Many marketers ask me some version of this: “Is MCP something I need a developer for, or can I actually use it myself?” They have seen the term everywhere, usually next to words like “agents” and “protocol,” and assume it is a coding project. It isn’t, at least not for most of what marketers want to do.
MCP is the reason you can now ask Claude or ChatGPT, “Which landing pages lost organic clicks last month, and what keywords did they drop for?” and get an answer built from your real data, not a guess. That is a big change for reporting, audits and research.
Let me explain what MCP is in plain words, which ready-made connections marketers can use today, how to set one up safely, and where it still goes wrong. This guide is about using existing MCP servers. If you are wondering whether your business should publish its own, I cover that separately in should your business publish an MCP server?
MCP in plain words: a universal plug between AI and your tools
MCP stands for Model Context Protocol. A protocol is just an agreed way for two pieces of software to talk. MCP is the agreed way for an AI assistant to talk to other tools, such as your analytics, your SEO platform or your CRM.
Before MCP, every AI tool needed a custom integration for every app. Now an app builds one MCP “server,” and any AI assistant that supports MCP can use it. Ahrefs puts it neatly:
“An MCP server is like a universal adapter.”
— Mateusz Makosiewicz, Ahrefs, What Is an MCP Server, and Why Should Marketers Care?
There are three parts to know. You don’t need to remember the names, but they help when you read setup guides:
- The host is the AI app you type into, such as Claude or ChatGPT.
- The client is a small connector the host creates for each tool you connect.
- The server is the part the tool provides. It lists what the AI is allowed to do, such as “get top pages” or “run a report,” and it does the work.
Servers come in two flavours. A remote server runs on the vendor’s side and you connect with a web address and a login. A local server runs on your own computer and usually needs a bit more setup. For most marketers, remote servers are the easy place to start.
Why MCP changes day-to-day marketing work
The big benefit is not magic insight. It is removing the copy-and-paste between tools. Most marketing analysis today looks like this: export from GA4, export from Search Console, export from your SEO tool, paste into a spreadsheet, then try to spot patterns. MCP lets the assistant pull those numbers itself and do the joining for you.
The second benefit is accuracy. When the assistant fetches real numbers, it doesn’t have to invent plausible ones. Ahrefs makes this point: “The AI pulls real data from your tools instead of hallucinating plausible-sounding answers.” That is true as long as you tell it to use only the data it fetched, which I come back to below.
The third benefit is speed for one-off questions. Questions that used to wait for the monthly report (“Did the new pricing page get any AI referral traffic?”) can be answered in a minute.
MCP servers marketers can connect today
Here are the connections that matter most for marketing work, based on official documentation and the list in Ahrefs’ guide. Availability changes quickly, so always check the vendor’s own page before relying on any of this.
| Tool | Type of server | What you can ask | Notes worth knowing |
|---|---|---|---|
| Google Analytics (GA4) | Official, from Google | Users, sessions, top pages, conversions by date or channel | Google says it handles read requests only and cannot change your GA settings |
| Google Search Console | Community-built (e.g. mcp-gsc, listed by Ahrefs) | Queries, clicks, impressions, page performance | Not an official Google product, so review it carefully before use |
| Ahrefs | Official, remote | Keywords, backlinks, competitor traffic, AI visibility data | Needs a paid plan from Lite up; calls use your API units |
| Semrush | Vendor-provided | Keyword, domain and audit data | I walk through this in running an SEO audit with Claude and the Semrush MCP |
| HubSpot, Salesforce, ActiveCampaign | Official (per Ahrefs’ list) | Contacts, deals, campaign results | CRM data includes personal data, so permissions matter most here |
| Zapier | Official bridge | Actions in other apps, e.g. Mailchimp and Clay | Useful when a tool has no MCP server of its own |
| Canva, Asana, Linear | Vendor connectors (per Ahrefs’ list) | Create designs or tasks from a chat | These are write actions, so approve each one |
A quick word on Ahrefs, since it is one of the more mature options. Its documentation says “Ahrefs MCP lets AI agents securely access the Ahrefs API.” It lists Claude, ChatGPT, Microsoft Copilot Studio and several other clients, and notes that keys created through the sign-in flow default to 200,000 API units per month, which an admin can change (Ahrefs MCP docs, accessed October 2026).
