Exa MCP in Claude: search the web for what Linkeddit shows the AI engines cite

Exa's listing in the Agent Index shows a read-only connector with two tools: web_search_exa, which searches the web and returns page content, and get_code_context_exa, which finds code examples and documentation. Exa's docs add two more that you switch on yourself: web_search_advanced_exa for domain, category and date filters, and agent_run for research that takes more than one search.

That is a search engine for Claude. It sees whatever the open web shows today. It cannot tell you what an answer engine said to a buyer yesterday, which sites those answers cited, or whether a competitor's pricing page changed last week. Those are measurements and stored history, and no web search returns them.

Linkeddit measures what the answer engines say. On its test project, plausible.io, Linkeddit read back the latest citation share for each engine. In Google AI Mode, plausible.io and usefathom.com were each cited in 8 of 66 cited-domain slots (12%). In ChatGPT the top cited domain was matomo.org at 10 of 183 (5%), and plausible.io was cited 8 times. Two engines, Perplexity and one Claude measuring instrument, returned only errors in those runs and are shown as unmeasured, not as zero.

Linkeddit also keeps competitor history. Compete stores dated snapshots of competitor pricing, changelog and landing pages and grades each competitor move. One plan covers it all at $60 a month. Linkeddit does not write the analysis. Your own Claude does that, using this data.

Start from Linkeddit's cited domains, then use Exa to read them. Ask Linkeddit for the domains each engine cited for your category, pick the ones that are neither you nor a competitor (on the test project these included bootstrap.build, itechguides.com and analytics-alternatives.com), and have Claude run web_search_exa for each domain's pages on your topic. The result is a list of pages that already shape the answers, with their text in front of Claude.

People ask for this by name. A news site publisher wrote that cited pages sit at roughly 700 while brand mentions sit at five. Linkeddit supplies the citation counts. Exa's search is how you read those pages and see why they are cited, before you decide whom to contact. Exa's advanced search takes an includeDomains field, which narrows a search to the domains you name.

Use Linkeddit for the stable facts and Exa for the open question. Linkeddit's competitor view lists who ranks in the Google top 10 for your tracked keywords. On the test project, matomo.org ranked in the top 10 for 11 of 59 checked keywords at an average position of 3.73, and ChatGPT cited it 21 times in the latest run. Exa can then search for what the web says about that competitor right now, such as a launch or a review.

Competely's guide to competitor analysis with Claude gives the same caution for any web-search method: verify prices against the live pricing page and ask for a source URL beside every claim. A search connector hands Claude pages, and Claude can still misread them. Linkeddit's page snapshots carry a content hash and a timestamp, so a change is a recorded fact.

Start with web_search_exa. Exa's docs say to keep it for ordinary searches because it gives the model a smaller tool surface, and to enable web_search_advanced_exa only when you need explicit domain filters, date ranges or freshness control. For lists of results checked against criteria, Exa documents agent_run, which needs OAuth or an API key and is usage-based on Exa's side.

To connect, add Exa and Linkeddit in Claude. Exa supports OAuth sign-in and an x-api-key header. If tools do not appear after you edit the URL, Exa's troubleshooting says an explicit tools parameter replaces the default list, so every tool you want must be in it. Neither connector writes to your site, so there is nothing to approve.

Measure share of the answers, not just the number of pages you found. Linkeddit reports presence, recommendation rate, share of voice and citation share per engine. On the test project, Plausible Analytics was present in 85% of admissible Google AI Mode samples (17 of 20) and 100% of Google AI Overviews samples (15 of 15), a count that moves when new pages enter the cited set.

Re-read those numbers after your outreach or new pages have had time to be picked up. A page found with Exa that never changes a citation count was not a lever. Keep that list so the next round starts from evidence.