How to Create a ChatGPT Agent for Marketing

Igor Ivitskiy

Igor Ivitskiy

  • #6 in The Top 50 Most Influential PPC Experts
About the author ↓

Published 12 min read

In this article
  1. A project instead of a new custom GPT
  2. Which files and instructions the agent needs
  3. Which task you give it
  4. What the agent must not decide
  5. How not to confuse this with Google Ads setup
  6. When this does not apply
  7. Questions and answers
  8. A new GPT on Plus, Pro, Go, and Free cannot be created
  9. The old agent mode is already closed in the help page
  10. An old GPT chat does not move into the project
  11. You publish the draft, the agent does not
In short: A new custom GPT on Free, Go, Plus, and Pro cannot be created. The marketing agent lives in a project that holds your files. You keep price, promise, client facts, and publishing.

I build a marketing agent in a ChatGPT project. It is a chat with files and instructions. On Free, Go, Plus, and Pro you cannot create a new GPT. That line is in the help page on creating GPTs, which I opened on 16 November 2026. GPTs that already exist stay until 11 December 2026. FedRAMP workspaces, and Enterprise workspaces with a confirmed deferral, have a retirement date of 11 February 2027. The ChatGPT agent page says the old agent is no longer available and points longer work to ChatGPT Work. You still decide the price, the promise, any fact about a client, and whether anything is published.

An empty chat remembers the thread only while it is open. The next job starts without your prices and bans unless those live in files. That base is what an AI marketing system: what it is and where to start is built on. A wider look at building an agent without code sits in AI agents for marketing: how to build without code. This page is the ChatGPT project around your own files.

I have spent 19 years in Google ads, since 2006. People still search for how to create an agent and land on a guide whose first click is Create. On a personal plan that click is gone. I made the same point from another side at The AI Era in Marketing and Business: Igor Ivitskiy at London Business Club 2025: a tool with none of your facts answers with confidence and misses the job.

A project instead of a new custom GPT

I do not build a new custom GPT on a personal plan. The help page names the plans: Free, Go, Plus, and Pro. Creating and publishing a new GPT is closed there. A GPT you already have can stay, and you can still edit it when the plan and the permissions allow it.

I checked the dates on 16 November 2026 in the custom GPT retirement FAQ. The standard retirement is 11 December 2026. The later date, 11 February 2027, is only for FedRAMP workspaces and affected Enterprise workspaces with a qualifying, approved deferral. A deferral does not keep the right to create new GPTs. For affected Enterprise workspaces, new creation ended on 26 October 2026. FedRAMP is outside that stop. If your account notice names another day, follow the notice.

A GPT you already built keeps working until its own date. When migration shows up on the account, the replacement is a plugin: instructions and files move, and the original becomes read-only after migration. Migration does not open for every account on the same day, and unpublished drafts do not move. For weekly emails, a review, and a plan, the same FAQ points to a project.

Here is the setup. The labels come from the projects page. If your screen uses another word, follow the screen.

  1. Sign in to ChatGPT. There is no project without a sign-in.
  2. Choose New project in the sidebar. Add a name, an icon, and a color. Put facts in files, not in the name.
  3. Open the three-dot menu and choose Project settings. Write the role, the language of the reply, and what to do when a line is missing from a file.
  4. Add files: a PDF, a spreadsheet, a document, an image, or pasted text. No more than ten at a time.
  5. Start a new chat inside the project. An old chat with a GPT cannot be dragged in.

Test it with a question whose answer sits in only one file. A smooth general reply means the file is missing or the file name says nothing.

For a long job, the agent page sends you to ChatGPT Work. Lower on that page the old clicks are still written out, including the tools menu and the /agent command. I do not give those clicks. The projects page still lists agent mode among tools a paid plan may include. If your menu still shows the old item, the first paragraph of the agent page is the one to follow. Work is on the web and in the mobile app on the paid plans where it has been opened. A Work chat started from the project uses the project files and instructions, and you approve important actions. If project memory is project-only, the projects page says Work is not available inside that project.

Which files and instructions the agent needs

Instructions tell every chat in this project how to behave. A file tells what is true about the company. Instructions apply only here and override your global custom instructions. Do not hide rules and tone inside a file. They belong in Project settings.

FileWhat goes inWhat to leave out
Who we areOne or two pages: what you sell, to whom, and what you refuseA slogan with no facts
Offer listServices and terms. Put the price date in the file nameA price list with no date
CustomerObjections from real emails, with no person’s nameA portrait you invented
VoiceThree to five emails or replies you actually sentText just written in the chat
BansWhat not to promise, and what stays a draftPasswords, logins, card data

The help page does not cap how many projects you can create. It does cap files. I checked the limits on 16 November 2026. Free allows 5 files in a project, Go and Plus allow 25, and Edu, Pro, Business, and Enterprise allow 40. You can upload no more than 10 at a time. The page says projects have no separate charge: access is included with the plan you already have. Remove the extra files or combine their data. Split the work into a second project only when the role is different.

