Most lists of ChatGPT SEO prompts treat the prompt as the hard part. It isn’t. A prompt like “find low-competition keywords for my bakery” fails because ChatGPT has no keyword database, no Search Console access and no crawl of your site, so it fills the gap with plausible guesses. What makes a prompt reliable is what you paste into it and how you check what comes back. This library gives you each prompt with three things the usual listicles skip: the input it needs, the output format it must return, and a pass/fail test you run before trusting the answer.
Why Most ChatGPT SEO Prompt Lists Fail
Page one for this query is mostly long lists of one-line prompts, and some of their advice no longer holds: prompts that generate FAQ schema “for rich results,” for example, are now pointless: Google’s FAQPage documentation says the FAQ rich result stopped appearing in Google Search from May 7, 2026.
The deeper problem is structural. A one-line prompt asks the model to supply both the data and the judgement. It’s decent at judgement and bad at data it doesn’t have, and it won’t always say when it’s guessing. Every prompt below follows the same rule: you supply the data, ChatGPT does the reasoning, and a test catches the failures.
The Anatomy of a Prompt That Holds Up
OpenAI’s own prompt engineering guide recommends structuring instructions as identity, instructions, examples and context, and using Markdown headers or XML-style tags to mark where pasted material starts and ends. For SEO prompts, that becomes four parts:
- Context: your data, clearly delimited. A keyword export, a Search Console export, the article text, a list of your URLs.
- Task: one job per prompt. “Cluster these keywords” works; “cluster, prioritise and write briefs” produces mush.
- Format: the exact columns or structure you want back, so the output is comparable between runs.
- Test: the check you’ll run afterwards, often written into the prompt as a rule (“every keyword must appear exactly once”).
The test is what turns a list of SEO prompts into a library. Without one, you’re judging output by how confident it sounds.
Set Up a ChatGPT Project Before You Prompt
Don’t retype your house rules every session. OpenAI’s help centre describes Projects as workspaces with their own files and instructions, and those instructions apply to every chat inside the project. Create one project per site or client and paste this in as the instructions:
You support SEO work for [SITE], a [BUSINESS TYPE] serving [MARKET].
Rules:
- Use only data I paste or upload for any number (volume, clicks,
position, difficulty). If a number isn't in my data, write "no data".
- Never invent URLs, statistics, quotes or sources.
- Spelling: [UK/US]. Tone: [PLAIN DESCRIPTION].
- If you're unsure, say so in one line instead of guessing.
- Return tables in Markdown unless I ask otherwise.A word on client data. Per OpenAI’s Data Controls FAQ, you can switch off “Improve the model for everyone,” and Temporary Chats aren’t used to train models and are deleted after 30 days. Workspace plans carry their own policies, so check yours before pasting a client’s analytics exports.
Keyword Research Prompts
These are the most abused ChatGPT prompts for SEO. Never ask the model for search volumes or difficulty scores; it has no source for them. Export keywords from a real tool first, then let ChatGPT do the sorting.
1. Cluster a keyword export
Below, between ### marks, is a keyword export (keyword, volume, KD).
Group the keywords into clusters where one page could rank for every
keyword in the cluster.
Rules: every keyword appears in exactly one cluster; add no keywords
that aren't in the list; name each cluster after its highest-volume
keyword.
Output: cluster | keywords | total volume | likely intent
###
[PASTE EXPORT]
###Test: the keyword count out must equal the count in, and cluster totals must add up. Then spot-check three clusters in Google. The model groups by meaning and shared words, but Google sometimes ranks different pages for near-identical phrasings. If the top results for two keywords barely overlap, split them.
2. Classify search intent
Classify each keyword below as informational, commercial investigation,
transactional or navigational. Add a confidence column (high/low).
Mark "low" whenever a keyword could reasonably carry two intents.
[PASTE KEYWORDS]Test: check every low-confidence row against the live results page.
SERP and Content Brief Prompts
ChatGPT can search the web, but its results aren’t Google’s results for your market. For brief work, run the search yourself in a private window, copy the top ten titles and URLs, and paste them in.
3. Read the SERP
These are the titles and URLs ranking in Google [COUNTRY] for
"[KEYWORD]": [PASTE TOP 10].
What page type, angle and depth does Google reward here? What do all
ten appear to miss? Quote the title or URL behind each observation.4. Write the brief
Using that SERP analysis and this keyword cluster [PASTE], write a
content brief: working H1; intent in one sentence; the reader's five
main questions in priority order; an H2 outline; entities and terms a
subject expert would mention; what we can add that the top results
lack, using only this list of what we have: [YOUR DATA, CASES, EXPERIENCE];
internal links chosen only from these URLs: [PASTE].
Where a statistic is needed, write [STAT NEEDED]. Don't supply one.Test: every internal link URL appears in your pasted list, and the “what we can add” section names something you actually have.
Title Tag and Meta Description Prompts
5. Titles and meta descriptions
Write 5 title tags and 3 meta descriptions for the page below.
Focus keyword: [KEYWORD].
Titles: keyword near the front, roughly 50-60 characters, no repeated
words, no year unless the page is updated yearly.
Meta descriptions: 140-160 characters, state the payoff, match the page.
Return a table with a character count column.
