Windsor.ai MCP in Claude: add Linkeddit for rank tracking and AI visibility

Windsor describes Windsor MCP as one connector for every data source you connect to Windsor.ai, from Google Ads and GA4 to Shopify and HubSpot, with 350+ sources. Its tools let the assistant list your sources and accounts, look up valid field names for each source, and fetch data for the dates you ask about.

The connector listing for Claude shows five tools, get_connectors, get_accounts, get_fields, get_options and get_data, all of them reads. Windsor says the assistant can make changes on some ad platforms after you confirm each one, but nothing on these pages uses that. Windsor's setup guide says the connector works on Claude's free plan, though free usage limits run out sooner on long analyses.

Windsor documents both. For Claude, open the Connectors menu, search Windsor.ai and click Add. For ChatGPT it lists a plugin, and it also documents Gemini, Perplexity, Copilot Agent and Cursor.

The workflows on these pages are written and tested for Claude and Claude Code only. Clients handle tools differently, so do not assume the prompts carry over unchanged to another client.

Linkeddit supplies measurements Windsor does not collect. Answer Radar asks ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews your buying questions and records who is named, who is recommended and which sources are cited. Search records weekly Google ranks with the AI Overview shown above them. Both come with one plan at $60 a month.

Windsor supplies what happened after: GA4 sessions by channel group, including AI Assistant visits, and Search Console clicks and positions. Linkeddit writes no advice, so the analysis is Claude's, and the two data sets stay separate columns rather than blending into one score.

Two, and both are covered on this page. SEO rank tracking joins Windsor's Search Console positions with Linkeddit's weekly ranks and AI Overview flags, and shows where the two disagree. AI search visibility measurement pairs Linkeddit's measured answers per engine with GA4 traffic from AI assistants, and shows how to verify that a published change moved the answers.

The Linkeddit figures below come from read-only runs on Linkeddit's own test project, plausible.io, on 2026-10-29. The Windsor side is described from its documentation and connector tool list, not run on a live Windsor account.

Setup takes three steps, and Windsor's guide says to connect at least one data source in your Windsor.ai account before asking anything. Add Windsor.ai from the Connectors menu in Claude, connect Search Console and GA4 inside Windsor.ai, then add the Linkeddit connector at https://mcp.linkeddit.com/mcp and create a project with your domain.

Name the Search Console property and the date range in every prompt. Windsor's SEO guide lists this as its first tip for SEO teams, and it also advises cross-referencing Search Console data with GA4 before reporting to stakeholders. Claude can then call get_fields to find the right field names and get_data to fetch them.

Linkeddit needs no field selection. Its tools return named fields such as rank, ai_overview_shown and ai_overview_cites_us, so Claude reads them as they come. Because Linkeddit writes nothing and Windsor.ai is read here, you can run the whole setup without risk to either account.

Windsor's Search Console guide says a third of all Search Console calls on Windsor.ai come through Claude and ChatGPT, more than any other connector on its platform. That is a reason to expect rank questions to be asked in chat, and to give Claude measured data to answer them from.

Windsor's reporting guide says Search Console position should be read per query, weighted by impressions, and that its data takes 2 to 3 days to settle. Linkeddit's rank is the best organic position on one checked results page, from 1 to 10, and null means not in the top 10. Join them by keyword, never by averaging.

Two mistakes recur. One is reading a null rank as missing data, when it is a result. The other is comparing periods of different length. Ask Claude to use equal windows and to say which source each figure came from.

Windsor tested API-based tracking against the logged-in ChatGPT app. In its third round, Pipedream was named in 59% of API answers and 0% of app answers, and the API averaged 8.3 tools per answer against 4.4 in the app. Windsor's conclusion is that an API number and an app number are two different measurements.

Linkeddit states the same limit in its own results: they come from official provider APIs with grounded web search, a controlled proxy and not an exact copy of what each consumer sees. Pairing that with GA4 visits from AI assistants gives a second, independent reading to check it against.

One reading can mislead in either direction. A startup founder on Reddit wrote about checking AI answers: "Tried maybe 20 different phrasings, they're in almost every answer, we appear in like 2." Repeated checks across phrasings are what separate a real gap from a lucky or unlucky single answer.

Windsor's SEO guide for its MCP connects Google Search Console and GA4 to an AI client, and its example prompt asks for the last 7 days against the previous 7: keywords that improved, keywords that dropped most, and keywords between positions 8 and 15 with over 500 impressions and a CTR below 3%.

Windsor also sells a free Search Console keyword rank dashboard template for Data Studio, and it lists SE Ranking as a source for rank-tracker data. Its SE Ranking page says the SE Ranking API is a separate paid plan, so API access must be enabled before that data flows. The rankings Windsor reads are therefore Search Console averages or a rank tracker you already pay for.

