ChatGPT for SEO: how to use it without shipping slop.

1. The SEO jobs ChatGPT handles well.

ChatGPT is the default SEO assistant for a boring reason: most people reading this have a tab of it open already. The practical question is which SEO jobs to hand it and which ones to keep.

One rule sorts them. ChatGPT is very good at transforming text you give it and unreliable at recalling facts you did not. Every task below starts from material you supply.

• Keyword clustering and intent. Paste two hundred keywords and it will group them by shared intent, label each group informational, commercial, transactional, or navigational, and flag the terms that belong on the same page. With web search on, it can check the live results instead of guessing from the wording. It is also a fast way to fan a topic out into adjacent questions, alongside a blog post idea generator.
• Content briefs and outlines. Given a keyword and the pages that rank for it, it summarizes what they cover, what they skip, and the follow-up questions a reader will still have.
• First drafts. Fast, structurally sound, and generic until you add what only you know.
• Titles and meta descriptions. Ten variants of a title tag at the length you set, which you then cut down by hand or sanity-check against our SEO title generator and meta description generator.
• FAQ sections. It is good at listing the questions buyers ask. Google's Search updates changelog says FAQ rich results stopped appearing in Search on May 7, 2026, so an FAQ block now earns its place through the answers on the page alone.
• Schema markup. JSON-LD for Article, Product, Organization, or BreadcrumbList from a pasted page, ready to validate.
• Search Console summaries. Upload a performance export and ChatGPT's data analysis mode runs Python over the file: queries with impressions and no clicks, pages losing position, groups of queries no page targets. Search Console Help says a direct export from a report is truncated to 1,000 rows, so on a larger site the file is a sample of your data.

If what you want is a product built on top of ChatGPT rather than a way of working, the ranked list of ChatGPT SEO tools covers that. This page is the how-to.

2. Where ChatGPT SEO output goes wrong.

The failures are predictable, which is the useful part. There are three, and each one looks fine until somebody checks.

Search volume and difficulty scores are guesses. ChatGPT has no connection to a keyword database. Ask it how many people search a term and it produces a plausible number anyway. With web search on, it may quote a figure from a page that mentions one, which is a guess somebody else made. Treat every volume and difficulty number from a chat window as fiction until it matches a real source: your own Search Console for queries you already appear for, and a keyword data tool for the rest.

It makes up statistics and sources. ChatGPT's own help center lists fabricated quotes, studies, citations, and references to sources that do not exist among the errors it can produce. Web search reduces the problem without removing it: the help center article on searching the web with ChatGPT says results and citations can be incomplete, outdated, or incorrect, and tells you to open the cited source. An SEO draft with an invented statistic is worse than one with no statistic, because the number is the part that gets quoted. Our automated fact-checker exists because this step is tedious and cannot be skipped.

Drafts carry recognizable AI tells. Readers have learned the shape of an unedited model draft: an opener that restates the question, tidy lists of exactly three, a transition word at the start of every paragraph, a heavy hand with the long dash, and a closing paragraph that summarizes what you just read. None of that triggers a ranking penalty on its own. What it costs is trust, and it makes every article on your site read like every other article on the same topic.

3. Set up a Project before the first prompt.

Paste your style guide into a chat and article one follows it. Article twelve, written in a new chat three weeks later, has lost half the rules and links to a page that does not exist. Memory will not hold them for you: the ChatGPT help center says memory does not retain every detail from every conversation. Most of the drift goes away once the rules live somewhere other than your clipboard, and in ChatGPT that place is a Project. The help center page on Projects says project instructions apply only inside that project and override your global custom instructions, which is the scope you want.

Put five things in it.

1. Brand facts, which fix invented product details. What the product does, who it is for, the prices you are happy to state, and the claims you never make.
2. Style rules, which fix the AI tells. Spelling convention, sentence length, banned phrases, no long dashes, how numbers are written. Short and absolute beats long and nuanced.
3. An internal link list, which fixes links to pages that do not exist. Every URL you want linked, with one line on what it covers, and an instruction to link only from this list.
4. One reference article, which fixes drift in structure and tone. Your best published piece, as the example to match.
5. A source policy, which fixes invented statistics. Every statistic needs a URL you can open, and anything without one is marked [SOURCE NEEDED] instead of being filled in.

Older guides tell you to build a custom GPT for this instead. The help center's custom GPT retirement FAQ now says it plans to retire custom GPTs and help creators migrate them to plugins, so for new SEO work a Project is the place to start.

The full SEO checklist can be pasted almost whole into project instructions. If your SEO work lives in a code repository rather than a chat window, the agentic version of this setup is in our guide to using Claude for SEO. If you draft in Claude, our comparison of Claude models for writing covers which one to use for briefs, drafts and bulk rewrites. Deciding between the two assistants in the first place is a separate question, answered job by job in our head-to-head of Claude and ChatGPT for writing.

4. How to use ChatGPT for SEO, step by step.

Here is the whole workflow for one article, with the check a human has to run at each step. The model does most of the typing. The checks decide whether the page is worth publishing. The writing half on its own (the brief, section-by-section drafting, the claim check and the hand edit) has a separate walkthrough in how to use ChatGPT for content creation.

