ChatGPT SEO Prompts: 25 That Hold Up in Real Work

There is no shortage of ChatGPT SEO prompt lists. Most of them share the same defect: they are collections of instructions that produce impressive-looking output, tested by nobody, with no way of telling whether the answer was true. “Act as an expert SEO consultant and give me the top 20 keywords for my industry” reads like a prompt. It is actually a request for a language model to invent twenty plausible strings and, if you push it, twenty plausible search volumes to go with them.

This library is built the other way round. Every prompt below is organised around a single question: what data does the model need in front of it, and how will you know if the output is wrong? Prompts where the honest answer to the second question is “you won’t” are not included — they are listed near the end, under the prompts to delete.

A survey of SEO practitioners published by Keyword.com in 2026 found that 70% of respondents named poor-quality output, hallucinations, or the time required for quality control as their main limitation when using AI. That is a sample of 97 usable responses, so treat it as directional rather than definitive — but it matches what the work actually feels like. The bottleneck was never getting the model to produce something. It is getting something you would put your name on.

This is a prompt library, not a workflow guide. If you want the underlying method — the six workflows that pay for themselves, the full anatomy of a prompt, and the failure modes with their costs — that is how to use ChatGPT for SEO. If you want the broader tool-agnostic view of which SEO tasks are safe to delegate at all, read how to use AI for SEO.

What separates a prompt that works from one that reads well

Four things, and only one of them is wording.

1. It supplies data rather than requesting recall. ChatGPT has no connection to Google’s index, no keyword database, and no view of your rankings. Every reliable prompt in this library begins with you pasting something real: a keyword export, a crawl file, a Search Console download, your own page copy.

2. It permits the model to refuse. The single highest-value clause in SEO prompting is an explicit instruction to write UNKNOWN rather than estimate. Models produce confident text whether or not they have grounds for it, and the output reads identically either way.

3. It fixes the output shape. A prompt that specifies columns, row counts and character limits gives you something you can paste into a sheet. A prompt that does not gives you an essay you then have to reformat by hand, which eats the time you were trying to save.

4. It is checkable in under a minute. If verifying the output takes longer than doing the task yourself, the prompt has failed regardless of how good the answer looks.

How to read each entry

Every prompt below has three parts: Paste in (what real data the prompt needs), the prompt itself, and The check (the specific thing to verify before you use the output). Square brackets mark what you replace. The prompts are written plainly on purpose — role-play preambles like “you are a world-class SEO expert with 20 years of experience” add tokens and change nothing measurable.

Group 1: Keyword and demand analysis

The rule for this whole group: the keyword data comes from your keyword tool. The model does the sorting, grouping and judgement calls on top of it.

1. Cluster an export into page-level topics

Paste in: a keyword export with columns for keyword, volume and difficulty (200 to 800 rows works well).

Below is a keyword export. Group these keywords into clusters where each
cluster could be served by ONE page. For each cluster give me:
- a cluster name
- the keyword I should treat as the primary target
- every other keyword in the cluster
- the total volume for the cluster (sum the volumes I gave you, do not estimate)
- the search intent: informational, commercial, transactional or navigational

Rules: use only the keywords and numbers in my data. Do not add keywords.
Do not invent volumes. If a keyword does not fit any cluster, put it in a
final group called UNCLUSTERED rather than forcing it.

Output as a table. Data:
[PASTE EXPORT]

The check: sum two or three cluster totals against your own sheet. Arithmetic over pasted numbers is where models slip most often, and it is a five-second check.

2. Find the cannibalisation risk in a cluster set

Paste in: the cluster output from prompt 1, or a list of your existing URLs and their primary keywords.

Here is a list of pages and the primary keyword each one targets. Identify
every pair where two pages are likely to compete for the same query, and
for each pair say which page should keep the keyword and what the other
page should target instead. Judge on the keyword and URL only. Where you
cannot tell without seeing the page content, write NEEDS REVIEW instead of
guessing.

[PASTE LIST]

The check: open the top three flagged pairs. If the model has flagged pages that are obviously distinct, your list is too coarse — add page titles.

