You open ChatGPT, ask for twenty blog ideas, and get twenty titles that all sound the same. Sound familiar? That is normal, and it is not your fault. The problem is not the model. It is that you asked it to generate content ideas with AI before you told it what "complete" means for your cluster.
This guide fixes that. By the end you will have a repeatable way to find subtopics with AI across your whole cluster, then filter that pile down to a clean, non-overlapping spoke list you can hand straight to a writer.
Eight steps. Take them in order, and take them one at a time.
Step 1: Write the cluster brief before you write the prompt
A model can only be as complete as the boundaries you give it. So before you generate content ideas with AI, start with a short brief, not a prompt.
What to do. Write down the pillar topic and where it stops. Name the audience and the personas inside it. Note the product context, the buyer stages that matter, and any industry or technical limits. Then write the part most people skip: what is explicitly out of scope. List your existing pillar and supporting pages, and the competitors buyers weigh against you.
Last, set a completeness standard. Say out loud which coverage dimensions a finished cluster needs: definitions, how-to questions, use cases, comparisons, alternatives, cost and effort, implementation, risks, objections, troubleshooting, measurement, and decision criteria. Some will not apply. Ask the model to say which ones and why.
Done means. Someone else on your team could read the brief and describe what the cluster covers, who it serves, which stages matter, and what a valid spoke has to contain. It also says what should never become its own article.
Where people go wrong. Opening with "give me blog ideas about X." That produces a brainstorm, not a coverage model. You get generic listicles and keyword variants, and you miss the objections, the implementation questions, and everything at the decision stage.
Set these fields now, so the output arrives usable.
| Field | Why it earns its column |
|---|---|
| Candidate title | Makes the page legible to an editor |
| Primary buyer question | States the page's single job |
| Supporting questions | Shows what the page has to answer |
| Intent | Informational, instructional, comparative, evaluative, or transactional |
| Funnel stage | Awareness, consideration, or decision |
| Persona | Stops one generic audience from swallowing distinct needs |
| Evidence needed | Names the proof, data, or steps required |
| Existing page match | Flags a consolidation or a link target |
| AEO opportunity | Records the visibility reason |
| Overlap note | Explains why this is separate from its neighbors |
| Recommended format | How-to, comparison, glossary, or checklist |
Step 2: Give AI bounded context and ask for a structured inventory
Now you prompt. The content ideation AI prompts that work are built in labeled sections, so the model can tell your instructions apart from your source material.
What to do. Build the prompt in six blocks.
- Role. A content strategist hunting coverage gaps, not a copywriter chasing catchy titles.
- Task. A broad candidate inventory first, before any filtering, across every dimension in your brief.
- Context. The cluster brief, the persona, your product facts, your existing pages, competitor observations, real customer language.
- Rules. What to include, what to exclude, what it must not invent, and how to label anything uncertain.
- Output schema. One record per candidate, fixed fields, every time.
- Completeness check. A coverage matrix, plus the dimensions that came back thin or empty.
OpenAI's own prompting guidance points the same way: clear instructions, relevant context, explicit output requirements, and complex work broken into its sub-requests. It also notes that model output is not deterministic, so treat any single run as a draft of the inventory, not the inventory.
Ask for JSON or a table so the output is comparable run to run. A record that works:
{
"candidate_title": "",
"primary_question": "",
"question_variants": [],
"audience": "",
"funnel_stage": "",
"intent": "",
"why_it_belongs_in_cluster": "",
"evidence_to_include": [],
"existing_page_match": "",
"competitor_or_ai_gap": "",
"overlap_risk": "low | medium | high",
"recommended_format": "",
"confidence": "high | medium | low",
"missing_information": []
}
If you are calling an API, structured outputs can force the response to match a schema you supply, which is stronger than plain JSON mode. A schema-valid record can still be wrong or invented, so treat schema validity as a formatting check, never a truth check.
Done means. Every candidate comes back in the same shape, with a coverage matrix by dimension and stage and a missing-information list. No idea is accepted just because its title reads well.
Where people go wrong. Pasting source material above the instructions. Mixing your real pages in with hypothetical ideas and no labels. Asking for "all the topics" without ever defining what all means. The worst version is letting the model invent search volume, rankings, citations, or customer quotes. Tell it to write "not known" instead.
