Claude SEO Prompts: 5 That Work — and Where They Stop

Search “Claude SEO prompts” and you get the same artifact every time: a scroll of fifty prompts, each opening with “Act as an expert SEO consultant” and each promising that the right incantation turns a chat window into an SEO team. Here is what those lists get wrong: the wording is the least important variable in the whole exchange. Claude given no data produces plausible, generic text about your site — the same advice it would give any site. Claude given your actual Search Console rows, your actual page-one competitors, and a decision rule produces work. The prompts worth stealing are not clever phrasings. They are data contracts: what to paste in, what rules to apply, what shape to hand back.

This guide gives you five of those contracts — striking-distance audit, title and meta rewrite, content brief, cluster mapping, competitor teardown — in copy-usable form. And it draws the line the prompt lists never draw: everything below is agent-assisted, not agentic. A prompt returns text. It does not read your data, decide, ship, or verify. Where that distinction starts to matter, this piece hands off to our Claude for SEO pillar, which covers the full stack.

Why most Claude SEO prompt lists fail

A language model is a reasoning engine, not a records department. It does not know your rankings, your Search Console queries, your competitors’ current titles, or last month’s search volumes — and when a prompt implies it should, it will oblige with numbers and findings that look precise and are invented. This is the failure mode behind most published SEO prompts: “list the top keywords for my niche,” “audit my website at [URL],” “what is my domain authority.” Each asks the model to retrieve facts it does not have. The output reads confident because confident text is what the model is good at. That is worse than useless — it is wrong with conviction.

The prompts that earn their output invert the flow of information. You supply the facts — a GSC export, a pasted competitor page, a keyword list with real volumes — and the model supplies what it is actually good at: pattern recognition across hundreds of rows, intent judgment, classification against rules, structured synthesis. Reading a 900-row query export and spotting the twelve queries stuck at position 6 with collapsed click-through is exactly the kind of tedious, judgment-laced work a model does well and a human does badly at scale. The demand for this shortcut is real — the query this article targets is searched about 1,600 times a month at a keyword difficulty of 8 — but the shortcut only works when the data travels with the prompt.

A bad prompt asks Claude for facts it does not have. A good prompt hands Claude your facts and asks for the judgment you do not have time to apply to nine hundred rows.

The anatomy of a prompt that earns its output

Every prompt in this guide has the same four parts, and the pattern transfers to any SEO job you invent later. Data: the pasted export, page, or list that grounds the answer in your reality instead of the model’s averages. Job: one task, stated narrowly — diagnose, rewrite, outline, cluster, tear down — never “improve my SEO.” Decision rules: the opinionated part, where you encode how the call gets made — what counts as striking distance, what disqualifies a title, when two keywords share a page. Without rules the model substitutes the internet’s average opinion, which is how you get output indistinguishable from everyone else’s. Output contract: the exact shape of the answer, including permission to say “insufficient data” — the single line that most reliably cuts hallucination.

Five Claude SEO prompts worth stealing

Each prompt below is usable as pasted. The bracketed placeholders are yours to fill, and the decision rules are deliberately opinionated — change them to match your own bar, but do not delete them.

1. The striking-distance audit

The highest-yield prompt in the set, because it runs on the data you already own. Export your query-level Search Console report and let the model do the triage a consultant would bill a day for: find the queries close enough to page-one prominence that a targeted fix — not new content — moves them.

Below is a Search Console export: query, page, clicks,
impressions, CTR, position (last 90 days).

A striking-distance query ranks 4-15 with 100+ impressions.
Exclude branded queries (anything containing: [brand terms]).

For each candidate, diagnose exactly ONE of:
- snippet: position <= 8 but CTR far below that position's norm
- relevance: position 9-15 — the page only partially matches intent
- cannibalization: 2+ pages earning impressions for the same query

Output a table sorted by impressions: query, page, position, CTR,
diagnosis, a one-line fix, and the single GSC metric that would
prove the fix worked. Never suggest "create more content." If a
row cannot be diagnosed from this data, say so instead of guessing.

[paste export]

2. The title and meta rewrite

Title rewrites fail when they are written from imagination — what we wish the page were about — instead of from the SERP that will judge them. This prompt forces both inputs onto the table before a word gets written.

Rewrite the title and meta description for the page below.

Inputs: the current title + meta, the page's top 5 queries by
impressions [paste], and the page-one titles currently ranking
for the main query [paste].

Rules: intent as page one defines it beats what we wish the page
were about. Title <= 60 characters, the query's noun up front, no
claim the page cannot back. The meta description is a 150-160
character case for choosing THIS result — not a summary.

