ChatGPT Prompts for Smarter SEO Content Briefs

I’ve already made the argument that the brief, not the draft, is where AI content programmes live or die. This post is the practical half of that argument: the prompts I use to build the brief itself. Because there’s a quiet irony in how most teams work - they’ll happily let ChatGPT write the article, then write the brief by hand, badly, in ten rushed minutes. The brief is the better delegation target. It’s structured, it’s checkable, and every input it needs is something you can paste.

The economics favour the upstream work more than most marketers assume. In Semrush’s content marketing research, 38% of marketers not using AI say one long-form article takes two to three hours to write, while 36% of AI users get a draft in under an hour. Drafting, in other words, is already cheap. Meanwhile Optimizely’s survey of 2,003 B2B marketing leaders found 75% spending three or more hours every week fixing AI output. That cleanup tax is mostly a briefing failure being paid off downstream - and these eight prompts are how I keep from incurring it.

One scope note before the prompts: this is the library for briefing a single SEO post well. The four-stage workflow covers where briefing sits in a whole production system, and the evidence you’ll paste into several of these prompts comes out of that workflow’s collection stage.

1. The intent-read prompt

Search intent decides what the page has to be before a word is written, and the reliable way to read it is from the actual results, not from the model’s guess about them.

Keyword: "[keyword]". Here are the titles and meta descriptions
of the current top 10 results: [paste them].

From this evidence only: (1) classify the dominant intent -
informational, commercial investigation, transactional or mixed -
and cite which results prove it; (2) name the page type the
results reward: guide, listicle, comparison, tool, product page;
(3) flag any result that breaks the pattern and what it might
mean; (4) state what a page targeting this keyword must therefore
be, in one sentence.

If the results are mixed, say "mixed" and describe the split -
do not force a single answer.

Why it works: the model classifies text you give it far better than it guesses at search results it can’t see. Two minutes of copying titles turns intent from an opinion into an observation. The one-sentence conclusion in point 4 goes into the brief verbatim as its first constraint - I’ve killed whole planned posts at this step, cheaply, when the SERP made clear the keyword wants a page type we had no business building.

2. The gap-and-table-stakes prompt

This is a cousin of the competitor-gap prompt in my blog post prompts, tuned to produce brief fields instead of an angle.

Here are the H2/H3 headings of the top 3 results for "[keyword]":
[paste headings, labeled by result].

Produce two lists for a content brief:
1. TABLE STAKES: subtopics all three cover. These are mandatory.
   For each, one line on what "covered adequately" means.
2. GAPS: questions a [audience role] searching this would still
   have after reading all three. For each, why the gap matters
   and what evidence we would need to fill it credibly.

Rules: a gap we cannot support with real evidence is listed as
"gap - needs evidence", not silently included.

Why it works: table stakes and gaps map directly onto the brief’s required-coverage and differentiation fields. The “needs evidence” rule matters most - the model will happily propose gaps that sound impressive and can’t be filled, and an honest brief marks those as research tasks instead of promises the draft can’t keep.

3. The secondary-keyword prompt

The brief should name the two or three mid-tail phrases the post can also cover, and this is a clustering job, which models do well - with one hard rule carried over from my ChatGPT keyword research method: the model supplies groupings and intent, never volumes.

Primary keyword: "[keyword]". Here is a list of related queries
from [autocomplete / People Also Ask / my keyword tool export]:
[paste].

Cluster them by what the searcher is trying to do, not by shared
words. Then recommend: (1) 2-3 clusters this post should absorb
as secondary keywords, with the H2 that would cover each;
(2) clusters that deserve their own separate post, and why
covering them here would blur the page's focus.

Do not estimate search volume. If two clusters compete for the
same section, say so.

Why it works: the separate-post list is half the value - it’s your cannibalisation guard and next month’s backlog in one pass. Everything the model outputs here is a hypothesis to check against real data, but the clustering-by-job logic is the part your keyword tool can’t do at all.

4. The reader-situation prompt

A brief that says “audience: marketers” briefs nobody. If you’ve run my buyer persona prompts, this step is a lookup rather than a generation - paste the persona and ask which of its documented pains this keyword serves.

Persona evidence: [paste persona, or 5-10 verbatim customer
quotes about this problem area].
Keyword: "[keyword]" with intent: [from prompt 1].

For the brief, state: (1) the situation the reader is in when
they search this - what just happened, what they're trying to
decide; (2) what they already know, so the post doesn't
re-explain it; (3) what they should be able to DO after reading,
in one concrete sentence; (4) the objection they'll have to our
likely answer, quoted from the evidence if possible.

Mark anything not supported by the evidence as [ASSUMPTION].

Why it works: fields 2 and 3 are the ones writers thank you for - “what they already know” is the anti-filler instruction, and the do-after-reading sentence gives the draft a finish line. The [ASSUMPTION] tag keeps the model from inventing a reader to flatter the brief.

5. The thesis-forcing prompt

The thesis field separates a brief that produces a real post from one that produces a summary of the SERP. I hold the same rule here as in the workflow post: no defensible position, no post.

Topic: "[keyword]". Our evidence and experience: [paste your
data points, observations, results - real ones only].
The consensus answer, from the top results: [2-3 sentences
summarising what everyone already says].

