# ChatGPT for SEO: what it does well, what it invents

ChatGPT is a writing and reading assistant, not a data source. It turns a results page you paste in into a brief, drafts a section whose facts you supply, and rewrites for length or register. It cannot tell you a search volume, a position, or anything about your own site, and it will invent all three with total confidence if you ask.

Updated 2026-09-15 · Source: https://porteur.ai/guides/chatgpt-for-seo

## The jobs it does well

All of these have one thing in common: you supply the material and it does the shaping. None of them asks the model to know something about the world that you did not give it.

- A brief from a results page you paste in: the titles that rank, the questions under them, what every page covers and what none of them covers.
- An outline for a page, once you have told it the query and the reader.
- A first draft of a section whose facts you supply in the prompt.
- Rewriting: shorter, plainer, second person, no marketing register.
- Twenty title variants for one page, from which you keep one.
- The questions a buyer asks about a product you describe, which become the FAQ and the headings.
- Checking a finished page against a checklist you wrote.
- A translation you then have reviewed by someone who speaks the language.

The assistants with search can also fetch a page you name and summarise it, which is useful for reading a rival's guide quickly. Treat the summary as a summary, not as a quote.

## What it invents, and why

A model answers from patterns in language. Asked for a figure it was never given, it produces a figure shaped like the right answer. This is not a bug you can prompt away, and it is worst exactly where it matters.

| Ask it for this | What you get | Where the real answer lives |
| --- | --- | --- |
| The search volume of a query | A plausible round number, often wrong | A keyword tool, or Bing Webmaster Tools for free |
| Your position on a query | A confident guess | The Performance report in Search Console |
| How much traffic a page of yours gets | An invented figure | Search Console and your analytics |
| A tool's price or feature list | Last year's pricing, or a feature that never existed | The tool's own pricing page, today |
| What Google's policy says | A paraphrase that drifts from the wording | Google's own documentation |

> The rule that saves you: every number and every claim about the outside world comes from you or from a source you opened. The model arranges; it does not know.

## Feed it the results page, not your imagination

The single highest-value thing you can paste into a chat is the results page for the query you want. It contains the brief: what Google has decided the searcher wants, and what the pages that satisfied it look like.

1. **Search the query in a clean browser** Use a private window in the country you sell to. Your own history and location change the page otherwise.
2. **Copy the first ten titles and their sites** Also copy the People also ask questions and, if one opens, the pages the AI overview cites.
3. **Paste the lot with one instruction** Ask for the intent behind the query, the sections every ranking page covers, and the ones only one page covers.
4. **Ask what is missing** The gap in the ten results is the reason your page has a chance. It is usually a worked example, a table of real numbers, or a straight answer nobody gives.

That is a brief, built from evidence, in five minutes. Everything the model said about the results is checkable against the page you pasted.

## Four prompts worth keeping

A prompt is a brief: the audience, the facts, the constraints, and an example of the shape you want. These four cover most of the work.

```
Here are the top ten results for the query “how to import a CSV”:
[paste titles and domains]
And the People also ask questions:
[paste questions]

Tell me: 1) what the searcher wants, in one sentence; 2) the sections that
appear on most of these pages; 3) the questions none of them answers.
Do not suggest anything that is not visible in what I pasted.
```

What it is for: turning a results page into an outline. What it cannot do: tell you the volume of the query, or whether you can rank for it.

```
Draft the section “Where the import can fail” for a guide aimed at a solo
founder who writes their own code. Use only these facts:
- the file must be UTF-8; Excel exports Latin-1 by default
- rows over 50,000 are processed in the background
- a missing email column is the only fatal error
Second person, British spelling, no em-dashes, no exclamation marks,
sentences under twenty-five words. About 150 words.
```

What it is for: a first draft whose facts are yours. What it cannot do: add the fifth fact you forgot, which is why you check the draft against your own list.

```
Rewrite this paragraph in plain second person. Cut every adjective that
carries no information. No marketing register. Keep every figure exactly
as written. Return three versions, shortest first.

[paste paragraph]
```

What it is for: killing the register of a draft that reads like a brochure. What it cannot do: decide which version is true, since it will happily keep a wrong number exactly as written.

```
Give me twenty title options for a page that answers “how to import a CSV”.
Rules: under sixty characters, the words people search near the front, a
promise not a tease, no year, no colon-and-clickbait pattern, sentence case.
```

What it is for: options you would not have thought of. What it cannot do: know which one matches the page, so read the page again before you choose.

## The workflow that keeps facts honest

The pattern that works is the same whoever writes the draft. It has four parts and the model touches only one of them.

