Claude SEO prompts that run on your own data

Using rank_check, look up [domain] for [keyword] in sg, us, uk and au. Return one row per market with the position and the ranking URL. A position of 0 means not in the top 100, so write "not in top 100" instead of the number. Then say which market is strongest and whether one URL ranks in every market or different pages do.

Run ai_presence for [domain] on each of these keywords in [country]: [up to 10 keywords]. Build a table with three columns: keyword, whether an AI Overview or AI Mode answer showed, whether [domain] was cited in it. Finish with the keywords where an answer shows and we are absent, in the order I listed them. Report what the tool returned and nothing more; do not guess at why.

Call page_meta on these URLs: [my URL], [competitor URL 1], [competitor URL 2]. Lay out title, meta description, H1 and the H2 headings side by side. Tell me which subtopics their headings cover that mine do not, and whether my title contains [keyword] at all. Only if my title or meta fails on that count, draft one replacement title under 60 characters and one meta description under 155.

For [URL] and [keyword] in [country]: call page_meta on the URL and rank_check on the domain. Report the position and whether the URL that rank_check returns is the same page. If a different page on my domain ranks, name it and say which of the two the title and H1 suggest should carry the keyword. Do not recommend merging pages; that is my call.

Call list_articles with limit 50. Group the articles by their seed keyword and keywords into topics. List any two articles that share a keyword. Then suggest five keywords that sit next to these topics but have no article yet, as candidates only. Do not call generate_article.

Write one article for the [project name] project on [keyword]. Before calling generate_article, show me the brief you will pass: the angle, the reader, three points it must cover, under 4000 characters. Wait for my yes. After the call, poll get_article_status about every two minutes and tell me the status each time: queued with the queue position, drafting, revising, then done with the article_id and score, or failed with the error. Stop polling after done or failed.

Call get_latest_audit for [project name]. From the recommendations, keep the ones whose fix is a missing page or thin content rather than a tag or a header change. For the most severe of those, propose a keyword and a brief, and quote the recommendation's issue and fix text next to it. Wait for my approval, then call generate_article with that brief and report the job_id.

Call get_latest_audit for [project name]. Give me: the score, grade and verdict as returned; the totals (pages, checks, pass, warn, fail); each category with its score and grade, worst first; and the recommendations grouped by severity, with the url where one is present. Then name the three fixes to ship first, using only those recommendations. If audit_running is true, say a newer audit is in progress and this may be superseded.

Call get_latest_audit for [project name] and list every entry in unassessed with its reason. For each, say whether the reason text points at something on my side (a blocked crawler, a missing sitemap, a page that timed out) or at a limit of the audit itself. Do not score or guess at what was not assessed.

Call get_latest_audit for [project name] and pull every recommendation that mentions a title, a meta description or an H1 and carries a url. For each url, call page_meta now and compare what is live against the issue. Mark each one fixed, unchanged, or changed but still failing, and quote the live value. Remind me at the end that this does not rerun the audit, so the score is unchanged until the next run.

Call rank_history for [project name]. For each market (each location_code) list the keywords with current, best and first position and the last_checked date. A position of 0 means not in the top 100. Skip keywords marked pending and tell me how many there were. Sort by the gap between current and best, largest gap first. Then list the five keywords sitting at their best position ever. Write it so I can paste it into a Monday message.

Call rank_history for [project name]. Take the keywords where current is worse than best by 3 or more positions. For up to ten of them, call ai_presence for [domain] in [country]. Table: keyword, best, current, AI answer shown, cited. Say which slips coincide with an AI answer we are not cited in. Report this as a coincidence in the data, not a cause.

Call list_projects. Table: name, domain, country, latest_grade. Put the projects whose latest_grade is null in their own list, since those have no finished audit. Sort the rest worst grade first. Do not call get_latest_audit on each of them yet; ask me which one to open.

Call list_integrations for [project name]. Show each connection with provider, status, last_error and auto_publish_status. If there is exactly one, use it; otherwise ask me which. Then call publish_article for article [article_id] with status draft. Report sent, requested, state and external_url exactly as returned. If state is unverified, say that a draft has no public URL to verify and that this is expected.

The publish I just ran came back with state "checking" and external_url [URL]. Call page_meta on that URL. If the title matches the article, tell me the page is reachable and the product's verifier should mark it published on its next pass. If page_meta returns an error or a different title, tell me that and do not call publish_article again; a second publish creates a second post, and the verifier decides the state, not us.

Call list_projects. For each project with a latest_grade, call get_latest_audit and rank_history. Write one paragraph per client: the grade and verdict, the worst category by score, how many keywords sit at their best position and how many have slipped from it. Use no adjective the tools did not return. End with the one project that needs attention first and the two values that say so.