Most marketers I talk to have the same complaint about AI tools: “It writes like a robot and I spend longer fixing it than writing it myself.” So they use ChatGPT or Claude now and then for a quick caption, and conclude that AI is overrated for real marketing work.
The problem is usually not the model. It is that every chat starts from zero. The assistant doesn’t know your brand, your customers, your prices, your tone or what you published last month. You wouldn’t expect a new freelancer to write well without a briefing, and an AI assistant is no different.
In this guide I will show you how to set up a personal AI marketing assistant that actually helps: a brand context file, the right workspace (Projects, plugins, Gems or skills), standing instructions, a list of tasks worth delegating, safe data connections and a fact-checking routine. I draw on an Ahrefs write-up of 200+ hours with an AI marketing assistant and on the official help pages for ChatGPT, Claude and Gemini. Some of those tools changed a lot in autumn 2026, so I will flag what is new.
Assistant or agent: decide what you are actually building
These two words get mixed up, so let me separate them in plain terms.
- An assistant works with you. You ask, it drafts, you check and decide. It sits in a chat window with your context loaded.
- An agent works for you. You give it a goal and it takes several steps on its own: searching, clicking, pulling data, sometimes publishing.
This guide is about the assistant. It is where most marketers should start, because you stay in control of every output. Once your assistant is reliable, you can automate parts of it. Our guide to AI agents in marketing covers that next step, and if you want to try building one for SEO, see how to build an SEO AI agent.
The Ahrefs author, who logged about 248 chat sessions and 17,000 messages with his assistant since April 2026, makes a point I think every marketer should hear:
“The model isn’t the smart one in this duo.”
— Mateusz Makosiewicz, Ahrefs, AI marketing assistant
In other words, your knowledge and judgment are what make the assistant useful. The setup below is about getting that knowledge into it.
The brand context file: one document that improves every answer
Before you touch any settings, write one document. I call it the brand context file. It is the briefing you would give a new team member on day one, written so a machine can use it too.
Here is what I suggest you include, in this order:
- Company basics: what you sell, where, since when, and for whom. Two or three sentences.
- Products or services: names, what each does, price ranges you are happy to state publicly, and what you don’t offer.
- Audience: two or three customer types, their main problems and the words they use for them.
- Positioning: why customers pick you over the two or three competitors they usually compare you with.
- Voice: five adjectives, a few “we say / we don’t say” examples, spelling (US or UK) and formatting rules.
- Proof: real facts you can cite, such as reviews, awards, years in business, with sources.
- Red lines: claims you must never make (for example medical or financial promises), and topics to avoid.
- Two or three examples of your best past content, pasted in full.
Keep it to two to four pages. Longer files get diluted. If you need more detail, split it into separate files: one for voice, one for products, one for audience. Our guide to defining your brand voice helps with section 5, and an editorial style guide can become its own file.
Ahrefs describes doing something similar by sharing favorite articles and guidelines so the assistant builds a writing profile, then calling that profile as a reusable skill when writing. Same idea, different packaging.
Where to keep your context: Projects, plugins, Gems and skills compared
Each major AI tool has its own place to store context. This changed a lot in late 2026, so check the current state before you build anything.
| Tool | Feature | What it holds | Worth knowing (as of October 2026) |
|---|---|---|---|
| ChatGPT | Projects | Chats, files and project instructions | File limits: 5 (Free), 25 (Go/Plus), 40 (Pro, Business, Enterprise). Optional project-only memory, chosen at creation |
| ChatGPT | Custom GPTs | Instructions, knowledge files, actions | Scheduled to retire on December 11, 2026; being replaced by plugins |
| ChatGPT | Plugins | Skills (instructions + reference files), apps/connectors | Migrated GPTs become plugins; instructions turn into a skill |
| Claude | Projects | Project knowledge (files) and project instructions | Free users can create up to five; paid plans switch on retrieval when knowledge gets large; each project has its own memory |
| Gemini | Gems | Instructions plus files from your device, Drive or NotebookLM | From November 2026, Gems on personal accounts move automatically to skills |
Sources: Projects in ChatGPT, OpenAI custom GPT retirement FAQ, Claude projects, Gemini Gems help (all accessed October 2026).
What the custom GPT retirement means for you
If you were planning to build a custom GPT as your assistant, don’t start one now. OpenAI’s FAQ says custom GPTs are scheduled to retire on December 11, 2026, and that creators can migrate them to plugins. According to Campus Technology, a GPT’s instructions become a skill and its knowledge files become reference files, while conversations, the selected model and sharing settings don’t carry over. Custom actions may need rebuilding.
