Most SEO prompt listicles are strings nobody ran. Every prompt below was run in a sandbox
against its target model before publishing, and each one passed review by three independent
editorial AI agents. The run evidence lives on each prompt's entry in our
Marketing & SEO prompt directory — this page is the tour;
the directory is the source.
Who this is for: anyone doing hands-on SEO work — titles, briefs, keyword lists,
decaying pages — who wants prompts that force evidence and refuse to hallucinate rankings.
How to use them: paste the prompt, replace the {{placeholders}} with your real
material (the page copy, the SERP, the keyword list). These prompts deliberately demand
inputs — a prompt that works without your data is describing someone else's site.
A few are marked heavier model recommended: they involve multi-document judgment and were
run against a stronger model in our sandbox.
1. Rewrite a title tag against the live SERP
Most title-tag advice ignores the only context that matters: the titles you're actually competing against. This prompt makes the model read the live SERP first, find the promise nobody else makes, and refuse claims the page can't back.
Rewrite this page's title tag to compete in the SERP shown. First state what the competing titles collectively promise and what none of them offers that this page genuinely delivers. Then write three candidate titles that lead with that gap, describe the page honestly, and are short enough to display without truncation in search results. For each candidate, say which searcher it wins and which it loses. Do not promise anything the page content cannot support, and do not predict rankings or click-through changes. Finish by recommending one candidate and why. Page summary: {{page}} Target query: {{query}} Competing titles: {{serp}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
2. Meta description that matches the page
Descriptions that overpromise get clicks that bounce. This one generates three candidates in different styles and flags any claim it can't point to in your copy.
Write a meta description for this page. It must promise only what the page copy shown actually delivers, name the concrete thing the searcher gets, and read as a sentence a person would click — not a keyword sequence. Keep it short enough that search results will not cut it off mid-thought. Write three candidates: one benefit-led, one specificity-led (numbers, named items from the page), one question-answering. Flag any candidate that makes a claim you could not point to in the copy, then recommend one. Page copy: {{copy}} Target query: {{query}}
See this prompt's run evidence →
3. Rewrite an intro to answer the query first
If your page answers the query in paragraph four, searchers never get there. This rewrite moves the answer to sentence one — and tells you honestly if the page never answers it at all.
Rewrite this page's introduction so the searcher's question is answered in the first two sentences, using only facts already present in the page. The direct answer comes first; context and nuance follow it, not the other way round. Keep every fact traceable to the copy provided — if the page never actually answers the query, say that plainly instead of writing an answer for it. Preserve the page's voice; do not add marketing language that is not there. Output the rewritten intro, then one sentence on what you moved and why. Current intro and page facts: {{page}} Query: {{query}}
See this prompt's run evidence →
4. Diagnose low CTR on a page that already ranks
Ranking without clicks is a diagnosis problem, not a rewriting problem. This prompt forces a commitment to one cause with SERP evidence before proposing any fix.
This page ranks but is not clicked. Using the search result it currently shows (title, description, URL) and the competing results around it, diagnose why: title promises the wrong thing for the query, description repeats the title instead of adding a reason, a competitor owns the click with something concrete (a number, a current year, a named brand), or the query's clicks are going to a SERP feature rather than any result. Commit to the most likely cause and say what evidence in the SERP supports it over the alternatives. Only then propose the fix for that cause. Do not predict a CTR number. Our result: {{result}} Query and position: {{query}} Surrounding SERP: {{serp}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
5. Rewrite a section to be quotable on its own
Snippet-worthy sections answer the query completely in a passage that survives being quoted alone. This rewrite produces that passage from facts already on your page.
