Jorge LewisTo build an ICP from closed won deals, let your AI read the won accounts through the CRM connector you already have, ask it for the traits that most of them share, turn those traits into one audience sentence and up to seven criteria, and then test the result on real people. With GTM Bud in the same Claude or ChatGPT workspace, the last two steps happen in the same chat: GTM Bud checks each criterion and a prospect preview shows who the criteria find.
This guide is for a RevOps analyst whose sales team targets by opinion. The VP of Sales says “mid-market fintech”. The best rep says “anyone with a compliance team”. You have 18 months of closed-won records and one afternoon before the planning meeting on Thursday, where you must hand over an audience that someone can launch.
What each tool does, so nobody expects the wrong thing
Three parts work in one chat, and each has one job.
| Part | What it does | What it does not do |
|---|---|---|
| Your CRM connector | Lets the AI read deals, accounts and contacts | It knows nothing about LinkedIn |
| Your AI | Finds the patterns, writes the audience sentence and criteria | It cannot see who the criteria will find |
| The GTM Bud connector | Checks each criterion, saves a draft, runs a prospect preview | It does not read your CRM |
The CRM connector is one that you or your IT team already added to the workspace. GTM Bud has no CRM integration through its connector, and the connector has no file upload. The data path is this: the CRM data goes to your AI, and your AI sends GTM Bud a campaign name, an audience description and the criteria. The draft tool takes nothing else. Later, the reason you give for a fit mark is saved too, so keep customer names out of it.
Step 1: Have the AI read the won deals
Tomasz is the RevOps analyst at a 60-person company that sells expense audit software. He starts with the read, and he asks for evidence, not for a conclusion.
“Use the CRM connector. Pull every closed-won deal from the last 18 months with the account industry, employee count, country and the job title of the primary contact. Show me the table first. Do not interpret it yet.”
He reads the table before any analysis. Two things usually show up at this point: blank fields and job titles typed five different ways. He asks the AI to group the titles into roles and to count the blanks for each column. A pattern built on a half-empty column is not a pattern.
Then he asks for a comparison, because won deals alone flatter every theory.
“Now pull the closed-lost deals of the same period with the same fields. For each trait, show me the share among won and the share among lost.”
A trait that is common among won deals and equally common among lost deals describes who the team called. It does not describe who buys.
Step 2: Keep only what a profile can show
This is the step that most ICP exercises skip. A CRM holds private data. LinkedIn targeting can only use public data.
GTM Bud reads targeting in two steps. First it searches LinkedIn for people by role, industry, company headcount and location. Then it reads the LinkedIn profile and the company of each person and checks each criterion. A criterion must be one requirement that a profile, a company page or public web content can show.
So Tomasz sorts the traits from Step 1 into two lists.
- Usable. Role of the buyer, seniority, industry, headcount range, country. Also things that a company page or a public site shows, such as “has a published careers page with finance roles”.
- Not usable as written. Deal size, sales cycle length, lead source, the expense tool the customer used before. GTM Bud cannot see private data, for example the tools that a company uses.
An estimate is acceptable when the criterion names its evidence. “Revenue in the mid eight figures” is private. “Mid-sized by revenue, estimated from headcount” names what the judgment rests on.
Step 3: Write the audience sentence and the criteria
The AI writes these, from the usable list. Tomasz asks for the sentence first.
“Write one audience sentence from the usable traits. Then write the criteria, at most seven, one requirement in each. Mark each one must-have or nice-to-have and tell me which won-deal share supports it.”
What comes back in his case: “Finance controllers and heads of finance operations at professional services firms with 200 to 1,500 employees in the United Kingdom and Ireland.” Under it, five criteria, each with the share of won deals behind it.
The must-have mark deserves care. In GTM Bud, a person that fails one must-have criterion is not a match. A nice-to-have criterion is a preference and does not reject a person. Tomasz keeps role, industry and headcount as must-have. He moves “has a shared services center” to nice-to-have, because only a part of the won accounts had one.
Step 4: Let GTM Bud check the criteria
“Create a campaign draft named ‘UK professional services, finance’ with that sentence and those criteria. Show me the result of the criteria check and stop.”