Output: 3 title options with one line of reasoning each, 1 meta
description, and the query whose CTR should judge the change.

3. The content brief

The banned move here is letting the model guess what page one covers. Paste the actual top results’ structure, and the brief stops being a generic outline and becomes a map of the gap you can win.

Build a content brief for: [keyword]

Inputs: the current page-one results — URL, title, and each
page's H2s [paste them; do not guess] — plus the list of slugs
we have already published [paste].

State the search intent in one sentence, as page one defines it.
Then split the subtopics three ways: table stakes (every top
result covers it), the edge (only one or two cover it), the win
(none cover it and a searcher would still want it). Flag overlap
with our existing slugs and call it: new page, or expansion of
an existing one.

Output: the intent sentence, an H2 outline with a one-line job
per section, the internal links the piece should carry, and the
metric that judges it after it ships.

4. Cluster mapping

Clustering is intent judgment at scale — precisely the work models beat spreadsheets at, and the reason keyword-grouping tools that match on shared words produce pages that cannibalize each other. The full method, including SERP-overlap verification, is in our keyword clustering guide; this is the prompt-sized version.

Cluster the keyword list below into page-level groups.

Rule: two keywords share a page ONLY if one page could plausibly
rank for both — judge by intent, not by shared words ("best crm"
and "crm pricing" are different pages).

Each cluster gets: a primary keyword (highest volume that matches
the intent), a page type (guide / comparison / landing / tool),
and a mapped-or-new call against our existing URLs [paste].

Put anything ambiguous into an "unclustered" bucket rather than
forcing it into a group.

[paste keywords + volumes]

5. The competitor teardown

The most common teardown failure is the conclusion “write better content” — true, useless, and what you get when the prompt permits it. This one does not.

Tear down the page ranking #1 for [keyword] — full text pasted
below, followed by our page.

Answer: what intent does it satisfy in the first screenful; what
does its structure cover that ours does not; what queries is it
probably earning beyond the head term (infer from its H2s and
FAQ); what would a page have to ADD — not copy — to beat it.

Rule: "write better content" is not an answer. Every gap must
name a section, an asset, or a data point.

Output: a ranked gap list with an effort estimate and expected
impact for each item.

[paste competitor page]
[paste our page]

Where the prompt stops: agent-assisted vs agentic

Now the honest part. Run all five prompts well and you have five pieces of sharp text — and a process that is still entirely manual around the edges. You exported the data. You pasted it. You will apply the title change by hand, or not. And three weeks from now, nobody will re-read the CTR on the query the rewrite was supposed to move, because the conversation that contained the reasoning has scrolled into history. That is agent-assisted work: the thinking is delegated, the reading, shipping, and proving are not. Agentic is a different claim — the same reasoning connected to your data on one side and to execution and verification on the other.

Give Claude hands: from pasted exports to MCP

The gap between the two columns is not a better prompt — it is tools. The Model Context Protocol is the open standard that lets Claude call live data sources and execution surfaces instead of waiting for a paste: query Search Console directly, pull live SERP composition, read the task board, report the fix. The moment those tools exist, the five prompts above stop being things you run and start being things the agent runs — the striking-distance audit becomes a scheduled query instead of a monthly export ritual, and the teardown fetches its own competitor pages. How that server is put together is the subject of our guide to MCP servers for SEO, and connecting one to Claude Code is a one-command setup.

The second half of “hands” is execution. Claude Code running inside the website’s own repository does not describe the title fix — it edits the template, opens a pull request, and records exactly what shipped and when. On a site we audited, the striking-distance pattern was textbook: a commercial query parked on page one with hundreds of impressions and almost no clicks. The prompt version of that finding is a row in a table that a human may or may not act on. The agentic version is a task with a committed metric, a shipped PR, and a verdict three weeks later.

Close the loop: ship it, then prove it

Notice that every prompt in this guide ends the same way: name the metric that judges the work. That line is doing more than tidying the output — it is the seam where prompting connects to the discipline that actually compounds. A finding filed with its metric committed in advance, a fix shipped with a timestamp, a re-read of that exact metric about three weeks later, a verdict written down: confirmed, regressed, or inconclusive. Run that cycle and each fix teaches you something about your site; skip it and you are generating well-worded suggestions forever. The full mechanics live in the closed-loop SEO guide — the prompts here are the find step of that loop, run by hand.

Steal the prompts — they are the genuinely useful eighty percent of what the prompt lists promise. Just be honest with yourself about which side of the table you are operating on. If you want to see the find step run against your real Search Console data instead of a pasted export, the free SEO report connects your GSC and surfaces the striking-distance queries, the pages shown but never chosen, and the queries competing with themselves — the same audit, with the data already attached.