3. Turn a cluster into a publishing sequence

Paste in: your clusters, plus a note of which pages already exist.

Here are my keyword clusters and my existing pages. Propose a publishing
order for the clusters that do not yet have a page. Order by: clusters
that support an existing commercial page first, then clusters with the
lowest difficulty in my data, then everything else. For each item give the
proposed URL slug and one line on why it sits where it does. Do not
reorder based on assumptions about my business that I have not told you.

[PASTE CLUSTERS AND EXISTING PAGES]

The check: the sequence should be defensible in a sentence each. If a reason reads as generic filler, the model is padding.

4. Extract the question set from People Also Ask data

Paste in: PAA questions copied from the SERP or exported from your tool.

Below are People Also Ask questions for my topic. De-duplicate them,
group them by the underlying job the searcher is trying to do, and for
each group tell me whether it belongs on an existing page as a section or
justifies its own page. Use only the questions I gave you.

[PASTE QUESTIONS]

The check: count the questions in and out. Silent dropping is the failure mode here.

For the underlying method — where the keyword data should come from in the first place — see keyword research in Singapore.

Group 2: Briefs and content planning

The Keyword.com survey found content briefs and outlines were the single most-delegated task at 77%, ahead of keyword research at 68% and content drafting at 66%. That ordering is telling: practitioners hand over the scaffolding before they hand over the prose.

5. Build a brief from real SERP data

Paste in: the H1 and H2 structure of the top five ranking pages, plus the target keyword.

Here are the heading structures of the five pages currently ranking for
[KEYWORD]. Produce a content brief for a page that covers this topic more
completely. Give me:
- the sections every one of these pages includes (these are table stakes)
- the sections only one or two include (these are differentiators)
- the questions none of them answer, based only on what is visible in
  these headings
- a proposed H1 and H2 outline

Do not invent what these pages say beyond their headings. Where you are
inferring, mark the line INFERRED.

[PASTE HEADINGS]

The check: the “questions none of them answer” list is the valuable part and the easiest to fabricate. Open two of the ranking pages and confirm the gap is real.

6. Stress-test a brief before anyone writes

Here is a content brief. Act as a sceptical editor. List every claim in
this brief that would require a source, every section that would be hard
to write without first-hand experience, and every section that overlaps
with another. Do not rewrite the brief. Just list the problems.

[PASTE BRIEF]

The check: none needed — this prompt produces questions, not facts. It is one of the few where the model cannot really hurt you.

7. Map a brief against your existing content

Here is a new content brief and a list of my existing page titles and
URLs. For each section of the brief, tell me which existing page already
covers it, so I can link to it instead of repeating it. Where nothing
covers it, write NEW. Match on titles only; do not assume content I have
not shown you.

[PASTE BRIEF AND URL LIST]

The check: spot-check three NEW verdicts against your site search. This prompt is how you keep an internal link plan honest, and it pairs with the linking rules in the on-page SEO checklist.

8. Convert a subject-matter interview into a draft outline

Paste in: a transcript of a call with the person who actually knows the subject.

Below is a transcript of an interview with a specialist. Extract:
- every specific claim, number or example they gave
- every opinion they expressed that a competitor would not say
- the points where they hedged or were unsure

Then propose an article outline built around the specific claims and the
distinctive opinions. Quote them verbatim where you use their words. Add
nothing that is not in the transcript.

[PASTE TRANSCRIPT]

The check: this is the highest-value prompt in the library, because it converts genuine expertise into structure without diluting it. Verify quotes verbatim.

Group 3: On-page and metadata at volume

9. Title tags and meta descriptions from real page copy

For each page below I have given you the URL, the current title, the
primary keyword and the first 150 words of the page. Write a new title
tag (maximum 60 characters) and meta description (maximum 155
characters) for each. Rules: the primary keyword appears in the title
near the front; the description describes what is actually on the page
based on the copy I gave you; no superlatives; no "discover" or "unlock";
count the characters and show the count.

[PASTE TABLE]

The check: character counts. Models are unreliable at counting their own output, so re-count in your sheet before shipping.