A draft sends no messages. GTM Bud checks each criterion and keeps the result. The check flags one kind of problem: a requirement that cannot be read from a LinkedIn profile, a company page or public web content. For such a flag it returns a message and, in most cases, a suggested rewrite. It also flags a criterion that repeats what GTM Bud sourcing already does for every campaign, such as leaving out people the account already contacted. That flag has no rewrite. Its message says to remove the criterion.
In this example the check flagged one criterion: “processes more than 5,000 expense claims a year”. That is a number from the CRM notes, and no public page shows it. The suggested rewrite pointed to headcount. Tomasz removed the criterion, since headcount was already a must-have, and the AI updated the targeting of the same draft.
A failed must-have can reject most people, so fix or remove each flagged criterion before the preview.
Step 5: Test with a prospect preview
The preview is where opinion meets evidence from outside your own CRM.
“Start a prospect preview for the draft and show me the people when the search is done.”
The search takes one to three minutes. Each person comes back as a short card: name, role, company, location and a short reason for the match. For any card, the AI can open the full profile, which shows experience, company details, recent posts and the result for each criterion.
Tomasz reads each card and asks one question for each: would the best rep be glad to see this name? He marks the fits.
“Mark Aoife R. as a good fit. Mark Daniel K. as not a fit: he is a finance controller, but the firm is a recruitment agency, and we never won one. Exclude everyone like him.”
GTM Bud keeps each mark with the campaign. The reason is in his words, and the choice between “never people like this” and “sometimes acceptable” is saved with a not-a-fit mark.
A preview is one short list per run. It is a sample of who the criteria find, not a lead list. A new preview with an unchanged audience shows the same run, and each campaign has a daily preview limit. So a new search makes sense after the criteria change.
What the won deals cannot tell you
Three honest limits belong in the handoff.
- Won deals show who bought from this team. They carry the bias of where the reps prospected. The lost-deal comparison in Step 1 reduces this. It does not remove it.
- Existing customers are not excluded by the connector. GTM Bud keeps a company exclusion list on the Account page of gtmbud.com. It applies to all campaigns of the account, and sourcing removes people at those companies. No connector tool reads or edits that list. Add the customer accounts there before launch.
- The preview shows fit, not intent. It tells you the criteria find the right kind of person. It says nothing about who will answer.
The handoff for Thursday
Tomasz hands over four things on one page: the audience sentence, the criteria with must-have marks and the won-deal share behind each, the names of the previewed people that sales agreed are a fit, and the list of traits he dropped because no public page can show them. The draft exists in GTM Bud with its targeting saved. Messages, value offers and launch are the next person’s work.
The test of the page: a rep who reads it can name a real company that fits and a real company that does not. If both are easy, the ICP is specific enough.
For the thinking behind an ICP before any data, see how to build an ICP for outbound, and for the difference between a market, an ICP and a list, read TAM vs ICP vs target account list.
Frequently asked questions about building an ICP from won deals
How many closed-won deals do I need to build an ICP?
There is no fixed number. With fewer than about 20 deals, treat every pattern as a hypothesis and say so in the handoff. The prospect preview is the test in both cases, because it shows real people and not a summary of your own records.
Does GTM Bud connect to my CRM?
No. GTM Bud has no CRM integration through the connector. Your AI reads the CRM through a CRM connector that you already have in the same chat, and then it sends GTM Bud only an audience description and a list of criteria.
Which CRM fields can become LinkedIn targeting criteria?
Fields that a LinkedIn profile, a company page or public web content can show: job title, seniority, industry, headcount and location. Fields such as deal size, lead source or the software a customer uses are private data, and the criteria check flags them.
Can the campaign skip companies that are already customers?
Not through the connector. GTM Bud has a company exclusion list in your account on gtmbud.com, and sourcing removes people at the companies on that list. No connector tool reads or edits the list, so add your customers there yourself.
What happens after the prospect preview?
You mark people as a good fit or not a fit, with a reason, and GTM Bud keeps the marks with the campaign. If the preview shows the wrong people, your AI rewrites the criteria of the draft and you open a new preview for the changed audience.
From opinion to a tested audience in one afternoon
You can build an ICP from closed won deals in one afternoon when the AI does the reading, the criteria are limited to what a profile can show, and a preview of real people is the test. Add GTM Bud next to your CRM connector with the steps on the GTM Bud connector page. For the gates around what happens after the draft, see AI LinkedIn outreach for RevOps on GTM Bud.