Pro tip: Run two passes, not one. Pass one goes wide and labels its own uncertainty. Pass two audits that inventory for missing dimensions, unsupported claims, and repeated intent. Filtering inside pass one is what kills the odd, specific buyer question that would have been your best page.
Step 3: Expand the list across every question lens
One prompt gives you one angle. The best content ideation AI prompts make the model walk the cluster several times, through a different lens each pass.
What to do. Ask it to work through these separately:
- Definitions, terminology, and beginner prerequisites.
- "How do I" implementation questions.
- Use cases, real workflows, and audience or industry variations.
- Features, capabilities, and integrations.
- Alternatives and comparisons.
- Cost, effort, timeline, and resourcing.
- Risks, limits, trade-offs, and objections.
- Setup, migration, troubleshooting, and maintenance.
- Measurement and success criteria.
- Compliance, security, or governance, where those apply.
- Decision questions: who should buy, when not to, how to choose.
Tell it to write questions the way a buyer would say them out loud, not as tidy SEO noun phrases, and to tag each one with a stage and the evidence a credible answer would need.
One instruction does a lot of work here: "For each lens, list the questions a buyer, a practitioner, an evaluator, a skeptic, and an existing customer would ask. Keep two questions separate only when the answer, the evidence, or the decision differs."
That last clause is the whole game. It is how you find subtopics with AI without ending up with forty pages that say the same thing.
Done means. Every applicable lens has at least one candidate, every candidate has a primary question, and the matrix marks each dimension covered, thin, or not applicable.
Where people go wrong. Treating every phrasing as a page. "What is X" and "X meaning" are one page. "How does X work in a regulated enterprise" may genuinely be another, because the evidence and the audience change.
A candidate you can brief looks like "How should a marketing lead evaluate AI citation tracking across platforms." A candidate you cannot looks like "AI citation tracking tips."
Step 4: Ground the ideas in real buyer and search evidence
Here is the shift that separates a working inventory from a nice list. Use AI to organize evidence, never to manufacture it. This is where AI content brainstorming AEO work stops being guesswork.
What to do. Feed the model things that actually happened:
- Sales and customer success call notes.
- Support tickets and the problems that keep repeating.
- Your internal site search terms.
- Autocomplete, related searches, and People Also Ask.
- Community and forum language, including the edge cases.
- Product analytics on setup pitfalls and time to value.
- Search Console queries and pages.
- Competitor pages, and the pages AI answers cite today.
Keep the exact phrasing wherever you can. Tag every input with its source, its date, and how confident you are in it. Strip personal or confidential details before anything customer-facing goes near an external model.
Search Console's Performance report groups by query and by page and gives you clicks, impressions, CTR, and average position. Use it to spot demand you already have and pages picking up impressions without really answering the question. Do not read average position as a fixed rank.
Then prioritize the questions themselves on five things:
- Impact: how much the answer moves a deal, an activation, or a retention risk.
- Evidence: whether you can answer it credibly today, with steps, data, docs, or an expert.
- Effort: how hard a precise, on-brand answer will be to produce.
- Stage coverage: which gap it fills, awareness, consideration, or decision.
- Opportunity: whether the answers out there today are weak, stale, or missing you entirely.
Rate each from one to five, add them up, subtract an effort penalty. That is an editorial scoring method you own, not an industry standard.
Done means. Every high-priority candidate has a real evidence source, or is flagged as needing research first. The inventory carries actual buyer language, not just model paraphrase. Anything with no business relevance and no evidence path is downgraded or gone.
Where people go wrong. Using search volume as a stand-in for buyer value. A quiet question about integrations, security, pricing, or migration can decide a deal that a high-volume definition query never touches.
Step 5: Check the list against your coverage and your AI visibility
AI topic cluster ideas feel complete right up until you compare them with what already exists. Run the list against three inventories.
What to do. Check each candidate against your own site first. Which page answers this already, and is that answer current, complete, and aimed at the same reader and stage?
Then competitors. Which questions do they cover that you do not? Which of their pages get cited for the prompts your buyers actually ask?
Then AI visibility. For the prompts that matter, are you mentioned, cited, absent, or described wrong? Which of your pages show up as sources, and which competitor pages are taking the citations you want?