Rewrite this section so its opening directly and completely answers the query in a form that makes sense quoted alone, without the rest of the page. Definition queries get the definition first; how-much queries get the number and its range conditions; how-to queries get the steps as a tight list. Everything must come from the section's existing facts — no new claims. After the rewrite, show the exact passage a search engine would most plausibly extract. Do not promise the snippet will be won; ranking on page one is a precondition you either have or do not. Section: {{section}} Query: {{query}}
See this prompt's run evidence →
6. Classify a keyword list by real intent
Intent misreads produce pages nobody wanted. This classifier names which word in each keyword carries the signal, admits ambiguity instead of forcing calls, and groups keywords into page-sized sets.
Classify each keyword by the searcher's actual need: informational (wants to understand), commercial (comparing before a purchase), transactional (ready to act), or navigational (wants a specific site). Judge from the words themselves and say which word carries the signal — modifiers like best, vs, price, near me, how, why. Where a keyword is genuinely ambiguous, label it mixed and state both readings rather than forcing one. Then group the list into page-sized sets: keywords one page can satisfy together, keywords that need separate pages, and navigational terms to leave alone. Keywords: {{keywords}}
See this prompt's run evidence →
7. Cluster a keyword list into pages
The follow-up: turn a raw keyword list into a page plan, with a stated reason for every grouping and an explicit skip list for keywords that don't deserve a page.
Cluster these keywords into pages: each cluster is a set one page can rank for together because a searcher typing any of them wants the same thing. Name each cluster's page (working title and the primary keyword), list its member keywords, and state the shared need that justifies the grouping. Keywords that look similar but want different things go in different clusters — say which word told you. Mark keywords too weak or too far from the site's topic to deserve a page as skip, with the reason. Do not invent search volumes; if volumes are provided, use them to order clusters, not to create them. Keywords: {{keywords}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
8. Build a content brief from a keyword and its SERP
A brief a writer can actually execute: committed intent, an angle grounded in what the SERP lacks, and the evidence to gather before drafting — with no arbitrary word count.
Build a writer-ready brief for this keyword from the SERP evidence provided. Commit to: the dominant intent and the page type that serves it; the angle this page takes that the ranking pages do not, grounded in what the SERP actually lacks; the sub-questions the page must answer, ordered by how the searcher's need unfolds; and the evidence the writer must gather before drafting — the specific tests, numbers, or examples that make the page citable rather than another summary. Name what the brief deliberately excludes and why. Do not specify a word count; the coverage list is the length. Keyword: {{keyword}} SERP: {{serp}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
9. Outline a page from the SERP's dominant intent
Reads the SERP, states the dominant intent with proof, then outlines in searcher order: committed answer early, edge cases last. Secondary intents get marked as same-page or new-page, never silently bloated in.
From this SERP evidence, first state the dominant intent in one sentence and name which results prove it. Then outline a page that serves that intent: headings in searcher order — the committed answer early, supporting detail after, edge cases last — with one line per section stating what it must establish and what evidence it needs. Where the SERP shows a secondary intent worth serving on the same page, mark that section as serving it; where the secondary intent needs its own page, say so instead of bloating this one. Keyword: {{keyword}} SERP: {{serp}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
10. Find the honest gaps against a competitor page
Gap analyses that only flatter the competitor aren't trusted. This one cuts both ways — what they evidence that you assert, and what they do worse — then picks the three additions that change what a reader gets.
Compare our page against the competitor page that outranks it. List what they cover that we do not, what they evidence that we merely assert (their tests, numbers, screenshots versus our claims), and what they do worse — thin sections, stale facts, padding — because the gap analysis must cut both ways to be trusted. Then name the three additions to our page that would most change what a reader gets from it, each tied to a specific gap you found. Do not recommend copying their structure; recommend closing evidence gaps. Work only from the two pages as provided. Our page: {{ours}} Theirs: {{theirs}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
11. Refresh brief for a page losing traffic
Fading pages need surgical refreshes, not rewrites. This brief separates stale facts from surviving sections — and says plainly when the topic itself died and no refresh will recover it.