10. Rewrite a heading structure for answerability

Here is the heading structure and body copy of one page. Rewrite the
headings so each one states the question a reader would ask, and so the
paragraph immediately below it answers that question in the first two
sentences. Do not change the facts. Do not add sections. Return the old
and new heading side by side.

[PASTE PAGE]

The check: read the first two sentences under each new heading in isolation. If they do not answer the heading, the rewrite is cosmetic.

11. Find the thin sections in a long page

Here is a page. For each H2 section, tell me the word count, whether it
contains any specific fact, number or example, and whether it would
survive being deleted. Rank the sections from most to least substantive.
Do not rewrite anything.

[PASTE PAGE]

The check: the word counts. Verify one against your editor.

12. Draft alt text for a set of images

For each image below I have given you the file name, the page it sits on,
and a description of what the image shows. Write alt text for each:
one sentence, describes the image content, no keyword stuffing, no
"image of". If my description is too vague to write good alt text, write
NEEDS DESCRIPTION.

[PASTE LIST]

The check: any alt text that could apply to any image on the site is a fail.

Group 4: Technical triage

13. Turn a crawl export into a ranked action list

Below is a crawl export. Produce a triage table with columns: issue,
number of URLs affected (count from my data), likely impact on organic
performance (high, medium, low), effort to fix (high, medium, low), and
the first action. Sort by impact then effort. Use only issues present in
my export. Do not add best-practice recommendations I did not ask for.

[PASTE EXPORT]

The check: the URL counts. This is the prompt that saves the most hours on a large site, and the counts are the only thing that can be silently wrong. Pair it with the SEO audit process, which is where the triage output belongs.

14. Explain a redirect chain and propose the fix

Here is a list of URLs with their redirect chains and status codes.
For each chain, tell me the final destination, how many hops, and the
single redirect rule that would replace the chain. Flag any chain that
ends in a 404 or a loop. Work only from the data given.

[PASTE CHAINS]

The check: test two proposed rules before deploying anything.

15. Write a Search Console regex filter

Write a regex filter for Google Search Console that matches [DESCRIBE
WHAT YOU WANT]. Use RE2 syntax, which is what Search Console accepts.
Explain in one line what each part does, and give me two example queries
it would match and two it would not.

The check: paste it into Search Console and look at the match count. Instant verification, which is why this prompt is safe.

16. Sanity-check structured data

Here is a JSON-LD block from one of my pages. Tell me: whether it is
valid JSON, whether the required properties for this schema type are
present, and which properties reference content that must also appear
visibly on the page. Do not add properties. Do not tell me schema will
improve my rankings.

[PASTE JSON-LD]

The check: run it through Google’s Rich Results Test as well — the model is a useful first pass, not the validator. And on the wider question of what markup is worth adding, see schema markup in Singapore.

Group 5: AI-search visibility

This group has grown quickly, and it is the one where bad prompts do the most damage — because the outputs are hard to check and the topic is fashionable enough that nobody wants to admit they cannot verify them.

17. Audit a page for passage-level answerability

Here is a page. Google describes query fan-out as "a set of concurrent,
related queries generated by the model to request more information and
fetch additional relevant search results to address the user's query."
For each section of my page, write the specific question that section
answers, and mark whether the answer is complete within that section
without reading the rest of the page. List the sections that fail.

[PASTE PAGE]

The check: read three flagged sections yourself. This prompt is a structural review, not a prediction of citations — do not let it become one.

18. Build a prompt panel for visibility monitoring

My business is [DESCRIPTION] in [MARKET]. Write 25 questions a genuine
buyer might type into an AI assistant at different stages: unaware,
researching, comparing, and ready to buy. Write them as a real person
would type them, not as keywords. Do not include my brand name in any of
them except the final three.

The check: the panel is the input to a measurement process, not a result. How to run and read it is covered in AI brand visibility monitoring.

19. Compare your page against what an answer actually needs

Here is a question a buyer would ask, and here is my page. Extract only
the sentences from my page that could stand alone as an answer to that
question. If there are none, say so plainly. Do not write new sentences.
Do not improve my copy.