For a plumbing company I would keep one project for replies to people who ask about a job and worry about the price. The draft takes the price from the offer file. Put in a competitor map only after you have checked it. The same kind of base, in another product and without this screen, is Claude for Marketing: A Project With Your Knowledge Base.

Project-only memory does not pull in your other chats. I use it when prices and emails are in the project, and then Work is not available inside it, so the task stays an ordinary chat. I take a long draft to a separate Work chat, without those files. On Plus and Pro, default memory looks at the project chats and files first. Outside Enterprise and Edu, a project chat can still reference conversations outside the project. On Enterprise and Edu, chats stay inside the project even on default memory and do not reference conversations outside the project.

You can invite a colleague. Chat access lets them read and talk. Edit access lets them change files and invite others, and only the owner can remove someone. A shared project is forced onto project-only memory, so Work is not available inside it. Until the files are checked, I keep the project private and I do not put passwords in it.

Which task you give it

One project slot is one job you repeat: a reply draft, a review of a text against your file, or a short list of points. It is not all of marketing in one line.

Write the task in steps. For the plumber it sounds like this. Use the offer file. Write a reply to someone who asks about a clogged drain and is worried about the price. Take the price only from the file. If the price is missing, say so. Do not promise when the job will be done. Do not send the email. Return the draft and the list of lines that are missing.

A separate Work chat takes the long draft, when the screen shows Work. A short question stays in an ordinary chat inside the project. Name one file, the shape of the reply, and the stop: if the line is missing, say what is missing and do not fill the hole.

What the agent must not decide

The agent in this project prepares text. It does not run the business. I keep four decisions even when the draft sounds smooth. In Work you approve important actions, and that approval does not catch every quiet mistake in a fact.

DecisionWho makes itWhat the project does
PriceYouTakes a number only from a file whose name includes a date
PromiseYouDoes not promise a result or a deadline
A fact about a clientYouDoes not add anything the file does not contain
PublishingYouReturns a draft and stops

A price whose file name has no date is worse than a blank cell: the reply will sound calm and be old. A fact about a client is a claim about a named person or their job, not a sketch of the audience. Publishing is the button on the site, in the email, and in the ad account. You press it.

Once the files, the instructions, and this line are in place, the agent can sit on your own base rather than on someone else’s GPT. If you also run Google Ads, the step-by-step version for business owners is in my book.

Want the system, not scattered tips?

My e-book “19 Google Ads Secrets”: how to get the most out of Google Ads without draining your budget, drawn from real campaigns and tests.

See the book →

How not to confuse this with Google Ads setup

This project does not turn a campaign on and does not change a bid. A draft ad in the chat is not a change in the account. Whether a chat suggestion holds up when you open the account is already covered in ChatGPT for Google Ads Setup: What Holds Up in Your Account. I do not repeat that path here.

Search themes in Performance Max are a hint you give Google’s algorithm inside a campaign, not this agent with your files. How to give them is in Performance Max Search Themes: How to Guide Google’s AI. You can keep the bans file in the project so a draft ad does not break them. Turning the campaign on stays in the account.

When this does not apply

A project does not replace a button that is already gone, and it does not publish for you. I would not call the setup a finished agent in these cases.

  • You need a new custom GPT on Free, Go, Plus, or Pro. There is no create button. A guide that starts with Create describes an old screen.
  • The task is to change a price, promise a result, or state a fact about a named person. That stays with you, even after the draft is written.
  • Project memory is project-only and you expect ChatGPT Work inside it. The help page says Work is not available there. A shared project is on that memory by itself, and you cannot switch it.
  • You want the chat to change a live Google Ads account. A draft is not an edit in the account.

One more limit. An old chat with a GPT does not move into the project. If you do not see Move to project, start a new chat in the project and put the conclusion you need into a file.

Questions and answers

A new GPT on Plus, Pro, Go, and Free cannot be created

No. Creation is closed on these plans. In FedRAMP, creation is allowed. In Business, Enterprise, and Edu it depends on product availability, settings, and permissions, and in affected Enterprise workspaces it stopped on 26 October 2026.

The old agent mode is already closed in the help page

Yes, in the first paragraph of the ChatGPT agent page. Long tasks go to ChatGPT Work. Work starts from the project only when memory is not set to project-only.

An old GPT chat does not move into the project

No. It stays in your history. Open the new chat inside the project so it picks up the instructions and the files. Actions tied to an outside service do not come across by themselves. The replacement is a connected app or a custom MCP server. Do not assume it brings every capability of the old action.

You publish the draft, the agent does not

Yes. The project returns text and should stop when a fact is not in the file. You check the price, the promise, the fact about a client, and the publish button. An approval in Work is an extra stop, not a substitute for reading.