AI Visibility MCP Servers Compared: Profound, Peec, Otterly, Semrush and GEOHATS (2026)

Which AI visibility vendors ship an MCP server, what each one's tools can read and change, which plan you need, and how the full gap-to-page loop runs from Cursor or Claude Code.

Daniel Borodin·Updated ·13 min read
Contents (7)
  1. Why run GEO from the editor at all
  2. What MCP is, in 60 seconds
  3. Which AI visibility vendors ship MCP servers in 2026
  4. The full loop with GEOHATS's MCP server
  5. A copy-paste agent prompt that works
  6. Honest limits before you commit
  7. FAQ
Quick answer

Yes. You can run GEO workflows from Cursor or Claude Code today by connecting a remote MCP server over Streamable HTTP. For visibility data in chat, Peec includes MCP on every plan and Semrush on Semrush One from $199 a month. For changing prompts and Workflows from chat, Otterly (from Standard) and Profound (Enterprise) add write tools. For content, Trakkr includes an 82-tool server on every paid plan and AthenaHQ's server can start its content pipeline. GEOHATS, our own product, lets your agent save the page it writes, publish it to your CMS and fix issues the site crawl finds.

Why run GEO from the editor at all

GEO work is code-adjacent by nature: a citation gap becomes a brief, a brief becomes an HTML page, the page ships through a repo or a CMS, and a tracker measures whether AI assistants start citing it. Every step in that chain is something a coding agent in Cursor or Claude Code already does well. The editor is where the loop closes fastest.

Three practical reasons this beats a web dashboard for anyone technical:

  • Agents write better with full repo context. An agent sitting in your project sees your components, your routes, your tone, and your existing pages. It can match your site's design system exactly instead of producing generic output in a vendor's template.
  • No copy-paste tax. Without MCP, an editor-based GEO session means exporting CSVs from a dashboard, pasting data into chat, writing the page, then uploading it somewhere. With MCP the agent pulls gaps and briefs itself and hands back a finished draft.
  • Shipping is native. If your site is a Next.js, Astro or custom repo, the last mile of publishing is a file write and a git commit. That is the one step a SaaS dashboard fundamentally cannot do for you, and the one step your editor does best.

The catch: many AI visibility tools that added MCP in 2025 and 2026 treat it as a reporting channel. You can ask "what is my share of voice" from chat, which is useful, but the answer still ends with you leaving the editor to act. The difference that matters is read versus write scope, and the vendor comparison below sorts the servers by exactly that.

What MCP is, in 60 seconds

The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems, tools and data. The official docs describe it as "a USB-C port for AI applications": one protocol, many tools, so a server built once works in every compatible client (Model Context Protocol, 2026). Anthropic open-sourced it in November 2024 (Anthropic, 2024), and it is now supported by Claude, ChatGPT, VS Code, Cursor and dozens of other clients.

Two transport details matter for GEO tools:

  • Local servers run as a subprocess on your machine over stdio. Good for filesystem tools, wrong for SaaS data.
  • Remote servers use Streamable HTTP: a single HTTPS endpoint that accepts JSON-RPC over POST and can stream responses. This replaced the older HTTP+SSE transport in the 2025 revision of the spec (MCP Specification, Transports, 2025). All five servers in the comparison table below are remote servers you connect to with a URL (Profound also ships a local package).

Both editors covered here support remote servers natively. Cursor reads an mcp.json file with a url plus optional auth headers (Cursor Docs, MCP, 2026). Claude Code adds one with a single command, claude mcp add --transport http <name> <url>, with an optional --header "Authorization: Bearer ..." flag (Claude Code Docs, MCP, 2026). If you can paste one URL and one token, you can connect.

Which AI visibility vendors ship MCP servers in 2026

We read each vendor's own MCP documentation in August 2026 and re-checked it against their pages in November 2026. The short version: Peec and Semrush expose read-only data, while Otterly and Profound let an agent change settings such as prompts, audits and Workflows. GEOHATS's tools also save agent-written pages and publish them to a connected CMS. Four more trackers with MCP servers, including two that write content, follow after the table. Details below, with sources.

