Tavily MCP in Claude: crawl a competitor's site, keep the history in Linkeddit

Tavily's listing in the Agent Index shows five tools: tavily_search for a web query, tavily_extract to pull the content of given URLs, tavily_map to list the URLs of a site, tavily_crawl to walk a site and return its pages, and tavily_research for a longer research task. The listing describes the connector as read-only, with PII redaction and prompt injection protection built in on Tavily's side.

Together they answer one question well: what does this site say today? A map shows how a competitor organises its pages, a crawl returns the text, and an extract reads a single page you already know. None of them remembers last month, so none can say that a plan was renamed or a page was removed.

Linkeddit stores history and measurements. Compete keeps dated snapshots of competitor pricing, changelog and landing pages with a content hash, so a changed hash between two snapshots of one URL marks a change. It also grades competitor moves from blogs and monitored pages. On the test project, plausible.io, a read-only pull returned five signals for the week of 2026-10-26, all low priority: three from Simple Analytics and two from Umami (for example a Simple Analytics post on a German court ruling against Meta's tracking technology, and an Umami preview of its v3 release).

A founder described the manual alternative: many companies "still rely on spreadsheets, saved tabs, screenshots, and occasional alerts for competitor tracking." A crawl hands Claude a fresh copy each time you ask. Linkeddit is the record Claude compares that copy against. One plan covers it at $60 a month, and Linkeddit writes no analysis of its own.

Start with who matters, then look wide. Linkeddit's competitor footprint lists the domains that rank beside you. On the test project, matomo.org ranked in the Google top 10 for 11 of 59 checked keywords at an average position of 3.73, and was cited 21 times by ChatGPT in the latest run. Ask Claude to run tavily_map on that domain to see which sections it has, and tavily_extract on the pages that matter.

Then compare to the record. Ask Linkeddit for the stored page snapshots of the same competitor's pricing and changelog, and have Claude say what the live extract shows that the last snapshot did not. Competely's guide to competitor analysis with Claude gives the same advice for any web method: verify prices on the live pricing page and put a source URL beside every claim.

Use tavily_research for an open question that needs several searches, such as how a category's vendors describe a feature. Use tavily_search for one fact and tavily_extract when you already hold the URL. For a competitor you track, skip research and ask Linkeddit, because the stored change history is cheaper and dated.

Tavily's docs describe a DEFAULT_PARAMETERS header on the remote server for setting defaults on every request, and a session id it attaches to each API call. Both are Tavily settings. They shape what comes back from the web and have no effect on Linkeddit's data.

Linkeddit's results view lists the domains each engine cited. On the test project, ChatGPT's top cited domains included matomo.org (10 of 183) and itechguides.com (9 of 183), and Google AI Overviews cited analytics-alternatives.com in 6 of 107. Linkeddit shows who is cited. Tavily is how Claude reads those pages in one pass with tavily_extract, so you can see what they say that yours does not.

Keep the order. Measure first, then read. A crawl of ten sites tells you nothing about the engines until you know which of those sites they cite. In the same runs, Perplexity and one Claude measuring instrument returned only errors, so those two are unmeasured, and Linkeddit reports them that way.

Check the date and the page. A crawl reflects the moment of the request, and a site can serve different content by region or to a bot. Ask Claude for the URL beside each claim and compare prices with the live page yourself. If Linkeddit's snapshot and the live extract disagree, the difference is the finding.

Then measure the effect of whatever you change. Linkeddit reports presence, recommendation rate, share of voice and citation share per engine. On the test project, Plausible Analytics was present in 17 of 20 Google AI Mode samples. Re-read those numbers after you act, since a crawl cannot show them.