Custom GPT vs MCP Connector for Marketing: Why Pasted Brand Docs Lose to a Live Line into Your Marketing System

Most marketing teams built a Custom GPT in the first month they had ChatGPT Team. You know the one: “Brand Voice GPT”, instructions pasted from the style guide, the positioning doc and a few winning emails uploaded as knowledge. It helped. Then a teammate got an off-brand draft from it, then another, and someone said “just paste the brand doc into the chat” and you were back where you started.

That is not a prompt problem. It is an architecture problem. A Custom GPT is a snapshot: a fixed set of files and a fixed block of instructions, frozen the day you saved it, searched rather than read, and blind to what your marketing is doing this week. An MCP connector is a phone line: the assistant asks your marketing system for the brand, the calendar and the running automations at the moment it needs them, and hands a draft back.

This post lays out what each one can and cannot hold, runs the same prompt both ways so you can see the difference, and names the cases where a Custom GPT is still the right call.

What a Custom GPT actually is

A Custom GPT is a saved configuration of ChatGPT: a name, a set of instructions, up to a fixed number of uploaded files, and optionally some Actions that call an API. When you chat with it, ChatGPT reads the instructions, searches the uploaded files for passages that look relevant to your message, and answers with those passages in context.

Three properties follow from that design, and all three show up as “the GPT is inconsistent” in a marketing team.

The files are searched, not read. Retrieval pulls the chunks the model judges relevant. A long brand guide has many chunks, and the one with your banned-words list is not always the one that gets pulled. An agency that builds Custom GPTs for a living, Aspiration Marketing, puts it plainly: “sometimes the GPT references your brand guide, sometimes it ignores it entirely.”

The instructions are capped and they fade. OpenAI’s own community maintains a long-running thread on Custom GPT limits: a cap on the number of knowledge files, a character cap on instructions, and no way for the GPT to remember anything between chats. The thread’s word for what happens over a long session is “selective amnesia”: the model loses track of the rules you gave it at the top.

The files are frozen. Every developer-side comparison says the same thing. As Hjarni’s MCP vs Custom GPTs guide puts it, knowledge files “are frozen in time.” Your Q2 offer, the launch date that moved twice, the new pricing tier: none of it exists inside the GPT until someone re-uploads.

None of these are bugs. They are what a snapshot is.

What a Custom GPT cannot know at all

The three limits above are about the quality of the snapshot. There is a bigger gap, and it is the one that causes real damage: a Custom GPT has no idea what your marketing is doing.

It does not know that a webinar reminder is already scheduled for Tuesday morning. It does not know that the pilot-offer automation is live and has three steps, so the “follow-up email” it drafts for you duplicates step two. It does not know which campaign a piece belongs to, what send day it should land on, or that a draft is already waiting for approval. You can upload a screenshot of your calendar. By Thursday it is wrong.

This is the difference between an assistant that knows things about your marketing and one that can see your marketing. Only the second one can be trusted with anything that touches the schedule.

What an MCP connector is

MCP (Model Context Protocol) is the open standard that ChatGPT, Claude, Claude Code, Codex and Cursor all use to connect an assistant to another system. A connector exposes a set of named tools. When you ask the assistant something, it decides which tools to call, calls them, and works with the live result. Nothing is uploaded. Nothing goes stale, because the assistant fetches the current state on every turn.

OpenAI has been moving in this direction on its own: since November 2025, company knowledge in ChatGPT Business and Enterprise has been able to use custom MCP connectors as a source, per OpenAI’s help center. That is the read half. The interesting part for marketing is the write half: a connector can also hand work back into the system.

Marqeable’s connector, as an example of the shape, exposes ten read tools and six write tools. The reads cover your business context (brand, ICP, voice, offers), campaigns with their pieces and send days, semantic search over existing content, the content calendar with scheduled sends and market dates, and every automation with its plan. The writes create campaign drafts, content drafts, inline suggestions on existing copy, brief updates and automation drafts. Every write is a draft. There is no send, launch, publish or delete tool. That last property is the one to look for in any marketing connector, and there is a longer argument for it in the draft-only rule.

The same prompt, two ways

Take an ordinary Monday request: “Write the launch email for the Q4 webinar.”

In the Brand Voice GPT. The GPT searches its files, finds the voice rules (probably), finds the June webinar email (maybe), and writes a competent email addressed to the ICP as it was described in the upload. It uses the offer from the positioning doc, which changed in July. It does not know the webinar’s date, so it leaves a placeholder. It does not know that the reminder is already scheduled for the 16th. You copy the text into your email tool, fix the offer, add the date, notice the collision yourself if you are lucky, and set up the send.

With the connector. The assistant calls get_business_context and gets the current offer and the current ICP. It reads the calendar for the month and sees the invite on the 10th and the reminder on the 16th. It searches existing content, finds the June webinar sequence, and reuses the structure that worked. Then it creates the email as a content draft attached to the webinar campaign for the 17th, copy only, so no images are generated until you approve the words. The result is a link to the draft in Marqeable.