VendorRead toolsWrite toolsPrice gate for MCP
Otterly.AIBrand reports, citations, GEO audits, recommendationsPrompt and tag management, audit triggersStandard plan and up, $189 per month, 2,000 MCP requests
Peec AIMentions, share of voice, competitors, cited sourcesNone documentedAll paid plans, no extra cost
ProfoundVisibility, sentiment, citations, shopping, agent analyticsPrompts, Workflows, Knowledge Base documents, Projects (preview, then confirm); no page writingEnterprise plan only
SemrushTrends and Projects API data, including AI visibilityNone, read-only methodsSemrush One Starter ($199 per month) and up, SEO Classic Pro or Guru, or any Trends API plan
GEOHATSGaps, page briefs, visibility score, site context, full page HTMLSave agent-written pages, bulk-queue builds, track questions, publish to your connected CMS (9 integrations)All tiers, lifetime license from $47 one time

Otterly.AI

Otterly.AI hosts its MCP server at data.otterly.ai/mcp with OAuth sign-in and tools spanning workspaces and engines, brand reports and GEO audits. It goes beyond pure reads: creating brand reports, prompt and tag management plus on-demand audit triggers are available with write permission (Otterly Help Center, 2026). It cannot write or publish content. Every tier is broken down in our Otterly.AI pricing guide. MCP and API access start on the Standard plan at $189 per month with 2,000 MCP requests; each plan tracks 3 of Otterly's 7 engines, and Claude answers cost 3 credits instead of 1 (Otterly Pricing, 2026).

Peec AI

Peec AI runs a Streamable HTTP server at api.peec.ai/mcp with OAuth or personal access tokens, and documents setup for Claude Code, Cursor, VS Code and Windsurf (Peec Docs, MCP Setup, 2026). Peec frames it as a live data layer: brand mentions, share of voice and competitor movement queried from chat, included on all paid plans at no extra cost (Peec AI Blog, 2026). No content creation or publishing tools are documented. We compare the two products in GEOHATS vs Peec AI.

Profound

Profound ships a hosted MCP server at mcp.tryprofound.com/mcp (Streamable HTTP, OAuth 2.1 or an API-key bearer token) plus a local @profoundai/mcp package, and it is available on the Enterprise plan only (Profound Docs, MCP, 2026). It is no longer read-only: besides visibility, sentiment, citation, shopping and agent-analytics reports, its tools can create, update and delete prompts, build and run Workflows, add Knowledge Base documents and manage Projects, with write tools defaulting to preview-then-confirm. Its AI Marketer agents, the part of Profound that drafts pages, aren't available through MCP yet. What Profound costs now is in our Profound pricing breakdown.

Semrush

Semrush extended its API into an official remote MCP server that pulls search, competitive and AI visibility data into Claude, Cursor and other clients. Access comes with Semrush One Starter ($199 per month) and up or SEO Classic Pro and Guru, which include API units; Business and Advanced plans need an API units package, and any Trends API plan also includes it. The exposed Projects methods are read only (Semrush Developer Docs, 2026). It is an SEO-first data source with AI visibility bolted on, not a GEO workflow. If that is what you have today, see Semrush alternatives for AI search visibility.

Trakkr, AthenaHQ, Scrunch and Rankscale

Four more trackers ship MCP servers. Trakkr includes an authenticated server on every paid plan, with 82 tools and slash-command workflows, and it is the closest product to GEOHATS's content side: it writes articles and publishes them to WordPress, Webflow, Shopify and GitHub (see our GEO tools roundup). AthenaHQ's server has write tools that start its content pipeline and import finished articles, though its content API doesn't publish to a CMS. Scrunch's server has 51 tools for metrics, prompts and configuration and no content-writing tools; its brand pricing table lists MCP on Agency Core and Enterprise rather than Core, while its MCP docs say any plan, so check with Scrunch. Rankscale's MCP is in beta: reading is always on, and creating search terms or running credit-spending checks are opt-in permissions you grant when you connect.

GEOHATS

GEOHATS's MCP tools cover the write side of the page loop: saving pages your agent wrote (gated, free, no run cost), bulk server-side builds, growing the tracked question set, and publishing to a connected CMS with explicit confirmation. Every tool is tenant-scoped to your own workspace token. The full tool list and setup steps are on the GEOHATS MCP server page. Pricing is a one-time lifetime license starting at $47 (see GEOHATS pricing), with your own AI keys (BYOK) powering scans and builds. For how it stacks up as a monitoring tool too, see our 2026 GEO tools for SaaS roundup.

