13 Best MCP Servers for Marketing & Website Analytics (2026)

1. Plerdy MCP Server: best for finding where a website loses sales

Plerdy MCP Server connects your Plerdy account to Claude, ChatGPT, Cursor, VS Code and any other client that supports remote MCP. It gives the AI 19 read-only tools over the same data you see in Plerdy: traffic and channels, lost revenue by page and source, page reports with clicks, scroll and errors, a session finder, conversions, events, sales, forms, pop-ups, A/B test results, site errors, SEO issues and Google Search Console queries.

The difference from the other analytics servers on this list is that Plerdy answers in money and causes, not only in sessions. Ask “where am I losing money this week?” and the AI calls the lost revenue tool, gets pages and traffic sources that bring visits but no sales, then opens the page report for the worst one to check clicks, scroll depth and JavaScript errors on it. The data comes from the same script that powers Plerdy’s heatmaps, session recordings, error tracking and Traffic Value, so the AI can connect a drop in sales to a broken checkout script on the same day.

Ten ready-made prompts come built in, including “Weekly money review,” “Is anything broken right now,” “Why conversions dropped,” “Fix my form” and “Can I trust this A/B test.” We wrote up all ten, with when to ask each, in 10 questions to ask ChatGPT about your website every week.

Here’s what a real answer looks like. We asked Claude, through Plerdy MCP Server, which referring sites sent visitors to plerdy.com between September 27 and October 26, 2026. Among 2,353 referral sessions it found perplexity.ai with 17 sessions, 4 of which reached the site goal. The tool also flagged that row as too small to judge, because Plerdy doesn’t draw conclusions below 30 sessions. That flag matters more than it looks: it stops the AI from telling you “Perplexity converts twice as well as other referrals” off 17 visits.

Where it falls short: Plerdy only sees what its own script collects. It won’t read your Google Ads spend or your CRM pipeline, so for cost per lead you’d pair it with the Google Ads or HubSpot server. It’s also read-only by design. The AI can tell you a form loses people at the phone field, but it can’t edit the form for you.

2. Google Analytics MCP: best free option for GA4 reports

Google’s Analytics team publishes an open-source MCP server for GA4, labeled “Experimental” in the repository title. It runs locally with pipx, signs in through Google Application Default Credentials, and uses the analytics.readonly scope. Nine tools cover account and property details, standard reports, funnel reports, custom dimensions and realtime data, plus an optional tool for property annotations.

What it does well: anything you can build in GA4’s Explore reports, the AI can now request in a sentence. “Compare organic sessions and key events by landing page, last 28 days against the previous 28” becomes one run_report call. The funnel tool is useful for checkout steps if your events are set up cleanly.

What it doesn’t do: GA4 has no heatmaps or recordings, so the AI can tell you a page converts badly but not why. The setup is the other cost. You need a Google Cloud project, enabled APIs and credentials on the machine running the server. Several vendors (Windsor.ai, Coupler.io, Stape and others) sell hosted GA4 connectors to get around this. If you also use Plerdy, its server already returns traffic and conversions per page, so you may not need GA4 in the chat at all for weekly questions.

3. Microsoft Clarity MCP: best free behavior data, with tight limits

Microsoft’s open-source Clarity server, published on npm as @microsoft/clarity-mcp-server, runs locally through npx with a Clarity Data Export API token. Its tools query dashboard metrics in natural language, list session recordings with filters by URL, device, browser, OS and country, and search Clarity’s documentation.

Clarity is free and its recordings are good, so the MCP server is an easy first test. The limits are the catch. Ten requests a day sounds like a lot until the AI spends four of them on one question, and three days of data means you can’t ask “how did last month compare with the month before?” at all. Clarity also has no revenue data, so it can describe behavior but can’t rank pages by money lost.

If you hit those walls, Plerdy MCP Server covers the same behavior questions (scroll, clicks, rage clicks, sessions with errors) with history that follows your plan’s retention and a daily limit of 100 to 600 requests on the main plans. We compared the two products in more depth in Microsoft Clarity vs Plerdy.