1. Start from a real keyword list. Export candidates with measured volume from a keyword data source, or pull the queries you already get impressions for from Search Console. ChatGPT does not supply this list. It organizes it. Human check: every keyword has a real volume attached and a results page you have looked at.
2. Cluster. Paste the list into the Project with prompt 01 below. Human check: search two keywords from each cluster. If the results are different kinds of page, the cluster is wrong, whatever the model said.
3. Brief. With web search on, have ChatGPT read the pages ranking for the cluster's head term and list what they cover, what they miss, and the follow-up questions (prompt 03). For a wide or technical topic, ChatGPT's deep research mode goes further and returns a structured report with sources. Human check: open every page the brief cites and confirm it says what the brief claims.
4. Draft. From the brief, inside the Project, so the style rules and link list apply. Pick the model on purpose rather than taking the default: our guide to the best ChatGPT model for writing covers which one to use for long-form drafts. Human check: the query is answered in the first hundred words, and the draft contains at least one thing the ranking pages do not: a number from your own data, a screenshot, a mistake you made and fixed. The rest of the page shape is covered in how to structure a blog post for search.
5. Fact-check. Ask for a table of every factual claim with the URL that supports it (prompt 05 below). Human check: open every URL. Keep the claims the page supports, fix the ones it contradicts, cut the ones with no source.
6. Edit. Strip the tells from section two, cut the summary paragraph, rewrite the opener, add the first-hand material. Human check: read it aloud. Anything you would not say to a customer goes.
7. Publish. Title tag, meta description, schema, internal links from your list, and at least one older page edited to link to the new one. Human check: look at the rendered page, click every link, and run the schema through Google's Rich Results Test.

Budget two to four hours of human time for a 1,500-word article done this way, most of it in the fact-check and the edit. Running the same steps daily, with the checks still enforced, is what a gated SEO automation pipeline is for.

5. ChatGPT SEO prompts for seven everyday tasks.

Seven prompts, written to run inside the Project from section three, so none of them repeats your style rules. Square brackets are placeholders. Each one hands ChatGPT material to work on and ends with an instruction that makes the output easier to check.

6. What Google says about AI-written content.

Google's position has been stable since February 2023, and it is less dramatic than the forums suggest. Google Search Central's post on AI-generated content says appropriate use of AI or automation is not against its guidelines, and that Google focuses on the quality of content rather than how it was produced.

Google's current guidance on using generative AI content says generative AI can be particularly useful for researching a topic and adding structure to original content, and asks for accuracy in the parts people forget are content: title elements, meta descriptions, structured data, and image alt text. Those are exactly the jobs ChatGPT gets asked to do fastest. The line it draws is about volume. Google's spam policy defines scaled content abuse as many pages generated for the primary purpose of manipulating rankings and not helping users, and its first example is using generative AI tools to generate many pages without adding value. A ChatGPT-assisted article that answers the query, has been checked, and adds something the ranking pages lack is fine. Two hundred near-identical articles pushed out in a week is the pattern the policy describes, however good each one looks alone.

7. SEO for ChatGPT is a different job.

One naming trap before the example. People searching “SEO for ChatGPT” usually want the reverse of this guide: getting their own pages cited inside ChatGPT's answers. That is a separate discipline with separate levers: a direct answer near the top of the page, statistics tied to named sources, one consistent name for each entity, and pages that are easy to find in the first place, because an answer built from a web search can only cite what the search step found.

The mechanics are on our LLM SEO page, the step-by-step version is the answer engine optimization guide, and if you want to measure whether ChatGPT names you at all, the LLM SEO tools roundup compares the trackers.

A worked example: one seed keyword through every step.

Take a fictional time-tracking product for agencies, Hourgrain, starting from one seed keyword: billable hours. The company and every number below are invented. The prompts are the ones from section five.

Hands-on time was a little under three hours, and almost none of it was writing. It went into the three places ChatGPT cannot be trusted: deciding what belongs on one page, opening sources, and adding the material only the company had. Skip those and you publish faster, along with an invented survey figure and a 404.

Common ChatGPT SEO mistakes.

1. Planning content around volumes the chat produced. A plan built on those numbers is a plan for queries that may not exist. The model has no keyword database, so pull volumes from a real source before it sees the list.
2. Trusting a citation without opening it. A link in a ChatGPT answer shows that a page exists. Whether that page says what the sentence claims is a separate question, so open every source attached to a number before it goes live.
3. Letting it invent internal links. Without a list of real URLs, the model writes plausible slugs that return a 404. Put the list in the Project and tell it to link only from that list.
4. Keeping the rules in your clipboard. Rules pasted into one chat do not follow you to the next. Put them in project instructions, where every chat inherits them.
5. Pasting a Search Console export and trusting the total without reading the code. The export stops at 1,000 rows, so on a larger site any total computed from it undercounts, and a wrong filter in the Python can make it worse. Read the code and compare the row count with the file before quoting a number.
6. Running the draft through another prompt to hide the tells. A rewrite pass swaps one set of tells for another and often undoes the facts you just checked. Edit by hand, against your banned-phrase list.