The full loop with GEOHATS's MCP server

Disclosure: GEOHATS is our product. This section describes our own MCP server; the vendor comparison above cites each competitor's own documentation so you can verify every claim.

GEOHATS exposes 23 tools over one Streamable HTTP endpoint, authenticated with a per-workspace bearer token you mint in the dashboard (Settings, Coding tools). Tokens are scoped: read (data only), draft (can save pages and track questions), and publish (can push live). A read token cannot spend your AI budget; a draft token cannot touch your live site.

Connect once

Claude Code:

claude mcp add --transport http geohats https://geohats.com/api/mcp \
  --header "Authorization: Bearer gh_mcp_your_token_here"

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "geohats": {
      "url": "https://geohats.com/api/mcp",
      "headers": { "Authorization": "Bearer gh_mcp_your_token_here" }
    }
  }
}

Setup for VS Code and Claude Desktop, and the full list of scopes, is on the MCP server page.

Then run the loop

  1. list_websites and list_gaps. The agent pulls the buyer questions where your site is not cited by ChatGPT, Claude, Perplexity, Gemini, Grok or AI Overviews, plus which competitor domains are cited instead. Gaps come from real scheduled scans, not guesses.
  2. get_prompt_brief. For one gap, the agent gets everything needed to win it: your brand, ICP and design tokens, the recommended page type, and the exact competitor URLs currently being cited, the pages the new page must beat.
  3. The agent writes the page. This is the point of the whole setup. Your agent drafts a complete HTML document in your editor, with your repo open, matching your design system. No LLM call on our side, so this step is free.
  4. create_page_from_html. The agent submits its draft. GEOHATS runs the same automated quality gate as internally generated pages: internal links, cited external sources, JSON-LD, extractable structure, banned AI-tell phrases. Failures come back as a fix list; retries update the draft in place instead of duplicating it.
  5. Ship it, two ways. On a connected CMS (WordPress, Webflow, Shopify, Ghost, HubSpot and others), publish_page (publish scope plus an explicit confirm: true) pushes the gated page live and returns the URL. On your own codebase, skip publishing entirely: get_page returns the final HTML and the agent writes it into your repo as a file, then you commit and deploy like any other change.

Around that core loop sit the scale tools: queue_pages accepts up to 30 gaps and has GEOHATS build them server-side one at a time while you keep working (get_queue reports progress), and generate_questions plus track_questions grow the tracked question set from chat, with a deliberate two-step review so an agent never silently inflates your scan costs. get_visibility_score, list_assets and get_website_context round out the read side. A second set covers technical fixes: scan_site crawls your site the way AI crawlers do, get_site_audit and list_site_fixes return the issues and suggested fixes (titles, meta descriptions, alt text, schema, internal links), and apply_site_fixes writes them back to your CMS, or your agent applies them in the repo and calls mark_fixes_applied. To spot-check one page by hand, the free title tag checker shows its title and description the way Google displays them.

If you want the deeper background on why a tool should write pages instead of only scoring them, we covered the category shift in GEO tools that write the content, not just score it.

A copy-paste agent prompt that works

Paste this into Cursor or Claude Code after connecting the server. It runs one full gap-to-draft cycle and deliberately stops before anything goes live:

Use the geohats MCP server. Call list_websites, then list_gaps for my site.
Pick the gap with the highest buying intent and call get_prompt_brief for it.

Write a complete standalone HTML page from the brief:
- use the site's design tokens from the brief
- open with a 50-word direct answer to the buyer question
- include at least one comparison table and a 6-question FAQ
- cite 8+ real external sources with links; link 4 internal pages
- add Article + FAQPage JSON-LD with dateModified set to today

Save it with create_page_from_html. If the quality gate fails,
fix every listed issue and save again until it passes.
Do NOT publish. I will review the draft first.

Two habits worth keeping from that prompt: always let the gate iterate (a first draft failing on "no external sources" is normal and fixable in one pass), and keep the human review step. You are the publisher of every claim on your site, and no gate replaces reading the page.