[PASTE QUESTION AND PAGE]

The check: the constraint “do not write new sentences” is what makes this useful. If the output contains prose that is not on your page, discard the whole answer.

20. Draft a factual summary block for a long page

Here is a long page. Write a 60 to 80 word summary using only facts
stated on the page. Every number in the summary must appear in the page
text. If the page contains no specific facts, write NO FACTS FOUND
instead of writing a generic summary.

[PASTE PAGE]

The check: search the page for each number in the summary. The NO FACTS FOUND branch is deliberately included — it is a diagnosis of your content, and a useful one.

Group 6: Reporting and stakeholder communication

21. Turn a metrics table into a three-paragraph client update

Here is this month's data. Write a three-paragraph update: what changed,
why it likely changed based only on the data and the activity list I have
given you, and what happens next. Use only my numbers. Where a change has
no explanation in the data I provided, say the cause is not yet known
rather than proposing one.

Data: [PASTE]
Activity this month: [PASTE]

The check: the second paragraph. Invented causation is the standard failure, and “we believe this is due to increased brand awareness” is what it looks like.

22. Rewrite a technical recommendation for a non-technical decision-maker

Rewrite the following for a business owner with no technical background.
Keep every specific number and deadline. Remove jargon or define it in
the sentence where it first appears. Do not soften the recommendation.
Maximum 200 words.

[PASTE RECOMMENDATION]

The check: confirm no number changed.

23. Prepare the objections before the meeting

Here is a recommendation I am taking to a client. List the ten hardest
questions they could ask, ordered by how likely they are. For each, note
what evidence I would need to answer it. Do not write the answers.

[PASTE RECOMMENDATION]

The check: none needed. Like prompt 6, this one produces questions.

Group 7: The Singapore layer

24. Localise copy without breaking the facts

Rewrite the following for a Singapore audience. Change spelling to
British English, change currency references to SGD only where I have
given you an SGD figure, and flag any claim that is specific to another
country's regulations rather than rewriting it. Do not add local
statistics, grant names, or regulatory references that are not in my
source text.

[PASTE COPY]

The check: the flags. This prompt exists because the common failure when localising is not tone — it is a model helpfully inserting a grant name, a government agency or a regulation that does not apply. The instruction to flag rather than rewrite is the whole point.

25. Draft a language variant brief for a multilingual page

Here is an English page. I want a [LANGUAGE] version for the Singapore
market. Do not translate it. Instead, tell me: which sections would need
to be rewritten rather than translated because the reasoning is
culture-specific, which terms have no direct equivalent, and which
examples would not land. Output as a table.

[PASTE PAGE]

The check: have a native speaker read the table, not the translation. Using a model to plan a translation is far safer than using it to perform one.

The five prompts to delete from your library

These are the ones that circulate most widely, and every one of them asks a language model for information it does not hold.

The pattern is consistent: each of these asks the model to recall rather than to process. Rewrite any prompt in that shape so that you supply the input, and it usually becomes a good prompt.

The clauses worth memorising

Most of the reliability in this library comes from a handful of reusable constraint clauses rather than clever phrasing. These are the ones that earn their place.

Adapting a prompt to your own site

A generic prompt becomes a good prompt through three edits, in this order.

Add the constraint that matches your risk. If your output goes straight into a client report, add the UNKNOWN clause. If it goes into a spreadsheet, add the “use only my data” clause. If it goes to a developer, add “do not propose changes outside the files I have shown you”.

Fix the output shape to your actual destination. Name the columns you want, in the order your sheet uses. This is the edit that converts a prompt from a demo into a tool, and it is the one people skip.

Add the domain facts the model cannot know. Your commercial pages, the products you do not sell, the markets you do not serve, the terminology your industry actually uses. Two sentences of context here removes more bad output than any amount of prompt-engineering vocabulary.

Here is the same task before and after. Before: “Write meta descriptions for my service pages.” After: “Here are 14 service pages with their URL, H1 and first 150 words. Write a meta description for each, maximum 155 characters, describing only what the page copy says the service includes. We do not offer web hosting or PR — do not mention them. Return as a two-column table: URL, description, character count.” The second version is usable output. The first is a demonstration.