4. Contentsquare MCP: the official route to Hotjar data

Contentsquare, which owns Hotjar, runs a hosted MCP server that lets ChatGPT, Claude and other assistants query heatmaps, journeys and web analytics. hotjar.com now redirects to contentsquare.com/hotjar, and the AI connection is offered through the Contentsquare platform. Directory listings give the endpoint as api.contentsquare.com/mcp with OAuth sign-in.

Older lists say Hotjar has no MCP server. That’s half true: there’s no server with the Hotjar name, but Hotjar data is reachable through Contentsquare’s. Read the plan gates before you plan a workflow around it. 300 calls a month on Free runs out in about a week of daily questions, and the most useful tools sit on paid tiers, starting with Growth at $49 a month on Contentsquare’s pricing page. Contentsquare doesn’t publish the tool list without sign-in, so check it inside your account before you commit.

5. Google Search Console MCP (mcp-gsc): best for search queries, if you can run it

Google hasn’t released an official Search Console MCP server. The most used community option is mcp-gsc by Amin Foroutan, MIT-licensed with about 2,000 GitHub stars. It runs locally, signs in with OAuth or a service account, and covers search analytics, period comparisons, page-by-query breakdowns, URL inspection, indexing checks and sitemaps.

Questions like “which queries have high impressions and a CTR under 1%?” or “which pages lost clicks since the last core update?” are where this server earns its place. It’s also the only server here that can submit a sitemap, which is handy and a reason to keep the destructive tools off.

There’s a shortcut if you don’t want to run community code. Plerdy syncs Search Console into its SEO Checker, and Plerdy MCP Server exposes that data through its search performance tool: queries, clicks, impressions and CTR per page, read-only, with nothing installed. You lose URL inspection and sitemap tools, and you gain the same chat that already knows the page’s conversion rate and errors.

6. Ahrefs MCP: best for backlink and keyword research

Ahrefs runs a hosted MCP server at api.ahrefs.com/mcp/mcp over Streamable HTTP. Connecting goes through an OAuth consent screen, after which Ahrefs issues an API key scoped to MCP. Ahrefs publishes setup guides for Claude, ChatGPT, Copilot Studio, n8n and others, and its older local server on GitHub is no longer maintained.

Ahrefs is the strongest pick when the question is about other websites: who links to a competitor, which keywords they rank for that you don’t, what a domain’s traffic trend looks like. Watch the API units. An AI exploring a broad question can make many calls, and the units are shared with everything else on your account. Ahrefs knows nothing about what happens after the click, so it tells you which keywords to chase but not whether that traffic buys.

7. Semrush MCP: best for competitive research, now under Adobe

Semrush hosts an official MCP server on mcp.semrush.com with OAuth sign-in, or an API key in the Authorization header for clients without OAuth. It gives assistants keyword, domain and competitor data from Semrush’s databases. Access depends on your plan: some Semrush One and SEO plans include API units, others need a separate API package.

Semrush’s server does the same job as Ahrefs’ for most marketers, so pick the one you already pay for rather than running both. Two things to know before connecting. First, check the exact endpoint in Semrush’s own docs on the day you set it up. Directories list three different paths (v1, v2 and a Claude-specific one), which suggests it has moved more than once. Second, Semrush is now an Adobe subsidiary, and Adobe has said it plans to fold Semrush into Adobe Experience Manager and its LLM Optimizer. The product still runs as before, but if you’re signing a multi-year contract, ask about plan changes.

8. Google Ads MCP: best for reading campaign data safely

Google documents an official Google Ads MCP server in its developer toolkit, with code at github.com/googleads/google-ads-mcp. It’s read-only: three tools list accessible accounts, run GAQL queries for metrics, budgets and status, and return metadata about resource types. It runs locally over stdio or on your own Google Cloud Run service over Streamable HTTP.