1. A brief that carries the facts and where each came from. Anything not in the brief may not appear in the page.
2. A draft, by you or by the model, from that brief.
3. A human pass: every figure checked against the source, the first-hand detail added, the register cut.
4. A read-back that refuses invented numbers. This library runs one: any figure in a page that is not in its brief is flagged before the page is published.

Skip step one and the model fills the gap with invention. Skip step three and the page reads like a hundred others. Skip step four and one wrong number survives for a year.

## What it cannot know about your site

A chat has no access to your Search Console, your analytics or your database. When you ask it why your traffic dropped, it will produce a list of the usual reasons, which is a list you could have written yourself, and none of them will be checked against your data.

The honest sequence is to bring the data to the chat: export the queries and pages from the Performance report, paste the rows, and ask it to sort and read them. Now it is working with your figures rather than with the shape of the question.

We make Porteur, which does that reading from your own figures once and writes the findings out, if you would rather not paste rows into a chat.

## A worked example, query to draft

Say you sell an invoicing tool at yourproduct.com and you want the query “how to invoice a client abroad”.

1. **The results page** Ten results: four bank blogs, three accounting firms, two competitors, one forum thread. The questions underneath are about currency, VAT and getting paid.
2. **The brief** The model reads the paste and reports that every page covers currency and none covers what happens when the client pays late in another currency. That gap is your section.
3. **The facts** You write the facts yourself: what your product does with multi-currency invoices, what it does not do, and the two rules your support team repeats every week.
4. **The draft and the pass** The model drafts the sections from those facts. You add the sentence only you can write: the mistake your customers actually make. You check every figure. You write the title last.

Elapsed time: an afternoon. The part that made the page worth publishing is the part the model could not do.

## Questions

### Is using ChatGPT to write pages against Google's rules?

No. Google says it rewards quality content however it is produced, and there is no penalty for AI assistance as such. What its spam policies target is scaled content abuse: pages produced at scale, by automation or by people, that exist to rank rather than to help. One page written with a model, from your facts, edited by you, is not that.

### Can ChatGPT tell me the search volume of a keyword?

No. It has no keyword database and no live access to one, so it produces a plausible figure instead of a true one. Use a keyword tool, or Bing Webmaster Tools, which gives volumes free.

### Can it check my rankings?

No. Positions differ by country, device and personalisation, and the model cannot see any of that. The Performance report in Search Console gives your average position from Google's own data, and a rank tracker measures specific queries on a schedule.

### Which is better for SEO work, the free or the paid version?

The practical difference is the search and browsing features and the longer context, both of which matter because this work is mostly pasting large things in. Any model that can read a long paste and keep to instructions will do the jobs on this page.

### How do I stop it inventing facts?

Put the facts in the prompt and forbid anything outside them, then check the draft against your own list. A model told to use only the facts you supplied still drifts, so the check is the part you cannot skip.

### Should I tell readers a page was written with AI?

Google does not require disclosure and has no field for it in structured data. Its guidance is to use author bylines where readers would expect them, so the question to answer is whether a reader would feel misled, not whether a policy demands it.

## Read next

- [AI-written content and Google: the policy, in Google’s words](https://porteur.ai/guides/ai-content-and-google): Google does not penalise AI writing as such. It acts on pages made at scale with nothing of their own. Here is the policy and where the line sits.
- [Prompts for SEO work that are worth keeping](https://porteur.ai/guides/ai-prompts-for-seo): A prompt is a brief: audience, facts, constraints, shape. Six prompts written out in full, what each is for, and where each one lies to you.
- [A writing workflow with AI that does not produce filler](https://porteur.ai/guides/ai-content-workflow): Five steps: the facts, the brief, the draft, the human pass, the read-back. The shapes to copy, and the failure each step prevents.
- [What AI SEO means, and the three things it can refer to](https://porteur.ai/guides/what-is-ai-seo): AI SEO means three things: using AI to do SEO work, being cited in AI answers, and Google’s AI in Search. Here is what to do on each.
- [SEO for ChatGPT: how assistants choose what to cite](https://porteur.ai/guides/seo-for-chatgpt): How ChatGPT with search finds pages and decides what to cite, what to fix on your site, and how to test it each month in under 30 minutes.
- [ChatGPT search](https://porteur.ai/glossary/chatgpt-search): What ChatGPT search is, where results come from, how to allow or block its crawler, and how a small site gets cited and measures impact.
- [How ChatGPT search picks and cites sources](https://porteur.ai/guides/how-chatgpt-search-cites-sources): See how ChatGPT search finds pages and cites them, what your page needs to be picked, and how to check if your site is named.

Paste your URL and the free check reads your site, the searches around it and the rivals on them in about thirty seconds, with three findings written whole. No chat required, and no invented figures. Free check: https://porteur.ai/