[PASTE H1 AND FIRST 300 WORDS]Test: recount the characters yourself. Models read text as tokens, not characters, so their counts are often off. The targets are display guidelines, not rules. Google’s title link documentation says there’s no limit on title length and that titles are truncated to fit the device width, and it warns that repeating the same words “can make your results look spammy.” Its snippet documentation adds that snippets are primarily created from the page content itself, so a meta description that doesn’t match the page tends to get replaced.
Content Refresh and Internal Linking Prompts
6. Refresh from Search Console queries
Here is a Search Console query export for [URL] for the last three
months (query, clicks, impressions, CTR, position), then the article.
Find queries with meaningful impressions that the article doesn't
answer directly. For each: the query, the section it belongs in, and a
2-3 sentence addition. Ignore queries under [N] impressions. Don't
rewrite sections that already answer their queries well.Test: every query it cites must exist in your export. Invented queries are the classic failure.
7. Internal link suggestions
Here are our URLs with titles [PASTE], then a new article [PASTE].
Suggest up to 8 internal links: the anchor phrase exactly as it appears
in the article, the target URL from my list, and one line on why.
Anchors must be exact substrings of the article.Test: use Ctrl+F to find each anchor in the article. If one isn’t there word for word, the model paraphrased.
Technical SEO Prompts: Regex, Redirects and Schema
8. Search Console regex
Write a regex for the Search Console Performance report's Custom (regex)
filter. It uses RE2 syntax, so no lookaheads or backreferences.
Match queries that start with how, what, why, when, can, does or is.A good answer looks like ^(how|what|why|when|can|does|is)\b. Google’s help page confirms the filter uses RE2, matches partially unless you anchor with ^ or $, and is case-insensitive by default. The RE2 syntax reference lists lookaheads and backreferences as not supported. Test: paste it into the filter and check the first rows.
9. Redirect map
List A is old URLs; list B is new URLs with titles. Map each old URL
to the most relevant new URL based on slug and topic. If there's no
good match, write "no match". Don't default to the homepage.
Output CSV: old,new,confidenceTest: review every low-confidence and “no match” row by hand.
10. JSON-LD from your facts
Write JSON-LD for a [Article/Product/LocalBusiness] using only these
facts: [PASTE]. Leave out any property I didn't supply.Test: run it through Google’s Rich Results Test and a schema validator. Google’s guidance on generative AI content says AI-written structured data should comply with its general guidelines and be validated before it goes live.
SEO Prompts to Retire
- “What’s the search volume for…” The model has no live keyword data. Any number it gives is a guess with no source.
- “Write 30 blog posts about…” Google’s spam policies list “using generative AI tools or other similar tools to generate many pages without adding value for users” as scaled content abuse. Our take on why quality gates beat volume covers the alternative.
- “Audit my site.” ChatGPT doesn’t crawl your site the way a crawler does, so it can’t see your redirect chains, orphan pages or rendering issues.
- “Find me backlinks” or “What’s my domain authority?” Same problem: no link index behind it.
- “Generate FAQ schema to win rich results.” That rich result no longer shows in Google Search.
How to Test a Prompt Before It Joins Your Library
“Tested” should mean something specific. Before a prompt goes into your project, put it through four checks:
- Known-answer run. Use it on a case where you already know the right answer, such as a keyword set you’ve clustered by hand or a page whose queries you know.
- Repeat run. Run it three times in fresh chats. If the structure changes between runs, tighten the format section.
- Adversarial input. Feed it something messy, like duplicate keywords or a URL list with typos, and see whether it follows the rules or quietly “fixes” things.
- Date-stamp it. Note the date and model you tested on. OpenAI updates ChatGPT’s models regularly, and a prompt that passed in spring can drift by autumn.
Once a prompt passes, paste its best output into the prompt as an example; OpenAI’s guide recommends showing the model examples of the output you want.
ChatGPT SEO Prompts FAQ
What are the best ChatGPT prompts for SEO?
The ones that come with their own data and a test: clustering a real keyword export, refreshing a page from Search Console queries, writing regex and redirect maps. Prompts that ask the model to supply SEO data perform worst.
Can ChatGPT do keyword research?
It can brainstorm seed terms and sort keywords you give it, but it has no search volume or difficulty data. Pull those from a keyword tool, then use ChatGPT to group and prioritise them.
Will Google penalise content written with ChatGPT SEO prompts?
Not for using AI as such. Google’s guidance says generative AI can help with research and structure, but mass-producing pages without adding value can count as scaled content abuse. Google also says to fact-check AI output before publishing.
Should I use ChatGPT search or paste the SERP myself?
Paste it yourself when the SERP matters. ChatGPT’s web search is useful for research, but it doesn’t reproduce Google’s results for your country and device.
Where Prompts Stop and Data Starts
Half the prompts above begin with “paste your export,” which means exporting CSVs and checking the model didn’t drift from the numbers. SEO Rocket was built to remove that step. Its chat runs keyword research on real Ahrefs data, competitor gap analysis and a real-crawler technical audit with JavaScript rendering, with Search Console and GA4 connected. The volumes it reports come from that data, not the model’s memory. Its article writer runs validation gates, the same output tests described here, before a draft reaches you.
Whichever tool you use, a playbook proven across 1,000,000+ ranking pages applies to ChatGPT SEO prompts too: supply real data, ask for one job at a time, and never ship output you haven’t checked. If you also work in Claude, our guide to Claude for SEO covers the equivalent workflows, and how to build an AI SEO agent shows how to chain these prompts once they’ve passed.