Which AI assistants support MCP, and on which plans
This is where people get stuck, because support differs by app and by plan. Here is what the official help pages say as of October 2026.
| Assistant | How you add a server | Plan limits (per official docs) | Write actions? |
|---|---|---|---|
| Claude (web, desktop) | Customize → Connectors → Add custom connector, paste the server URL | Available on Free, Pro, Max, Team and Enterprise; Free is limited to one custom connector | Yes. Claude asks for approval; you can choose “Allow always” for trusted tools |
| Claude Desktop (local) | Settings → Extensions, install a reviewed extension or a .mcpb file | Team and Enterprise owners can allowlist or block extensions | Depends on the extension |
| ChatGPT | Add a custom MCP server from the apps/plugins area on the web | Business, Enterprise and Edu get full MCP on web; Pro is limited to read/fetch; mobile is not supported | Yes on business plans, with confirmations; some risky actions are blocked |
Sources: Claude custom connectors help, Claude local MCP help and OpenAI’s MCP connectors help. Ahrefs also lists Perplexity’s Mac app, the Gemini CLI and Copilot Studio as MCP-capable hosts.
Setting up your first MCP connection without creating a security hole
Here is how I would set up a first connection, using a read-only analytics or SEO server as the example. It takes about 15 minutes.
- Pick a read-only use first. Start with reporting, not anything that can change settings, send emails or edit your site.
- Use the official server where one exists. Find it on the vendor’s documentation site, not a random directory. Copy the server URL from there.
- Add it in your assistant. In Claude: Customize → Connectors → Add custom connector. In ChatGPT business plans: add a custom MCP server from the apps area. Name it clearly, for example “Ahrefs – read only.”
- Read the permission screen. When the sign-in (OAuth) screen appears, check what access is requested. Claude’s help page advises you to limit scopes and deny unnecessary access.
- Sign in with the right account. Use an account that only has access to the properties you need. If you manage client accounts, don’t connect your master login to everything at once.
- Switch off tools you don’t need. Most servers expose several tools. Disable the ones that write, delete or send.
- Test with a question you already know the answer to. For example, “How many sessions did the site have last week?” Compare the answer with GA4 itself.
- Write down what you connected. Keep a simple log of which servers, which accounts and which scopes. You will need it when someone leaves or a client asks.
If you are new to the AI assistant side generally, my guide to setting up a personal AI marketing assistant covers projects, instructions and memory, which work well alongside MCP.
Three marketing workflows that MCP makes faster
Here are three workflows I suggest trying, with example prompts. Keep each prompt focused on one job, because, as Ahrefs notes, overloaded prompts with many actions produce more errors.
Monthly reporting: from five exports to one conversation
Connect GA4 and your SEO tool, then ask:
Using only data from the Google Analytics connection, list the 10 landing pages
with the biggest drop in organic sessions, last 28 days vs the previous 28 days.
Show the numbers in a table. If any data is missing, say so instead of estimating.Then follow up: “For those 10 pages, pull the top keywords from Ahrefs and show which ones lost position.” You now have the first draft of a report section, which you check and annotate. My guide to what to include in an SEO report covers how to turn that into something a client reads.
Quick audits: find the pages that need attention
With a Search Console server connected, ask for pages with high impressions and low click-through rate, then ask the assistant to group them by likely cause (title doesn’t match the query, AI Overview present, page type mismatch). This is a starting point for a human review, not the review itself.
Research: competitor and keyword questions in minutes
With an SEO tool connected, try: “Which competitors grew organic traffic fastest in the last six months, and which pages drove it?” Ahrefs lists this as one of its favourite uses. For AI visibility specifically, you can also ask how often your brand is mentioned versus linked, which I explain in AI mentions without links.
Example: what one reporting workflow saves
This is a hypothetical example with assumed timings. Say an in-house marketer spends about three hours each month pulling GA4, Search Console and Ahrefs data into a spreadsheet for a 12-page report. With MCP, the pulling and joining might take 30 minutes of prompting, plus 45 minutes of checking numbers against the source tools. That is roughly 1 hour 15 minutes instead of 3 hours, or about 21 hours saved a year. The checking time doesn’t go away, and it shouldn’t.
Permissions and client data: what to allow and what to keep off
This is the part I would not skip. An MCP connection gives an AI assistant the same access as the account you sign in with. If that account can delete a campaign, so can the assistant, by mistake or because it was tricked.
“Only give tools the minimum permissions they need to work, because AI could accidentally delete things”
— Mateusz Makosiewicz, Ahrefs, What Is an MCP Server, and Why Should Marketers Care?
Here is a simple permission plan I suggest for marketing teams:
| Use | Allowed access | Approval setting | Keep in mind |
|---|---|---|---|
| Analytics and SEO reporting | Read-only | “Allow always” is reasonable for trusted, official servers | Check a sample of numbers against the source each month |
| CRM lookups | Read-only, limited fields where possible | Ask each time | CRM data is personal data under privacy laws |
| Task creation (Asana, Linear) | Write, scoped to one project | Ask each time | Low risk, but review before it creates 40 tasks |
| Ad platforms, email sends, site edits | Avoid write access for now | Never auto-approve | Mistakes cost money or reach customers directly |
On client data: before connecting a client’s GA4 or CRM, check your contract and their privacy policy. You are sending their data to an AI provider. My overview of GDPR, CCPA and India’s DPDP Act basics is a useful starting point, but get the client’s written OK.