Propose 3 candidate theses for this post. Each must be:
(1) supported by the evidence I gave you - cite which piece;
(2) something a competent competitor could plausibly dispute;
(3) stated in one sentence a practitioner would understand.

For each, add the strongest objection to it. If my evidence
supports no thesis beyond the consensus, say exactly that.

Why it works: requiring the citation and the objection turns “give me a hot take” into an argument audit. And the permission to answer “no thesis” is load-bearing - when the model can’t find a defensible position in your evidence, the honest conclusions are to gather evidence or skip the post, and I’d rather learn that at the brief stage than after 1,500 words exist.

6. The structure prompt

Only now, with intent, gaps, audience and thesis fixed, do I let anything propose headings - structure written before this point is decoration.

Brief so far: [paste outputs of prompts 1-5].

Propose the post's structure: H2s phrased as the reader's actual
questions, ordered so the thesis lands early; under each, 2-3
bullets of what must be said - claims, evidence to include,
the table stakes item or gap it covers. Add: which single H2
carries the thesis, and a "this post will not cover" list built
from the separate-post clusters in the keyword step.

Constraints: max 7 H2s. Every gap from the gap list must map to
a heading or be explicitly dropped with a reason.

Why it works: the mapping constraint is the whole prompt. Briefs rot in the handoff - gaps identified in research quietly vanish from the outline. Forcing every gap to be placed or explicitly dropped means the writer inherits decisions, not research debris.

7. The internal-link prompt

Links planned at the brief stage get written into the copy naturally. Links added after publication read like what they are.

Here are our live URLs with their target keywords and a one-line
description each: [paste - your content inventory or sitemap
notes].
The new post: "[keyword]", covering: [one-line summary].

Recommend: (1) 3-5 of our pages this post should link to, each
with the section it belongs in and a natural anchor phrase;
(2) 2-3 of our pages that should add a link TO this post, with
the sentence on that page where it would fit.

Rule: recommend only from the list. No invented URLs.

Why it works: the “no invented URLs” rule exists because models fabricate plausible internal paths with total confidence. The inbound list in part 2 is the step I see skipped most often - a new post with no inbound links is invisible to crawlers and readers alike, and the brief is the last moment when adding them is a task instead of an afterthought.

8. The brief QA prompt

The last prompt attacks the brief the way the skeptic’s pass attacks an outline.

Here is a finished content brief: [paste].

Audit it as a sceptical senior editor:
1. Could a competitor's writer produce essentially the same
   article from this brief? If yes, the brief has failed - name
   what's generic.
2. Which fields would a writer have to guess at? List the
   questions they'd come ask you.
3. Does the evidence listed actually support the thesis?
4. What would you cut - where does the brief over-specify?

Verdict: READY / NEEDS WORK, with the single most important fix.

Why it works: question 1 is the only test of a brief I fully trust - if the brief’s fingerprint isn’t unique to you, the draft’s won’t be either. Question 4 earns its place too: a brief can fail by over-constraining, and a writer boxed in on every sentence produces exactly the stiff copy the brief was meant to prevent.

Running the chain

In practice the eight prompts take me 25 to 35 minutes for a post that matters, and less once the inputs are systematised - the SERP paste for prompts 1 and 2 is one browser session, the persona for prompt 4 is on file, and the URL inventory for prompt 7 updates monthly. That’s a real cost, and I’d point at the industry numbers to defend it: in CMI’s 2026 B2B research, 97% of B2B marketers now have a content strategy and 95% of their organisations use AI - yet only 58% say AI has improved their content quality. Near-universal adoption, middling quality gains. The differences between teams aren’t in the drafting tools anymore. They’re in what gets fed in.

Two habits make the chain stick. Run prompts 1 and 2 in one session and everything else fresh - the SERP context helps the gap analysis and pollutes the thesis work, where you want the model reasoning from your evidence, not the competitors’. And save every QA verdict: the “questions a writer would ask” lists, accumulated over a quarter, are a precise map of what your brief template is missing.

The full production picture - what happens before these prompts (evidence collection) and after (drafting and the editing pass) - is in the AI content brief workflow. The prompt anatomy all eight are built on is in the prompt library, the keyword decisions upstream of the whole thing are in the ChatGPT keyword research guide, and if SEO is your full-time seat, the AI guide for SEO specialists shows where briefing fits in the week. For the widest view, start with the complete guide to AI marketing.

FAQ

Do I need all eight prompts for every post? No. My full chain is for posts with commercial weight. For a quick informational post, prompts 1, 2 and 6 plus the QA pass get you 80% of the value in ten minutes. The one prompt I never skip is the QA - a thin brief that’s been honestly audited beats a thick one that hasn’t.

Can I combine them into one mega-prompt? You can, and the output will be worse in a specific way: each stage’s output is an input you’re supposed to check before the next stage uses it. One mega-prompt launders the model’s own unverified guesses into “evidence” for its later steps. The chain is slower because the checking is the point.

Does this work with Claude or Gemini instead of ChatGPT? Yes - nothing here depends on the model. The prompts run on pasted evidence and explicit constraints, which is the model-agnostic part of prompt design. Whichever model you use, keep the rule from the keyword step: groupings and intent from the model, numbers from your tools.