Here is roughly what that exchange looks like:

You: Run whoami, then get_business_context, then tell me what is on the calendar this month.

ChatGPT: You are signed in as the marketing lead on the Northwind account. Current offer is the 30-day pilot; ICP is the first controller at a 50 to 200 person company; voice is direct and specific. September calendar: nurture email on the 4th, webinar invite on the 10th, reminder on the 16th, LinkedIn case study on the 18th. The pilot-offer automation is live with three steps.

You: Write the post-webinar email. Reuse what worked in June.

ChatGPT: I found the June follow-up; its two-question recap had the strongest structure, so I kept it and rewrote for this topic and the pilot offer. I created it as a content draft on the webinar campaign for the 17th, the day after the reminder, copy only. It is waiting for your approval here: [link to the draft].

Same request, same model. The difference is entirely in what the assistant could see and where the result landed.

Snapshot vs phone line: what each one can hold

The middle column matters. Projects and company knowledge fix freshness for documents, and you should use them for the guidance that genuinely lives in documents: the strategy memo, the messaging pillars, the long-form voice guide. What they cannot do is the third and fourth rows. Your marketing is not a document. There is a fuller treatment of that split in why a marketing knowledge base is not enough for AI agents.

When a Custom GPT still wins

It would be dishonest to pretend the snapshot is always the wrong tool. Three cases where it is the right one:

1. A shareable persona for people with no login. A tone checker for freelance writers, a headline brainstormer for the sales team, an internal FAQ over the onboarding docs. If the answer never depends on what is scheduled or live, the GPT is faster to set up and safer to hand out.
2. A stable reference set. A GPT over your case-study library or your competitive battlecards is useful for months, because those files change slowly. Re-uploading quarterly is fine.
3. Zero-write jobs. If you only want text back and you will always carry it somewhere yourself, the connector’s write half buys you nothing.

The failure mode is using a GPT for the jobs in the table’s bottom rows: anything that touches the calendar, an automation, or a campaign that already exists. That is where the inconsistency you have been blaming on prompts actually comes from, and it is the same mechanism behind brand voice drift in long ChatGPT sessions.

How to connect ChatGPT to your marketing

The connector lives at https://app.marqeable.com/mcp. It is a remote server over HTTP with OAuth 2.1, which means you sign in with your normal Marqeable login in a browser window and ChatGPT sees exactly what your account can see. There are no API keys to paste into a chat.

As of September 2026, in ChatGPT on a plan that supports custom connectors: open Settings > Connectors, and under Advanced turn on Developer mode. Click Create, name the connector “Marqeable”, paste the address, choose OAuth as the authentication method and save. Sign in to Marqeable when prompted, then enable the connector from the tools menu in any conversation. OpenAI has moved this setting more than once this year (it has also appeared under Apps and under Plugins), so if the path differs, search Settings for “Developer mode”. On Business plans an admin may need to allowlist the connector first, and ChatGPT may ask you to confirm write actions per conversation. Because every write is a draft, that confirmation is a second gate, not the only one.

If your team has been pasting ChatGPT output into five tools to make a campaign real, the before-and-after is written up in ChatGPT wrote my campaign, now what?.

Frequently asked questions

What is the difference between a Custom GPT and an MCP connector?

A Custom GPT is a saved configuration of ChatGPT: instructions plus a fixed set of uploaded files that the model searches when it answers. An MCP connector is a live connection from ChatGPT to another system, so the model can call that system for current information and hand work back into it. The GPT is a snapshot; the connector is a phone line.

Why does my brand voice Custom GPT give inconsistent output?

Knowledge files are retrieved by search, not read in full, so the model pulls the passages it judges relevant and sometimes misses your brand guide entirely. Instructions are capped in length and their influence fades over a long conversation. And the files are frozen at upload, so any offer, date or positioning that changed since then is wrong in the GPT.

When is a Custom GPT still the right choice for marketing?

When the job is a stable, shareable persona with no need for live data: a tone-of-voice checker for freelancers, a headline brainstormer, an internal FAQ bot over a document set. If the answer never depends on what is scheduled or live, a GPT is cheaper to set up and easier to hand to a contractor who has no login to your systems.

Can a ChatGPT connector send my campaigns without me?

Marqeable’s connector cannot. It has read tools for your business context, campaigns, content, calendar and automations, and write tools that only create drafts. There is no send, publish or launch tool, so every draft waits for a person to approve it in Marqeable.

The bottom line

The Brand Voice GPT was a reasonable first move, and for shareable, stable, text-only jobs it still is. But the inconsistency your team keeps hitting is built into the snapshot: searched files, capped instructions, frozen offers, and no view of the calendar or the automations. A connector replaces the snapshot with a live line, and a good one hands the result back as a draft you approve rather than text you carry. Keep the GPT for the persona. Connect the assistant for the marketing.

See the live line in practice: Marqeable’s agents build and run your campaigns, run behavior-triggered automations across text and email, answer and qualify every visitor with AI website chat, and tie revenue to the exact message with attribution.

Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access