By the end of this hour you will have a ChatGPT account with the data settings chosen deliberately rather than left at their defaults, one project holding your recurring work, and one competitor scan finished.
The first 10 minutes are the settings. Do those before any operator material goes into a chat, because the setting you want is off by default in the sense that it is on by default and most people never open the menu.
This is the widest-reach pillar in the set. Marketing, CRM, commercial, product, compliance.
The setup
An account at chatgpt.com. Every plan has projects, with the file limit changing by tier, so free is enough to follow this.
Before you start, know which account you are on. A personal account and a Business or Enterprise workspace behave differently on data, on memory and on what you are allowed to build. That difference matters again in the third ChatGPT pillar in this series.
Step 1: turn off model training
Click your profile icon, then Settings, then Data Controls. The setting is called “Improve the model for everyone”. Turn it off.
On mobile the path is the side menu, then your profile icon, then Data Controls.
Do this before you paste a single line of anything internal. Campaign plans, supplier terms, unpublished results, draft policies, all of it.
Step 2: decide what memory does
Go to Settings, then Personalization, then Memory.
Memory picks up context from conversations, files and connected apps and carries it forward so you repeat yourself less. That is useful and it is also the source of the failure in this pillar, covered further down.
Two things to know now. Saved memories are stored separately from your chat history, so deleting a conversation does not delete what ChatGPT learned from it. And the memory summary shown in Settings will not include everything ChatGPT remembers, because some details are left out of the summary.
Three practical choices:
Leave memory on if you work on one brand and want continuity. Turn it off from the three-dot menu, using “Delete and turn off memory”, if you switch between brands or clients. Use a Temporary Chat for anything sensitive, since Temporary Chats do not use existing memories, do not create new ones, and are deleted from OpenAI’s systems after 30 days.
Step 3: create a project
Click New project in the sidebar. Name it, pick a colour.
Open the three-dot menu and add project instructions. These apply only inside that project and they override your global custom instructions there.
You support a marketing team at a licensed online gambling operator.
Rules:
- We are licensed in the UK (UKGC), Malta (MGA) and Romania (ONJN) only.
Refuse any request for copy targeting a market not on that list.
- Never use: risk-free, guaranteed, no-lose, beat the house, easy money.
- Never imply gambling relieves stress, boredom or loneliness, or solves
financial problems.
- Never write anything that could appeal to under-18s.
- Never attach a revenue, conversion or retention promise to a claim.
- GGR and NGR are different numbers. Use whichever the source uses and
never substitute one for the other.
- State your source for every figure. If you do not have one, say so.Then set the project memory mode. Default memory lets the project reference your saved memories and other chats. Project-only memory keeps everything inside that project. For anything brand-specific, project-only is the setting that prevents the leak described below. Shared projects use project-only memory automatically and cannot be changed.
Step 4: add files, within your tier’s limit
Upload the operator context file if you built one for the Claude pillar. The same six blocks work here.
File limits per project are 5 on Free, 25 on Go and Plus, and 40 on Pro, Business, Edu and Enterprise. Projects themselves are unlimited.
Step 5: run a competitor scan
Now do something real. Pick one market and three competitors.
For [market], compare these three operators on what is publicly visible
on their sites today: welcome offer structure, wagering requirement,
minimum deposit, payment methods listed, RG tools linked in the footer,
and where the licence number is displayed.
One table. One row per operator. Cite the page you took each item from.
If something is not visible, write "not visible" rather than inferring it.Then check three cells against the actual sites. Public financials work the same way if you point it at published results, and something like Bragg’s Q2 results and withdrawn guidance gives you a scan you can verify line by line against the release.
Where it breaks
Memory carries one brand into another
This is the documented behaviour that costs marketing teams the most. Memory pulls context from conversations and files across your account. Work on Brand A in the morning and Brand B in the afternoon, and the model has both. It will use a Brand A bonus mechanic in Brand B copy without flagging it, because as far as it knows both are yours.
Two things make it worse. The memory summary in Settings does not show everything ChatGPT remembers, so you cannot audit it fully by reading that page. And saved memories survive the deletion of the chat they came from, so clearing your history does not clear them.
The fix is project-only memory for brand work, or memory off entirely if you handle several clients.
“Not visible” turns into a guess
Ask for a competitor table and you will get a complete-looking table. Cells the model could not verify tend to arrive filled in with what a typical operator does. The instruction in step 5 reduces it. Checking three cells against the live sites catches the rest.
The settings are per account, not per team
Turning off training on your account does nothing for the colleague who pasted the same document into their own. On a personal plan there is no central control over this. If your team handles operator material daily, that is the argument for a Business workspace rather than 6 personal accounts.
The limits
No player-level data goes into any of this. Not in a chat, not in a project file, not in a Temporary Chat.
A competitor scan is a starting point for a commercial conversation. It is not market intelligence you present as fact, and it is not a compliance assessment of anybody’s site.
Under the EU AI Act you are a deployer rather than a provider when you use ChatGPT this way, and the duties differ. Article 50 transparency obligations have applied since 2 August 2026, with a limited grace period to 2 December 2026 for marking obligations on systems already on the market. The deployer duty covering AI-generated text applies to text published to inform the public on matters of public interest, and text that has had genuine human review and editorial control is exempt. A superficial check does not count as review. Where your published marketing sits against that line is a question for your legal team.
AI for iGaming is a recurring series from The iGaming Europe. Bartosz, our Head of Content, shares the prompts, setups and workflows he has tested on real iGaming work. One workflow per issue, including what went wrong. No tools to buy, no affiliate links.
Source: OpenAI Help Center, European Commission