The bigger lesson: keep your brand context file as a plain document you own. Then moving from a GPT to a plugin, or from a Gem to a skill, is a copy-and-paste job instead of a rebuild.
One project per job, not one for everything
Claude’s help page makes an important point about how context works inside a project:
“Context is not shared across chats within a project unless the information is added into the project knowledge base.”
— Claude Help Center, What are projects?
So if you discover something useful in one chat, such as a better headline style, add it to the project files. I suggest separate projects for separate jobs, for example “Blog and SEO content,” “Social posts,” “Ads copy” and “Reporting,” all sharing the same core brand file plus a job-specific file.
Writing standing instructions the assistant will actually follow
Project instructions (or a skill’s instructions) are the rules the assistant applies to every chat. Most people write them too vaguely: “Write in a friendly, professional tone.” Every brand says that.
Better instructions are specific and testable. Here is an example for a blog-writing project:
You write for [Brand], a [what you do] for [audience].
Always read brand-context.md before answering.
Use US spelling. Short paragraphs of 2–4 sentences.
Open with the reader's problem, not with "In today's world".
Never invent statistics, quotes, customer results or features.
If a fact is not in the files, say "[NEEDS SOURCE]".
Prefer plain words; explain any technical term once.
End drafts with a list of claims that need checking.Notice the two lines about facts. They do more for quality than any tone instruction, because they tell the assistant to flag gaps instead of filling them with confident guesses.
Daily tasks worth handing to your assistant
Not every task is a good fit. I judge each one on two questions: how much time does it take, and how bad is it if the assistant gets it wrong?
| Task | Good fit? | Risk if wrong | Your check |
|---|---|---|---|
| First drafts of social posts from a blog article | Yes | Low | Read before scheduling |
| Rewriting ad headlines into 15 variations | Yes | Low | Check character limits and claims |
| Summarizing call notes or reviews into themes | Yes | Medium | Spot-check against the source |
| Content briefs from keyword lists | Yes | Medium | Check search intent yourself |
| Writing statistics-heavy articles | With care | High | Verify every number and link |
| Replying to customers or reviews | Drafts only | High | Human sends every reply |
| Legal, medical or financial claims | No | Very high | Expert review required |
For more ideas, our overview of AI in marketing lists uses by channel, and our roundup of AI marketing tools covers specialist options. Our list of ChatGPT prompts for SEO is a good source of ready-made tasks.
Weekly and monthly workflows that save real time
Daily tasks save minutes. Repeatable workflows save hours. The Ahrefs author built things like a weekly Reddit listening digest, a keyword research hub and a dashboard of AI citations with a weekly summary. You don’t need to build apps to get value. A saved prompt plus a fixed input does most of the work.
A worked example: the Monday content repurposing routine
Here is a hypothetical routine for a small agency publishing one blog post a week. Times are estimates for illustration.
- Input: paste the finished blog post into the “Social posts” project (1 minute).
- Ask for: 3 LinkedIn posts, 5 short social posts, 1 newsletter section and 3 ad headline ideas, following the brand file (2 minutes to run).
- Review and edit: fix tone, cut weak lines, check every claim against the article (25 minutes).
- Schedule: load into your scheduling tool (10 minutes).
If writing all of that from scratch used to take about 2.5 hours, and the new routine takes about 40 minutes, you save around 1 hour 50 minutes a week, or roughly 95 hours a year. The review step is not optional, and it is where most of the 40 minutes goes. That is the right balance. Our guide to content repurposing has more formats to plug in.
Monthly jobs to set up next
- Summarizing monthly analytics exports into a plain-English report draft.
- Comparing your top pages with competitor pages for content gaps.
- Reviewing a month of customer reviews for new pain points and phrases.
- Checking what AI tools say about your brand, using a fixed list of prompts. Our AI search strategy guide explains how to set that list up.
Connecting your data without leaking it
Connecting Google Drive, analytics or your CRM makes an assistant much more useful. It also creates risk. Here is how I would approach it:
- Start read-only. Let the assistant read reports before you ever let it write, send or publish.
- Share folders, not your whole drive. Create a dedicated “AI assistant” folder with only what it needs.
- Keep customer personal data out unless your plan and company policy explicitly allow it. Remove names and emails from exports.
- Check your plan’s data settings. Business and enterprise plans usually offer different data controls from personal plans. Read them before uploading client material.
- Know what migrates and what doesn’t. OpenAI notes that custom actions don’t transfer automatically to plugins and may need a custom MCP server. Our guide to MCP for marketers explains what that means in plain words.
