The AI Recommendation Index for marketing MCP servers

This is the first AI Recommendation Index: what five AI engines say when a buyer asks them which MCP server to use for a marketing job. We asked ChatGPT, Claude, Gemini, Perplexity and Google's AI Overviews 33 questions across ten categories on 24 October 2026, and counted which brands each answer named and which sites it cited. The 165 answers, 1,614 citations and every count below are in a public GitHub repo, so you can check our work or count differently. We run it again every month with the same questions.

Across all ten categories, Ahrefs was named in 31 of the 165 answers, Semrush in 29, Composio in 23 and DataForSEO in 20. That says more about how often SEO comes up in marketing questions than about any one category: each category below has its own leader, and in five of them no brand was named in more than half the answers.

The index by category

Each table counts the answers in that category that named a brand ("Named in") and the answers that cited a page on its site ("Cited in"). Five engines answered every prompt, so a category with four prompts has 20 answers. Ties with the sixth row are kept, and our own row is shown even where we were not named.

SEO data

Ahrefs, DataForSEO and Semrush were named in 13 or 14 of the 20 answers, on all five engines. No other brand reached half. The one community server near the top is mcp-gsc, a Search Console server, named in 6.

Search Console

The most named Search Console server is a community project on GitHub, AminForou/mcp-gsc, named in 8 of 20 answers. Google does not publish a Search Console MCP server, as ChatGPT pointed out in two answers. Perplexity's top pick for "best Google Search Console MCP server" was a server from Get MCP Ads.

Google Analytics

Where the platform has its own server, the engines pick it. Google's official Google Analytics MCP server was named in 10 of 15 answers, and first in 7 of them. Windsor.ai was named in 7 and cited in 9, more than any other brand here, because the engines read its GA4 setup guides.

AI visibility

Peec AI, Semrush, Otterly.AI, Ahrefs and Profound were each named in 7 or 8 of 20 answers, on all five engines. Two of the four prompts do not say MCP, and their answers list AI search trackers, so the leaders are the tools that sell that tracking.

Reddit

Reddit has no leader. Three brands tie at 5 of 20: eliasbiondo/reddit-mcp-server and Hawstein/mcp-server-reddit, both community servers on GitHub, and Brand24, a listening tool, named in all five answers to the monitoring prompt.

LinkedIn

Composio was named in 8 of 20 answers and stickerdaniel/linkedin-mcp-server, a community server, in 7. For "LinkedIn MCP without login", ChatGPT and Perplexity named none of the brands we track, and the other three engines one or two each.

Social and video

The longest tail of the ten. No brand was named in more than 3 of 15 answers. The five answers to "best YouTube MCP server" named five different GitHub servers between them, and only Claude and Gemini agreed.

Leads

Clay was named in 8 of 15 answers, then ZoomInfo, Apollo.io and HubSpot in 7. We did not count Apollo in the answers to "Apollo MCP alternatives", because the prompt names it. Claude and Gemini also read that prompt as Apollo GraphQL's MCP server, a different product.

Paid ads

Pipeboard was named in 6 of 10 answers: all five Meta Ads answers and Claude's Google Ads answer. But the platforms' own servers came first where they were named. Four of the five Meta Ads answers put Meta's own Ads MCP server first, and three of the five Google Ads answers put Google's official server first.

Marketing in general

The two broad prompts drew the suites: Ahrefs in 8 of 10 answers, Semrush in 7, HubSpot and Zapier in 6.

The engines disagree

The same brand can lead on one engine and be missing on another. Claude named the most brands from our list per answer, 4.9 on average; AI Overviews named the fewest, 2.1. mcp-gsc was named in 5 Gemini answers and no AI Overview.

GitHub is the most cited site

GitHub was cited 296 times, 18% of the 1,614 citations, and in 76 of the 165 answers. Perplexity cited it in 26 of its 33 answers and ChatGPT in 21; Claude in 8. The next most cited sites were OpenAI's help centre (52, all in answers to prompts that name ChatGPT), Reddit (50) and YouTube (34). Directories of MCP servers come next: mcpservers.org was cited in 19 answers, glama.ai in 12 and mcpmarket.com in 10.

For an MCP server, that makes a GitHub README and a directory listing worth as much as a landing page. How to get cited by ChatGPT covers what the engines cite for buyer questions in general.

Where Manifold stands

We make an MCP server in most of these categories, so here is our row. Manifold was named in 5 of the 165 answers, 4 of them Claude's and one ChatGPT's (for Reddit), and in none for Search Console, Google Analytics, LinkedIn, social, leads or ads. Pages on manifoldmcp.com were cited in 12 answers, more than we were named: the engines read our roundups, such as the best SEO MCP servers, and then recommend the servers in them. In a separate check earlier the same day, ChatGPT named Manifold third for "what is the best MCP server for SEO data"; in this run it did not name us at all, which is why one run is a snapshot, not a ranking.

Method

• Prompts. 33 questions a buyer would type, in ten categories, written from the searches people make about MCP servers. They are listed in the repo and stay the same each month; any change is noted here.
• Engines. ChatGPT (gpt-6) and Gemini (3.5 Flash-Lite) as their apps answer a person, Claude (claude-sonnet-5-5 with web search) and Perplexity (sonar-reasoning-pro) through their APIs, and Google's AI Overviews. Asked once each, from the United States, in English, on 24 October 2026, through Manifold's AI visibility MCP server. All 165 answers came back.
• Named. A brand counts as named when the answer's text names it, with links and images taken out so a link's address is not a mention. We match a list of about 100 brands, each with its own pattern, and skip a brand the prompt itself names. A platform's own name (Google, Reddit, LinkedIn) is not counted, because every answer about it names it; its own MCP server is, where the answer says it is the official one. Brands outside the list are not counted, and the full answers are in the data for anyone who wants to.
• Cited. A brand counts as cited when the answer links a page on its site or its GitHub repo.
• Limits. AI answers change from one day to the next, and one run per prompt and engine is a sample of one. Read a gap of one or two answers as a tie.

Get the data

The repo at github.com/manifoldmcp/ai-recommendation-index holds one folder per run (this one is data/2026-10-24), with three CSV files and the prompts and brand list it was counted with:

• answers.csv: one row per prompt and engine, with the full answer, the brands named in order and the brands cited.
• mentions.csv: one row per brand in each answer, with its place and whether it was cited.
• citations.csv: every cited URL, with its title and the brand it belongs to.

The data is free to use under CC BY 4.0. Please cite it as "Manifold AI Recommendation Index, October 2026" and link this page. If your MCP server is missing from the brand list, open an issue on the repo and it will be counted from the next run.

Run the same check for your category

The index uses the same tool any Manifold user has. Give your agent your own buyer questions and brands, and it can return the same tables for your market.