To join them with Linkeddit, ask Claude to call get_connectors, get_fields and get_data for your Search Console property over two equal periods, weighting position by impressions per query. Then have it call search_visibility_get with view=ranks, paging until it has every tracked keyword, join the two on keyword, and flag rows where the ranks disagree by more than 3 places, saying which source it trusts for each.

It is an average, not a rank. Windsor's template page explains that Search Console reports the average position of your topmost result across all searches, devices and countries, and its reporting guide says a page's overall average is diluted by long-tail queries it barely competes on.

It also hides rare queries. Windsor's worked example on its own site found that Search Console's named queries covered 58% of impressions, and that brand queries brought 82% of the clicks on named queries, so they were excluded before grouping by position. In that example the 1 to 3 bucket held 12.1% of queries and 36.6% of clicks, at a 5.51% CTR.

None of this tells you what the results page looked like, and it says nothing about an AI Overview. Windsor's Search Console guide also notes that Search Console has no crawl-date field, so a before and after comparison on a page you just edited is only valid once Google has recrawled it.

A measured check of the results page itself. Linkeddit's Search product records, per keyword and week, the project's best organic position from 1 to 10, the change, the ranking URL, whether Google showed an AI Overview, whether it listed your domain as a source, and who holds position 1. The test project tracks 25 keywords, first checked on 2026-10-28.

On that project, cookieless analytics returned a rank of null, so the domain was not in the top 10, yet the AI Overview Google showed listed the domain as a source. The same held for gdpr compliant analytics. For privacy friendly analytics the domain ranked 3 and Google showed no AI Overview. A rank-only report would have called two of those four keywords misses.

Linkeddit's change value is previous rank minus latest rank, and it is null when either check was outside the top 10, so do not read a null change as zero.

Sources can disagree about the same keyword for a legitimate reason. A SaaS marketer on Reddit wrote: "In all possible keywords where our page is ranking number 1 to 3 on Google, we are not showing up on Google AIO, as if we got shadow banned." A Google rank and an AI Overview citation are two readings, so keep both beside each other instead of picking one.

Count the answers, not the guesses. Visibility in AI search is how often an engine names or cites your brand when a buyer asks a real question, and where in the answer it appears. That needs a fixed list of questions, the same engines asked repeatedly, and a record of every sample, including the ones that failed.

In one audit on the test project, dated 2026-10-28 with nine prompts asked of each engine, ChatGPT named the brand in 8 of 9 usable samples with a mean position of 1.375, and Claude named it in 5 of 9. Two engines answering the same nine questions disagreed, and a single blended score would hide that. Google AI Overviews showed an answer in 7 of 10 samples, so three had no Overview and are excluded from presence.

A failed engine is not averaged in. Perplexity returned errors for all 10 samples, so Linkeddit marked it unmeasured with the reason instead of scoring it zero. Ask Claude to repeat the disclosure fields every time: samples asked, samples admissible, and the reason for each exclusion.

The question is already common. A marketer on Reddit asked in 2026: "how do you track whether your website appears in AI-generated answers? Are there any reliable methods or metrics you use?" An agency owner on Reddit wrote that clients had started asking whether their brand shows up on ChatGPT, Perplexity or Google AI Overview, and "Honestly didn’t have a great answer at first." The answer is a method you can repeat and state.

A lot, in Windsor's tests. Using one direct prompt shape in its third round, Zapier was named in 86% of API answers and 70% of app answers. An open prompt made the app answer with no tool named in 57% of cases, against 2% for the direct prompt.

Same-day tool lists from the API and the app had nothing in common in 51 of 103 pairs in round 2, and in 3 of 53 pairs in round 3. So a visibility figure should name its method, and a second reading from another source is worth having.

Visitors. Windsor's guide to its MCP shows its own GA4 data for 8 September to 7 October 2026: AI Assistant visits were 21.1% of sessions and recorded key events on 24.2% of sessions, against 15.7% of sessions and 7.4% for Organic Search. That is Windsor's own site, not a benchmark, but it shows the question to ask of yours. Ask Claude for GA4 sessions and key events by session_default_channel_group for the 28 days before and after a page change.

To check a published change, run the same questions after the page is live and compare with the earlier run. Linkeddit's answer_verify_start does this for a changed page and spends credits, so Claude should propose it and wait for your approval, then read the result with answer_verify_get next to the GA4 window, using equal-length periods. Allow time: Windsor notes that a before and after test on an edited page is only valid once Google has recrawled it.