This page used to earn traffic and is fading. Write the refresh brief: which facts have gone stale and what current information replaces them; which sections the queries it still gets impressions for actually want, versus sections nobody asked for; what the page must add to beat whoever took its position — grounded in the competitor evidence provided, not guessed. Keep the URL and everything that still works; list what survives untouched so the refresh does not become a rewrite. If the evidence says the topic itself died, say that instead — a perfect refresh of a dead query recovers nothing. Page, stats, and competitor evidence: {{evidence}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
12. Keep, refresh, consolidate, or remove a page
Keep, refresh, consolidate, or remove — verdict first, evidence second, and what would change the call. Removal requires salvage-value reasoning, not just low traffic.
Give a verdict on this page: keep as is, refresh, consolidate into another page, or remove. Use the evidence provided — its traffic history, whether the topic still gets searched, whether the content is accurate today, and what else on the site overlaps it. State the verdict first, then the evidence that drove it, then what would change your mind. For refresh, name the specific sections that are stale. For consolidate, name the receiving page and the redirect. For remove, confirm nothing links to it that would break, and note removal is for pages with no salvageable value — not just low traffic. Page and evidence: {{evidence}}
See this prompt's run evidence →
13. Assign one owner page per query
When your own pages compete, someone has to own each query. This triage assigns ownership on performance evidence — and notes where apparent cannibalization is actually fine.
These pages from one site get impressions for overlapping queries. For each query group, name the page that should own it and the evidence — which page already performs better, which page's content actually satisfies the query, not which was published first. Then per page: what it keeps, what it cedes, and the change that enacts the decision — retitle away from the ceded query, add the section that justifies ownership, or consolidate and redirect if a page ends up owning nothing. Note where apparent cannibalization is actually fine because the queries differ more than their wording suggests. Pages with their query data: {{data}}
Heavier model recommended — this one was sandbox-run on a stronger model.
See this prompt's run evidence →
14. De-stuff copy without losing the topic
Rescues copy written for a crawler. Every fact survives, the mechanical repetition doesn't, and the worst offender's repetition count is listed so the change is auditable.
This copy was written for a search engine and reads like it. Rewrite it for a person: keep the topic's natural vocabulary — the words a knowledgeable writer would use anyway — and remove the mechanical repetition, the variant-phrase stacking, and the sentences that exist only to hold a keyword. Information content must survive: every fact in the original appears in the rewrite. Show the rewritten copy, then list the phrases you cut and the repetition count for the worst offender so the change is auditable. Copy: {{copy}}
See this prompt's run evidence →
15. Place internal links that read like prose
Internal links that read as inserted get skimmed past. This pass finds the sentences where a reader genuinely wants the target page — and says so when there's no honest placement.
Given a target page and the copy of a source page, find the places in the source copy where a link to the target genuinely helps the reader — a moment where they would want the thing the target page has. For each placement: quote the sentence, show it rewritten with the link, and use anchor text that says what the target delivers in natural words — no bare 'click here', no keyword-exact anchors that read as inserted. If the source copy has no honest place for the link, say so; a forced link is worse than none. Target page: {{target}} Source copy: {{source}}
See this prompt's run evidence →
FAQ
Are these prompts free to use?
Yes. Copy them, adapt them, ship them. Attribution is appreciated, never required.
What does "actually run" mean?
Before publishing, each prompt was executed in a sandbox against its target model with a
realistic input fixture, and the output was checked by three independent editorial AI agents
for accuracy, safety, and usefulness. Prompts that failed did not get published.
Which model should I use?
Any current frontier model handles most of these. The ones marked heavier model recommended
involve multi-document judgment. Not sure which model fits your work?
Our which-LLM picker takes two minutes.
Why do the prompts demand so much input?
Because SEO advice without your data is astrology. The placeholders force the context that
makes the output specific to your site instead of generic best practices.
All 30 Marketing & SEO prompts, with run evidence and review verdicts, live in the
PromptZone prompt directory. New categories ship every few weeks —
each one gated on the same evidence standard.
Top comments (0)