Read-only is a feature here. An AI that can’t pause a campaign also can’t pause the wrong one. The real limit is that Google Ads stops at the click: it reports conversions you’ve imported, but not what visitors did on the landing page. That’s the gap Plerdy’s Traffic Value fills. It ties each channel and landing page to sessions, conversions and revenue per session, and labels the problem as a traffic problem or a landing page problem. We showed the method in how to find wasted ad spend by landing page.

9. Meta Ads MCP: best for Facebook and Instagram ads, with write access

Meta hosts its own Ads MCP server at mcp.facebook.com/ads. Meta’s developer documentation says it works with any MCP-compatible agent and covers reporting and insights as well as creating and editing campaigns, ad sets and ads, through permission-scoped tools. You sign in with your Meta account.

Third-party guides date the launch to late April 2026. It’s the only official ads server on this list that writes, which makes it the most useful one and the riskiest. A sensible rule is to use it for reporting first (“which ad sets had a rising CPM and a falling CTR this week?”) and only allow changes after you’ve watched how the assistant phrases them. Like Google Ads, Meta can’t see the landing page, so pair it with a behavior tool when the question is “why don’t these clicks buy?”

10. HubSpot MCP: best for CRM and pipeline questions

HubSpot’s remote MCP server at mcp.hubspot.com became generally available to all HubSpot accounts on April 13, 2026. It reads CRM records, marketing and content objects and activity history, and it writes contacts, companies, deals, tickets, line items and products, plus calls, meetings, notes, tasks and emails. Sign-in uses OAuth 2.1 with PKCE only.

HubSpot closes the loop the other servers leave open: a lead came in, but did it turn into a deal? Questions like “which source brought the deals we closed this quarter?” or “list contacts from the webinar who haven’t been called” work well. Don’t confuse it with HubSpot’s separate developer MCP server, which is for building apps, not reading your CRM. And since it writes, give it to people who’d be trusted to edit the CRM by hand anyway.

11. PostHog MCP: best for product-led and developer teams

PostHog hosts its MCP server at mcp.posthog.com/mcp and routes you to the US or EU region based on your login. Tools cover analytics queries, SQL, feature flags and experiments, session replay, error tracking and CDP destinations, and agents can both read and change things in PostHog. Supported clients include Claude, Cursor, Codex, VS Code, Windsurf and Zed.

PostHog’s server is the widest on this list. That’s a strength for a product team and a trap for a marketer, because a broad toolset means the AI has more ways to answer a vague question wrongly. Ask narrow questions (“signups by utm_source last week, excluding internal users”) and check the SQL it writes. For marketing sites without PostHog already installed, this is not the first server to add.

12. Amplitude MCP: best for teams that build charts and cohorts

Amplitude’s MCP server runs at mcp.amplitude.com/mcp, with an EU endpoint, and signs in through OAuth with your existing Amplitude permissions. It reads charts, dashboards, notebooks, experiments, cohorts, events and session replay, and it can create and edit charts, dashboards, cohorts, experiments, feature flags and metrics. ChatGPT gets a curated subset of about 25 tools.

The write side is what sets Amplitude apart for analysts: “build a dashboard with weekly signups, activation and retention by plan” produces an actual dashboard you can share. For marketers, the read side is enough. Ask for a funnel by campaign, then ask the AI to save it as a chart so the team sees the same numbers next week.

13. Mixpanel MCP: best for event analytics on a free plan

Mixpanel hosts MCP endpoints in the US, EU and India, with OAuth sign-in and a beta service-account option for automated agents. The AI can read data and make changes such as creating dashboards and editing Lexicon metadata, and every change is logged with an MCP origin in the audit log. The limit is 600 MCP requests per hour per user.

Mixpanel’s audit log is a detail more vendors should copy: when an assistant renames an event, you can see that it was the assistant. The free-plan access makes it the cheapest way to try a write-capable analytics server. Like PostHog and Amplitude, it only knows the events you send, so the answer is only as good as your tracking plan.