Prompt injection and other ways MCP setups go wrong
MCP is powerful, and that also makes it a target. The main risk has a name: prompt injection. This is when text the AI reads, such as a web page, an email or a tool description, contains hidden instructions that the AI follows as if you had typed them.
Anthropic’s help page is blunt about this:
“Malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions.”
— Anthropic, Get started with custom connectors using remote MCP
OpenAI’s help centre gives a similar warning, saying unsafe or untrusted MCP servers “may increase exposure to security risks (including prompt injection).” Microsoft’s developer blog, in Protecting against indirect injection attacks in MCP (April 2025), also describes tool poisoning, where the description of a tool is written to trick the model, and warns that a tool’s definition can change after you approved it.
Beyond security, here are the everyday failure modes I would watch for:
- Wrong date ranges. “Last month” can be read as the last 30 days or the previous calendar month. Always state exact dates.
- Silent sampling or limits. APIs return limited rows. The assistant may not tell you it only saw the top 1,000.
- Filling gaps with guesses. If a call fails, some assistants carry on and estimate. Tell it to stop and say so.
- API costs. Tools like Ahrefs count MCP calls against your plan. A long exploratory chat can use more units than you expect.
- Changed tools. Vendors update servers. A prompt that worked in June may behave differently in October.
None of this means you should avoid MCP. It means you treat the assistant like a fast junior analyst: useful, but checked.
Where MCP is heading for marketing teams
A few trends are already visible in the official docs. More vendors are shipping official remote servers instead of leaving it to the community. Admin controls are growing: ChatGPT’s enterprise plans let admins switch individual actions on or off, with new actions off by default, and Claude’s Team and Enterprise plans have allowlists for desktop extensions. And assistants are adding confirmations for write actions rather than letting them run freely.
The other direction is websites becoming usable by AI agents directly, which is a different but related idea. I explain it in my guide to WebMCP. For teams that want to go further and automate whole tasks, see Claude Code for SEO and how to build a simple AI agent for SEO.
My honest view: in a year, connecting your analytics to your AI assistant will feel as normal as connecting Search Console to GA4. The teams that start now, carefully and read-only, will have their checks and habits in place when it does.
Questions marketers ask about MCP
Do I need to code to use MCP?
No, not for official remote servers. You paste a URL into your assistant’s connector settings and sign in. Local and community servers may need more technical setup.
Is MCP free?
The protocol is free, but the tools are not. Your AI plan must support connectors, and data tools like Ahrefs count MCP calls against your API allowance.
Can MCP change my Google Analytics settings?
Not with Google’s official GA MCP server. Google says it supports read requests only and cannot modify your configuration or settings.
Is it safe to connect client accounts?
It can be, with read-only access, official servers, limited scopes and the client’s written agreement. Avoid write access to ad accounts or email tools.
Does ChatGPT support MCP on every plan?
No. According to OpenAI’s help centre, full MCP support is on Business, Enterprise and Edu plans on the web. Pro users are limited to read and fetch actions, and mobile is not supported.
Your first week with MCP: a safe starting plan
If I were starting this week, I would connect one read-only source, probably GA4 or my SEO tool, to Claude or ChatGPT. I would ask five questions I already know the answers to, and compare. Once the numbers match, I would build one monthly report section through MCP and keep checking it against the source.
That is enough to learn how it behaves before you trust it with anything bigger. If you would rather have a team set up AI-assisted reporting and audits for you, our SEO services and PPC management teams already work this way. Get in touch for a free quote and tell us what you would like to automate.
Further reading on MCP for marketers
- AI in Marketing Statistics: Adoption, Spend and Results
- AI Marketing Tools Worth Using: Picks by Task
- AI Mentions Without Links: How Often Assistants Link to Brands They Name
Sources for this MCP for marketers guide
- Ahrefs — What Is an MCP Server, and Why Should Marketers Care?
- Ahrefs — MCP documentation: introduction
- Google Analytics — Try the Google Analytics MCP server
- Anthropic — Get started with custom connectors using remote MCP
- Anthropic — Getting started with local MCP servers on Claude Desktop
- OpenAI — Developer mode, apps and full MCP connectors in ChatGPT
- Microsoft for Developers — Protecting against indirect injection attacks in MCP