Before you use any output: the four checks

1. Every number traces to something you pasted in. If you cannot point at the source cell, delete the number.
2. Nothing was silently added. Count items in and items out. Additions are more common than omissions and much harder to spot.
3. Every local or regulatory claim is checked against the primary site. Agency names, grant conditions and legal requirements are the highest-risk category in a Singapore context, and the model will produce them fluently.
4. A named person is accountable for the deliverable. Not “the AI produced it”. That accountability is what the whole verification habit hangs on.

On the data you put in: Singapore’s PDPC ran a public consultation on proposed Advisory Guidelines on the Use of Personal Data in Generative AI, with material published on its consultation page in July 2026. Keyword exports, crawl data and your own published copy are unproblematic. Customer lists, enquiry exports and CRM records are not, and no prompt in this library asks for them.

Google’s position on AI-assisted work is narrower than most people assume. Its guidance on generative AI features states plainly that “the best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems”, and its spam policies target scaled content abuse rather than the use of a tool. A researched, edited, genuinely useful page is fine regardless of what helped draft it.

Where this leaves you

The useful prompts in SEO all have the same shape. You bring the data, you state what “I don’t know” looks like, you fix the output format, and you check the result against the input. Everything else — the role-play openings, the “act as”, the promise of a hundred prompts — is decoration on top of those four moves.

Start with three from this library rather than twenty-five: the clustering prompt, the crawl triage prompt, and the client-update prompt. They cover the three places where the hours actually go. Once those are running reliably against your own data, the rest of the library is mostly variations on the same discipline.

If you would rather have this run as a process than as a habit — with the verification steps built in and someone accountable for the output — that is what our AI SEO service in Singapore is for. You can see how we report on this kind of work in our client case studies.

Where to go next

• AI SEO in Singapore — the pillar guide covering both senses of the term
• How to use ChatGPT for SEO — the workflows and failure modes behind these prompts
• ChatGPT SEO tools — when to move from a chat window to purpose-built software
• The AI SEO tools landscape — what each category of tool actually does
• AI SEO agents — what changes when the tool acts rather than answers

Frequently asked questions

What is the best ChatGPT prompt for SEO?

There is no single best prompt, but there is a best clause: instructing the model to write UNKNOWN rather than estimate when it lacks the information. Adding that one line to any prompt in this library removes most of the fabrication risk, because the default behaviour of a language model is to produce an answer whether or not it has grounds for one.

Can ChatGPT do keyword research?

It can do the analysis layer, not the data layer. Export real keywords and volumes from a keyword tool, paste them in, and it will cluster them by topic and intent quickly and well. Ask it for volumes directly and it will produce numbers with no source, which is how invented figures end up in client documents.

How many SEO prompts do I actually need?

Three to five, used repeatedly, beats a library of fifty used once. In practice the recurring jobs are clustering a keyword export, triaging a crawl file, drafting metadata at volume, and turning monthly data into a written update. If you find yourself running the same prompt on the same data every week, that is a process rather than a prompt, and it belongs in a tool or a script.

Do longer, more detailed prompts produce better SEO output?

Only when the extra length is data or constraints. Adding your keyword export, your page copy or a rule about what not to include improves the output measurably. Adding personas, seniority claims and encouragement does not. The useful test is whether a line you are adding changes what the model can check its answer against.

Is it safe to paste client data into ChatGPT?

Keyword exports, crawl data and your own published page copy are fine. Personal data is a different question: Singapore’s PDPC consulted publicly in 2026 on proposed Advisory Guidelines on the Use of Personal Data in Generative AI, and the responsibility for data passing through these systems sits with the organisation deploying them. Customer lists, enquiry exports and CRM records should not go into a general-purpose assistant.

Will using AI prompts to write content hurt my rankings?

Not the method itself. Google’s spam policies target scaled content abuse — generating many pages without adding value — rather than the use of a tool, and its guidance on generative AI features confirms that standard SEO best practices still apply because those features run on Google’s core ranking systems. The risk is not that you used a prompt. It is publishing unverified claims at volume.