Honest limits before you commit

Where this approach (and our tool) falls short today

  • Integrations beyond WordPress are new. GEOHATS's publish_page ships to the CMS connected to that site: WordPress, Webflow, Shopify, Ghost, HubSpot, Framer, Wix, GitHub or a webhook. Only the WordPress path has a long track record; the others were added in October 2026. Squarespace has no publishing API, so there you use get_page and paste the HTML.
  • BYOK means you watch two bills. Scans and server-side page builds run on your own AI provider keys. Costs are shown per scan and per build, and there is no token markup, but it is your OpenAI or Anthropic invoice, not ours.
  • English-only content generation. The quality gate and templates are tuned for English pages today.
  • MCP itself has a trust surface. Any MCP server you connect can feed content to your agent, so prompt injection is a real risk class; Claude Code's own docs tell you to verify you trust each server before connecting (Claude Code Docs, 2026). Use scoped tokens, and give publish scope only to sessions that need it.
  • Citations take time regardless of tooling. Publishing a better page starts a clock; AI engines typically take 2 to 8 weeks to pick up and cite new pages. We walk through that timeline in our playbook on getting cited by ChatGPT.

FAQ

Can I run GEO workflows from Cursor or Claude Code with an MCP server?

Yes. Connect a remote MCP server over Streamable HTTP with one command (Claude Code: claude mcp add --transport http) or one mcp.json entry (Cursor). With Peec, Semrush, Otterly or Profound you can query visibility data from chat, and Otterly and Profound also let an agent change settings such as prompts. With a write-capable server like GEOHATS, your agent can also pull gap briefs, save the pages it writes through a quality gate, and publish, so the entire GEO loop runs inside the editor.

What is the difference between a read-only and a write-capable GEO MCP server?

Read-only servers answer questions: share of voice, citations, sentiment, crawler activity. Write-capable servers take actions: saving a page draft, queueing builds, tracking new questions, publishing. Profound's Enterprise server can now manage prompts, Workflows and Knowledge Base documents but doesn't draft pages; Otterly allows prompt and tag management but not content; GEOHATS exposes the full page write path. Read tools inform you, write tools close the loop.

Do I need a CMS to publish pages from my editor?

No. If your site is a codebase (Next.js, Astro, anything you deploy yourself), the agent calls get_page to pull the final gated HTML and writes it into your repo as a file. You review the diff, commit, and deploy through your normal pipeline. No CMS connection and no publish scope are needed for that route; only publish_page requires a connected CMS.

Is it safe to let a coding agent publish to my live site?

Only with guardrails, which is why scoped tokens exist. In GEOHATS a publish action needs a publish-scoped token, a page that passed the automated quality gate, and an explicit confirm flag from the agent. The safer default is a draft-scoped token: the agent can prepare everything and you press publish after reading the page. Whatever tool you use, treat publish credentials like production deploy keys.

How much does MCP access cost across GEO tools?

It varies widely. Peec includes MCP on all paid plans at no extra cost. Otterly gates MCP behind its $189 per month Standard plan with 2,000 requests per month. Profound offers MCP on its Enterprise plan only, and Semrush includes it from Semrush One Starter at $199 a month. Trakkr includes MCP on every paid plan from $100 a month. GEOHATS includes MCP on every tier of its one-time lifetime license, from $47, with usage running on your own AI keys.

Does an agent really write better GEO pages inside the editor?

For sites that live in a repo, yes, for a structural reason: the agent can read your actual components, styles, existing pages and internal link targets, so the output matches your site instead of a generic template. It also skips the copy-paste round trips that lose context. The brief still matters most, which is why the workflow starts from real scan data (which competitor pages are cited, for which question) rather than from a blank prompt.

How this page was made

Researched with AI assistance, then reviewed, edited and fact-checked by Daniel Borodin. Prices and product details were checked against each vendor's own pages on ; vendors change plans often, so confirm on their site before you buy. Spotted an error? Tell us.

Daniel Borodin · Founder, GEOHATS

Daniel builds GEOHATS and runs AI-search visibility for his own products and client sites. Posts here are researched with AI assistance; he reviews, edits and signs every one.

Free check

Can ChatGPT read your site?

See if GPTBot, ClaudeBot and PerplexityBot are blocked, and whether your pages load without JavaScript. Under a minute, no signup.

Run the free check

Also available as plain Markdown for AI assistants and readers.