PostHog MCP with Claude, plus the AI visibility data PostHog cannot see

PostHog describes its server as a free, hosted endpoint that lets an agent use PostHog through plain text prompts. Its own examples include shipping a feature flag, reviewing stack traces, running HogQL queries, triaging support tickets and setting up CDP destinations. It works with any MCP compatible client, including Claude Code and Claude Desktop.

The Agent Index listing for the PostHog connector shows 57 tools: 33 reads, 20 writes, 1 admin and 3 it does not classify. Reads include projects-get, read-data-schema, query-run, execute-sql, insight-query, insights-get-all, dashboards-get-all, list-errors and logs-query. Writes include create-feature-flag, dashboard-create, experiment-create and survey-create. Connecting and calling tools is free, but PostHog notes that some tools use language models internally and may be billed as PostHog AI spend.

PostHog records what happens on your site and in your product. It does not record where you rank for a buying keyword, whether Google's AI Overview names you, or which brands the AI engines recommend when a buyer asks. A drop in visits can come from any of those, and a PostHog insight alone gives Claude no way to tell which.

An ecommerce SEO on Reddit described the pairing that works: "Right now I'm checking a fixed list of buying questions every month, noting whether an AI Overview appears and who it cites, and pairing that with branded search and GA4 referrals from AI platforms." Linkeddit does the first half of that routine on its own clock. PostHog can hold the second half, if your events record the referrer.

PostHog is the behavior side: events, insights, HogQL and the dashboards your team already trusts. Linkeddit is the visibility side, at one plan of $60 a month. Its search_visibility_get tool returns each tracked keyword's best Google position and whether an AI Overview listed your domain, and its answer tools record which brands the five engines named. Claude reads both and reports where they agree or disagree.

Keep the roles separate when you prompt. The PostHog server can create flags, dashboards and experiments, so a prompt that only needs a report should tell Claude to read and change nothing. An independent comparison from Scalekit, a company that sells a PostHog connector layer, notes the server offers read only mode, tool filtering and pinning to one organization or project. Use them.

For interactive questions, the MCP server is enough. Scalekit's comparison says it suits analytics style agents: running HogQL, reading insights and dashboards, managing flags and experiments, and triaging errors. It also says event capture is API only, bulk or scheduled export is limited on MCP, and tool schemas change as PostHog updates the server.

So Claude can read your numbers and write you a report, but it will not ingest events or run a pinned nightly pipeline. A Linkeddit plus PostHog report is the interactive kind: you ask, Claude reads both, and you get the answer. Scalekit is a vendor in this space, so read its advice as a view, not a verdict.

On Linkeddit's test project, plausible.io, search_visibility_get returned 25 tracked keywords, 12 of them checked on 2026-10-28. Five of those 12 ranked in the top 10: google analytics alternative at position 2, matomo alternative at 2, privacy friendly analytics at 3, simple analytics alternative at 3 and plausible vs fathom at 5. The other seven were not in the top 10.

Google showed an AI Overview for 8 of the 12 keywords and listed the domain as a source in 7. Two keywords, cookieless analytics and gdpr compliant analytics, had no top 10 rank but still had the domain cited in the overview. A rank report alone would have missed that. The run was read only, and the PostHog side was not run. Every row was a first check, so Claude should call it a snapshot and repeat it to show change.

Linkeddit's answer_results_get returns, per engine, how often your brand was named and recommended, each vendor's share of voice, and which domains the answers cited. For plausible.io, Google AI Overviews named the brand in 15 of 15 usable samples and recommended it in 12, while ChatGPT named it in 16 of 18 and Gemini in 16 of 20. Plausible held the largest share of vendor mentions in Google AI Overviews at 15 of 91, about 16 percent.

Two segments came back unmeasured: Perplexity returned errors for all 21 samples, and one of the two Claude measurement setups returned errors for all 10. A good report names those gaps. Claude can then set the picture beside a PostHog insight for the same weeks.

Treat events and answers as two measurements of one change. After you publish a fix, Claude can ask Linkeddit's answer_verify_start to re-check the buying questions the change targeted, then read the result with answer_verify_get. Starting a check spends credits, so Linkeddit returns an estimate first and Claude waits for your yes.

Then Claude runs the matching PostHog query for the same period. Linkeddit supplies the answer side only; the conclusion is the customer's to draw.