpose.
The rule that makes the markers worth having is that an inference never becomes a decision without a human. An agent may propose the promotion; only a person performs it. Where that rule is missing, every downstream draft treats a guess as settled, and nothing in the output shows which sentences were built on one.
Open questions are load-bearing rather than embarrassing. A brain with none, at a company that has not yet talked to its market, is describing a certainty nobody has.
Who may write what
Commit rights differ by store because the stores differ in kind. Judgment is committed by people. Observation can be committed by agents, under review.
The consequence is a working split rather than a policy statement: agents propose changes to the decisions layer and never merge them, while writing directly to the observational stores through the same review path. Enforcing that mechanically rather than by convention is what keeps a brain from quietly becoming a record of what an agent inferred last week.
How agents read it
A brain that must be loaded whole on every run stops being affordable at about the size where it starts being useful. Three load classes keep the cost bounded:
- Always. A small set read on every run: how the company sounds, what it claims, who it sells to, the boundaries agents may not relitigate.
- Named by the task. Full files read when the work at hand calls for them.
- On demand. Individual records, fetched by identifier: one competitor, one persona, one customer story.
The convention that makes this cheap is a short summary at the top of each file — a handful of lines holding only the load-bearing facts, so an agent can read the summaries of everything and the bodies of very little.
What a GTM brain is not
Four things get filed into brains and should not be:
- Process. How a comparison page gets written is a playbook. The brain holds what goes into the page.
- Live signals. A competitor raising a round is perception. That competitor's positioning, as of a date, verified against their own site, is state.
- Raw analytics. A traffic table is data. Definitional pages outconvert listicles for signups is a learning. A brain commits conclusions, not exports.
- Secrets and personal data. Every agent reads the brain by design, so anything placed in it is effectively disclosed to every run.
The test for admitting a file is whether some named reader loads it before producing work, and whether its absence would make the output generic or wrong. A file no one reads is a note.
How to tell whether one is working
Not by its size. The measure is whether the learnings store grows, because that is the only part that cannot be bought, copied, or written in advance. It is a record of one company's own outcomes.
A brain whose decisions are edited occasionally and whose learnings never grow is documentation. The loop is what makes it a brain.
This definition is written from a working implementation rather than assembled from other definitions. The brain it describes is a public repository, and the four-store split, the status markers and the load classes above are in daily use rather than proposed.
Frequently asked
- Is a GTM brain the same as a CRM?
- No. A CRM records what happened — accounts, deals, activity. A GTM brain records what is true and how confident anyone is in it, including judgment a CRM has no field for: positioning, voice, which claims may be published. Most GTM brains mirror a CRM for entity data and own everything else.
- How is a GTM brain different from a knowledge base or a style guide?
- Provenance and write-back. A knowledge base holds documents; a GTM brain holds claims that each carry a status marker and a decision path, and results from published work write conclusions back into it. A style guide that nothing updates is a document. A GTM brain that nothing updates is broken.
- What goes into a GTM brain?
- Four kinds of state, kept separate because they have different commit rights and different decay rates — decisions (positioning, voice, ICP, offer), market map (competitors, segments, vocabulary), entities (companies, contacts, deals, customer stories), and learnings (what converted, what a reviewer corrected, what an experiment settled).
- Who is allowed to change a GTM brain?
- It depends on the store, and the split is the point. Decisions are committed by humans only, because they encode judgment. Agents may propose changes to them but not merge. Market map, entities and learnings accept agent-written updates through review.
- Can a GTM brain be built from markdown files?
- Yes, and it is a reasonable starting substrate. Files in version control give provenance, diffs, review and history without new infrastructure, which is most of what a GTM brain is for. What they cost is query and scale, neither of which binds at one company.
- How do AI agents use a GTM brain?
- They read it before drafting, rather than working from a prompt alone. In a well-organized brain an agent loads a small always-read set on every run, the specific files its task names, and individual records like one competitor or one persona on demand — so that context cost stays bounded as the brain grows.