Ritesh Mathur · 8 min read · October 2, 2026
Stop asking ChatGPT whether your message is good. Run it against a defined audience of simulated buyers instead. Here is how to set up icp-gtm-sim in ChatGPT and Codex and test your GTM message in minutes. "Is this cold email good?" Paste that into ChatGPT and you will get polite, reasonable feedback. Tighten the subject line. Lead with the pain. Add a clear call to action. That is copy editing. It is useful, but it does not tell you the thing you actually need to know: who will this land with, and why will everyone else ignore it? A GTM message simulation answers that question. It puts your exact message in front of hundreds of simulated buyers, each defined by attributes you could target, and shows you where the response changes. This guide shows how to run one in ChatGPT using icp-gtm-sim, our free and open-source skill.
Why a simulation beats a single prompt
Asking ChatGPT to "pretend to be 50 customers" gives you 50 answers. It does not give you an experiment. A message simulation has structure: A defined audience. Every valid combination of role, company size, industry, revenue and geography, each included once. A controlled stimulus. Every persona reads the exact same message, in the format it will actually ship in. A fixed response scale. Relevance, clarity and trust on a 1.00–5.00 scale, plus an intent: ignore, reject, inspect details, start a trial or bring it to the team. Checks before conclusions. Flat scores and stereotyped responses are flagged before any segment claim is made.
The output is not "this email is a 7/10". It is "this email lands with founders at 11–50 person companies and loses RevOps leaders on trust, not relevance."
Setting up the skill in ChatGPT
The skill is a single SKILL.md file, written in the open Agent Skills format that ChatGPT and Codex both support. There are three ways to use it.
Option 1: ChatGPT Skills
Skills are available on eligible ChatGPT Business, Enterprise, Healthcare and Edu workspaces, depending on your workspace settings. Step 1: Download SKILL.md from the icp-gtm-sim repository. Put it in a folder named icp-gtm-sim and zip the folder. Step 2: In the ChatGPT sidebar, select Plugins, then open the Skills tab. Step 3: Select Create, then Upload from your computer, and choose the zip. Step 4: ChatGPT scans uploaded skills before they become available. Once it is ready, start a new chat and ask: "Use icp-gtm-sim to test this message", followed by your copy. If you do not see Skills, your workspace admin may need to enable skills and skill uploading.
Option 2: A ChatGPT Project
If your plan does not have Skills, a Project works well. Step 1: Create a new Project, for example "GTM message tests". Step 2: Add SKILL.md to the project's files. Step 3: Set the project instructions to: "For every request in this project, follow the instructions in SKILL.md." Step 4: Start chats inside the project and paste the message you want to test.
Option 3: Codex
If you use Codex in the terminal or your IDE, install the skill with skills.sh: npx skills add ResearchifyLabs/icp-gtm-sim Then ask in plain language, or call the skill by name with $icp-gtm-sim . Codex can run code, which the skill uses to build the persona grid and compute segment averages instead of estimating them by eye.
Running a message test
Step 1: Paste the message exactly as it will ship
Not a summary, not "an email about our AI tool". The actual subject line and body, the actual landing page hero, the actual LinkedIn ad.
Tell the skill three things, or let it ask: The format. Cold email, landing page, deck slide, in-product message. A cold email gets four seconds of attention. A landing page gets scrolled. The price or commercial ask. Without it, personas cannot tell you whether they would start a trial themselves or need approval. Your hypothesis. "I think this lands with VP Sales at 51–200 person companies." The report will give you a verdict on it, even when the verdict is unflattering. If your message names its own audience ("Built for RevOps teams"), the skill will point it out. Every RevOps persona will score it higher simply because it names them. Decide whether you want to test the message as-is or discover who it resonates with.
Step 2: Confirm the audience
The skill proposes 4–5 factors inferred from your message, drops impossible combinations and tells you the persona count. You adjust it conversationally: Drop manufacturing. Add funding stage: bootstrapped, seed, Series A, Series B+. Only US and India. Keep the first grid modest in a chat, a few hundred personas, and widen it once the setup looks right.
Step 3: Read the report
The report comes back in a fixed structure: The answer: who the message lands with and how strongly. Who it lands with: best and worst segments, with the number of personas behind every figure. What is holding it back: ranked, and split into wrong audience versus wrong message. Your hypothesis: supported, not supported or not answerable here. What I would do next: one real-world check and one message variant worth testing. Run stats: personas, time taken and tokens used. You also get a CSV of every persona's reaction as an audit trail.
Diagnosing what is wrong with a message
The most useful part of a message simulation is not the ranking. It is the diagnosis. Low relevance everywhere. The message is aimed at a problem this audience does not have. No amount of copy editing fixes that. Change the audience or change the problem you lead with. Low clarity everywhere. People do not understand what you do. Simplify. Say what the product is before you say why it matters. Low trust everywhere. The claim is not credible yet. Add proof: a number, a customer type, a mechanism, a guarantee. High interest, but "bring to team" instead of "start a trial". The segment likes it but cannot buy alone. Either your pricing motion needs a sales-assisted path for this segment, or you target the person who can say yes.
Strong in one segment, weak in the others. The message is fine. The targeting decision is the real finding.
Comparing two messages
Most message decisions are A versus B. Run both.
- Run message A against your audience.
- Run message B against the same audience grid.
- Ask the skill to compare the two runs. The skill will first confirm that both runs used the same format and price, because a different stimulus moves results more than any segment does. Then it maps the segments that overlap, puts the number of personas next to every figure, and shows you where the two messages agree and where they diverge. Disagreements are the interesting part. They become your shortlist of what to test with real people. A good prompt for the comparison: Compare the last two runs. Same audience, same price, two subject lines. Which segments flip between A and B, and which variant should I send to each?
From simulation to the real world
A message simulation does not replace a real campaign. It decides which campaign is worth running. Take the strongest message for each segment into the market with a small, clean test, for example 50 accounts per cohort. You go into that test with a sharper hypothesis and fewer variants to burn through. Read more about this approach in Run ICP and GTM experiments like science.
When to use the hosted version
The skill is free and open source. If you test messages regularly, Tesemble runs the same method with more accurate, cheaper and faster results, and adds a visual report, AI-generated message revisions and one-click re-runs against the same validated audience. Every new account starts with 10,000 free credits. Using Claude instead? See how to run an ICP simulation in Claude.
FAQ
Can ChatGPT test marketing messages? Yes, with structure. A single prompt gives you opinions.
The icp-gtm-sim skill turns ChatGPT into a message simulator: it builds a defined audience of synthetic buyers, has each one react to your exact copy on a fixed scale and reports which segments the message lands with.
Does ChatGPT support skills? Yes. ChatGPT Skills are available on eligible Business, Enterprise,
Healthcare and Edu workspaces under Plugins → Skills. On other plans, you can add the skill file to a ChatGPT Project and instruct the project to follow it.
Can I use icp-gtm-sim in Codex? Yes. Install it with npx skills add ResearchifyLabs/icp-gtm-sim
and ask in plain language, or call it by name.
What kinds of messages can I test? Cold emails, landing page copy, positioning statements, ads,
sales pitches, launch posts and in-product messages. Tell the skill the format so personas judge it the way real buyers would.
How do I compare two messages? Run each message against the same audience grid, then ask the
skill to compare the runs. It checks that the stimulus format and price match, compares overlapping segments and highlights where the messages disagree.
Are simulated buyer reactions reliable? They are directional, not predictive of any individual. Use
them to narrow your options and form sharper hypotheses, then validate with real customers. Read more in Synthetic audiences: when they work and when they don't.