Give every candidate one of five statuses:
| Status | What it means |
|---|---|
| Covered | A suitable page already answers it |
| Thin | A page exists but misses subquestions, evidence, or intent |
| Untapped | You have nothing suitable |
| Consolidate | It overlaps an existing page and should strengthen it, not add a URL |
| Monitor | Plausible, but the evidence or the priority is not there yet |
Google has said its AI Overviews and AI Mode may run multiple related searches across subtopics and sources, and can surface a wider set of supporting links than classic Search. Read that as a reason to cover related questions well, not as proof that every related query deserves its own URL. Google has also said no special AI file, AI-only markup, or schema is needed to be eligible. Normal crawlability still carries the weight.
This is where a tracker pays for itself, because guessing at your own visibility is slow and usually wrong. DeepSmith's AI Visibility watches the prompts you choose on a schedule and reports mention rate, citation rate, share of voice, answer history, which of your pages get cited, and which competitor pages win the ones you lose. Content Map does the other half. It crawls your site and unlimited competitor sites onto one shared topic taxonomy, then separates coverage gaps, where you publish but a competitor publishes more, from untapped topics, where they publish and you have nothing. Page counts and funnel spread per topic make a top-heavy cluster visible instead of assumed.
That turns your AI content brainstorming AEO list into a prioritized one, backed by observed data.
Done means. Every candidate has a coverage status, a decision about the existing page, a competitor or visibility note where one applies, and an evidence requirement.
Where people go wrong. Reading "a competitor has a page" as "we need a page." Look at the question, the intent, the evidence, and the audience. Their page might be a gap to fill, a weak answer to beat, or a paragraph inside something you already have.
Step 6: Merge the overlaps and cut the weak spokes
This is the step everyone skips, and the one that decides whether your cluster reads as a library or a pile. AI topic cluster ideas arrive redundant by default, because the model has no memory of the page it proposed four records ago.
What to do. Ask AI to audit its own inventory, then review that audit yourself. Compare candidates on primary question, intent, audience and stage, required evidence, format, existing page target, and the next action you want from the reader.
Then make one of three calls on each pair:
- Keep separate. The pages answer materially different questions, or serve different stages, audiences, evidence sets, or actions.
- Merge. One page can answer both without losing its focus or burying the main answer.
- Keep it as a section or an FAQ. The question is worth answering, but it is too narrow or too dependent on the main answer to hold its own URL.
Make the model state its reason for every merge and every keep-separate, then check the reasoning. Google has written about filtering duplicate documents and consolidating URL properties, which supports not shipping near-duplicates. It does not hand you a threshold for when two editorial ideas are the same idea. That call stays yours.
A quick overlap test. Two candidates are probably one page when nearly all of these match: same audience, same stage, same primary question, same format, same evidence, same next action. They are probably distinct when swapping one for the other changes the answer, the proof, the workflow, or the decision.
Done means. Every surviving spoke has one primary question, one dominant intent, one audience and stage, an evidence plan, and a written reason it does not duplicate its neighbors. Every rejected candidate has a disposition: merged, section, FAQ, deferred, or removed.
Where people go wrong. Publishing another page because you found another keyword. Keyword variation on its own is not a separation test, and it is how clusters bloat.
Step 7: Score and sequence what is left
You have a clean list. Now put it in order, in a way you can defend.
What to do. Score each approved spoke on a visible worksheet rather than an opaque ranking.
| Score | The question you are answering |
|---|---|
| Business impact | Could this influence a deal, an activation, or a retention risk? |
| AEO opportunity | Are you absent, uncited, or losing to a competitor on a tracked prompt? |
| Coverage value | Does it close an untapped, thin, or missing-stage area? |
| Evidence readiness | Can your team support the answer today? |
| Audience clarity | Is the reader and their job specific? |
| Production effort | How much research and review will it need? |
| Overlap risk | Could it still duplicate something approved? |
A high score is a prioritization signal, not a prediction of ranking or citation, and nobody can sell you one. Put the quick, evidence-ready, strategically useful gaps into production now, and schedule the research-heavy topics once you know who owns the evidence.
Sequence so the cluster does not come out all awareness. Balance definitions with implementation, comparisons, objections, and decision support. Then plan the links: the pillar points to the spokes, and the spokes point back or across when that helps someone navigate.