Here is the access level I would start with for the most common connections:
| Data source | Starting access | Why | When to expand |
|---|---|---|---|
| Google Drive or SharePoint | One dedicated folder, read-only | Keeps contracts and HR files out of reach | When you need it to save drafts back |
| Analytics exports (GA4, Search Console) | Uploaded CSVs, not live access | You control exactly which rows it sees | After a month of reliable summaries |
| CRM or email platform | None at first | Contains personal data | Only with anonymized fields and company approval |
| CMS or social scheduler | Drafts only, never publish | A wrong post goes live instantly | Rarely; keep a human on the publish button |
Be aware of how memory works too. In ChatGPT, a project can use project-only memory, which keeps it from referencing your other chats. It has to be chosen when the project is created, and shared projects use it automatically. If you handle several clients, I suggest one project per client with project-only memory, so nothing leaks between them.
Checking the work: a fact-check routine for AI drafts
An assistant that is wrong 5% of the time is more dangerous than one that is wrong 50% of the time, because you stop checking. Build the check into the workflow.
- Ask for a claims list. Every draft ends with the facts it relies on. Your instructions already require this.
- Verify numbers at the source. Open the original page. If the assistant can’t give a source, remove the number.
- Check names, prices and dates against your brand file and competitor websites.
- Read it out loud. AI drafts often have a rhythm that sounds off when spoken.
- Add something only you know: an opinion, an example from your work, a local detail. This is what stops the content becoming generic.
That last step matters for search too. Google’s spam policies describe scaled content abuse as generating many pages mainly to manipulate rankings. Publishing unedited AI drafts at volume is a fast route into that territory. Our guide to AI content and SEO covers where Google draws the line.
What tends to go wrong after the first 100 hours
Once the novelty wears off, a few problems show up again and again. The Ahrefs author puts it well: creating something is easy, keeping it useful is the real work. Here is what I would watch for:
- Stale context. Prices, products and team members change. Put a “last updated” date at the top of your brand file and review it monthly.
- Instruction creep. Every problem gets a new rule until the instructions contradict each other. Rewrite them from scratch every quarter.
- Same-sounding content. Over time, every post starts with the same structure. Rotate your example articles and ask for different formats.
- Too many projects. Twenty half-maintained projects are worse than four good ones. Archive what you don’t use.
- Skipped checks. When outputs look good for weeks, people stop fact-checking. That is exactly when a wrong number goes live.
- Tool lock-in. Platforms change, as the GPT and Gem changes show. Keep your files and prompts in a folder you control.
Questions people ask when setting up an AI marketing assistant
Which tool should I use: ChatGPT, Claude or Gemini?
Use the one your team already pays for and likes. All three support stored instructions and files. Your context file matters far more than the tool choice.
Should I still build a custom GPT?
No. OpenAI has scheduled custom GPTs to retire on December 11, 2026, and is moving them to plugins. Use a Project now and a plugin or skill when ready.
How long does the setup take?
A first useful version takes an afternoon: two to three hours to write the brand file, one hour to set up projects and test them.
Can the assistant write my blog posts on its own?
It can draft them, but a person should check every fact and add real experience. Unedited AI content at scale can hurt quality and search performance.
Is it safe to upload client data?
Only if your plan’s data settings and your client agreements allow it. Remove personal data and use project-only memory for client work.
Your first afternoon of AI assistant setup
To recap: an AI marketing assistant is only as good as the briefing you give it. Write a short brand context file, store it in a Project, plugin or skill, give it specific instructions, delegate the right tasks, connect data carefully and check every fact.
Set aside one afternoon this week. Write the brand file, create one project for your most repetitive task, and run it on three real jobs. Fix the instructions after each one. By Friday you will know whether it saves you time.
If you would rather have experts set this up for your team, or you want help producing content that holds up, our content writing team and social media management service use these workflows every day with human editors. Get a free quote and tell us what you want to speed up.
Further reading on ai marketing assistant
- Agent-to-Agent Marketing: When AI Agents Talk to Other AI Agents
- Agentic AI vs Generative AI: The Difference for Marketers
- From SEO Specialist to Search Visibility Lead: Skills for the AI Era
Sources for this AI marketing assistant guide
- Ahrefs – AI marketing assistant (200+ hours) (Mateusz Makosiewicz, June 15, 2026, updated July 3, 2026)
- OpenAI Help Center – Projects in ChatGPT (accessed October 2026)
- OpenAI Help Center – Custom GPT retirement and migration FAQ (accessed October 2026)
- Campus Technology – OpenAI discontinuing custom GPTs and moving to plugin architecture (October 5, 2026)
- Claude Help Center – What are projects? (accessed October 2026)
- Google Gemini Help – Use Gems in Gemini Apps (accessed October 2026)
- Google Search Central – Spam policies for Google web search (updated August 28, 2026)