If you would rather start from evidence than a blank worksheet, DeepSmith's Opportunity Agents read your own AI Visibility and Content Map data and return ideas with the reason attached. Pick a direction, growing AI visibility, building topical authority, or closing an awareness, consideration, or decision gap against competitors, and each idea arrives carrying the data point that justifies it. You set the window (30, 90, or 180 days), how many ideas you want, and instructions that override the defaults. That automates the evidence gathering, not the editorial call you just made.
Done means. The backlog is ordered, every score has a reason next to it, and you could defend the first batch in a planning meeting without hedging.
Where people go wrong. Ranking by ease alone. Do that for two quarters and your cluster becomes a wall of easy informational pages with no comparisons, no objections, and nothing where competitors already win citations.
Step 8: Turn approved spokes into briefs, then keep the map alive
Almost there. This last step is what makes the whole thing repeatable instead of a one-off afternoon.
What to do. Write a brief per spoke: the working title and primary question, the one-sentence answer the page owes the reader near the top, the audience and stage, the buyer-language variants, the required subquestions, the evidence needed, the pages to link to and from, what not to claim, and the format.
Draft from the brief. Then run a coverage check on the draft: ask AI to map every required question to a section, or mark it unanswered. Have someone who knows the subject verify the facts and the product details.
Structure for answers without pretending there is a formula. Clear titles, descriptive headings, the answer near the top of each section, sections that stand alone, scannable lists and tables, accurate metadata. Google has said generative AI features need no extra technical setup or special schema. The page still needs normal Search eligibility and crawlability, and indexing and serving are never guaranteed.
When a page ships, update its inventory status and reclassify anything that now competes with it. Keep a dated record of the prompt set, the model, the evidence inputs, the scores, and the decisions. Next cycle starts from a live map instead of a blank page.
Production is where a lot of good backlogs quietly die. DeepSmith's Content Studio moves an approved idea from New Ideas to Planned Content to Produced Content, and the Writer turns a planned idea into a researched, internally and externally linked article with a cover image and publish-ready metadata. Autowrite runs it on its scheduled date with nobody in the app. Deep IQ holds your company, product, persona, brand voice, and content type context, so every run is grounded in your information instead of re-briefed from scratch. That is execution. It does not make an unverified idea a good one.
Done means. Every approved spoke has a brief, an evidence owner, a production status, and a plan for measuring it after publication. Your inventory can show which questions are covered, thin, deferred, consolidated, or still open.
Where people go wrong. Stopping at publish. Your site changes, competitors publish, and buyers start asking something new. Recheck the inventory and the visibility data on a set cadence.
The prompt, in one block
Adapt this. The logic matters more than the wording, and no single prompt guarantees completeness.
ROLE
You are a content strategist auditing coverage for one topic cluster. Maximize
useful recall first, then find gaps and overlaps. Do not invent search volume,
rankings, customer statements, citations, product capabilities, or evidence.
CLUSTER BRIEF
Pillar topic, personas, product context, funnel stages, limits, in and out of scope.
KNOWN EVIDENCE
Existing pages, sanitized customer language, search data with dates,
competitor and AI-answer observations, product facts and claims to avoid.
TASK
1. Build a broad candidate inventory across every lens in the brief.
2. Per candidate: primary question, audience, stage, intent, job to be done,
evidence needed, format, existing-page match, gap reason, overlap risk,
confidence, missing information.
3. Produce a coverage matrix by lens and funnel stage.
4. Flag anything thin, unanswered, unsupported, out of scope, or duplicated.
5. Do not filter out unusual buyer questions before explaining their relevance.
OUTPUT
Records, then: covered, thin and unanswered dimensions, proposed merges,
candidates needing human evidence, assumptions.
VALIDATION
Every applicable lens and stage represented, one primary question per record,
no two records with the same intent and answer, unsupported claims labeled unknown.
Then run a stricter second pass over the output: "Review this inventory as a strict editor. Find missing buyer questions, duplicate intent, unsupported assumptions, weak evidence, vague titles, and candidates that should be sections rather than pages. For each issue, give the record, the reason, and a recommended action."
What to do next
Pick one cluster. Just one. Write the brief today, run the two passes tomorrow, and give yourself an hour for the overlap audit before you score anything. You will finish the week with a spoke list you can defend, which is more than most teams have.
When you want the evidence half to run itself, that is what DeepSmith is for. Start a 7-day free trial and point it at the cluster you just mapped.
