The five ways ChatGPT damages SEO, and the check that catches each one before you publish
ChatGPT is not bad for SEO. Five specific things it does are, and each is catchable. A verdict first prompt that fails any draft with an unsupported number.
- Works in
- ChatGPT, Claude, Gemini
- You need
- The full draft, headings included · The brief it was written from, if there was one · One sentence on your business
- Written for
- is chatgpt bad for seo
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
ChatGPT is not bad for SEO. Five specific things it does are, and every one of them is catchable before publication. The prompt below runs that check and returns its verdict on the first line, failing the whole draft on a single unsupported number rather than averaging one fabrication against a page of decent prose. For the other half of the argument, see where it genuinely helps.
The five failure modes
They are specific and separable, which is what makes them fixable. Listed in the order they cost you money.
Fabricated specifics. Statistics, dates, studies and quotations that arrive fully formed and attributed to nothing. This is the expensive one, because it survives into decks and client emails and is found by the reader whose trust mattered.
Search figures it cannot have. Volume, difficulty, traffic and rankings are not in a language model. Asked to prioritise, it produces numbers that look exactly like real ones, and a content plan built on them is worse than a plan built on none.
The category average. Ask for a page on a topic and you get the mean of everything written about it. Read on its own it seems fine. Read against the three pages already ranking, it is the same page. At volume this is the thing Google’s spam policies call scaled content abuse: large amounts of unoriginal content that adds no value.
The wrong intent, confidently. No assistant can see the live results for your query. It writes the type of page it assumes the query wants, and when that assumption is wrong nothing on the page needed fixing.
Technical advice about a site it never fetched. Canonicals, redirects, robots directives and indexing, asserted about a configuration it has not seen. This one gets implemented, which is what makes it different from the rest.
Why the verdict fails on one instance
A single unsupported specific fails the whole draft, and that is deliberate rather than proportionate. It is calibrated to what one wrong figure does to everything around it.
A reader who catches one invented statistic does not discount that sentence. They discount the page, and reasonably, because they have no way to tell which of your other claims came from the same place. Grading a draft as broadly accurate hands them that problem instead of solving it, which is why rule 2 forbids the phrase outright.
What to actually do about it
Add a checking step and keep using the tool. The failure modes are properties of unchecked output, not of the assistant.
Run the check in a different model from the one that drafted, cut the passages that would fit any competitor rather than rewriting them, and open every cited source yourself. Then run what survives through the AI content readiness check, because a draft that lost a third of its sentences to this process is usually better and occasionally now too thin to quote.
You are a fact checker for a search team. I am giving you a draft written
with an AI assistant, and the brief behind it. Decide whether it is safe to
publish.
Your first line is one of exactly two strings, alone on the line, with
nothing before it:
DO NOT PUBLISH
PUBLISHABLE AFTER THE EDITS BELOW
Write DO NOT PUBLISH if the draft contains even one unsupported specific:
a statistic, a date, a named study, a quoted person, a cited source, a
search volume, a difficulty score, a price or a version number that is not
attributed to something I could open and read. One is enough. Do not weigh
it against how good the rest of the draft is.
Then write these four sections, under these headings and no others.
UNSUPPORTED SPECIFICS
Every number, date, name, quotation and citation in the draft, one per
line, each with the sentence it sits in, marked attributed or unattributed.
Do not skip one because it sounds plausible. Plausible is how they arrive.
CLAIMS ABOUT MY SITE
Every sentence asserting something about my site, my product, my customers
or my results. You cannot verify any of these. Mark each: confirm, or
remove.
CLAIMS ABOUT HOW SEARCH WORKS
Every sentence stating how a search engine or an assistant behaves. Mark
each as documented, meaning there is a source I can check, or received
wisdom. Received wisdom is not an error. It is a sentence that needs
hedging or cutting.
SENTENCES THAT WOULD FIT ANY COMPETITOR
Passages that could appear unchanged on any other page in this category.
Quote them. These are the sentences that make a page rank for nothing.
Rules you must follow:
1. Do not rewrite anything. You are checking, not editing.
2. Never conclude that the draft is "broadly accurate". Accuracy here is
per sentence and you report it per sentence.
3. If the draft cites a source, do not assume the source says what the
draft claims. Mark it unattributed unless you opened that page in this
session and it contains the claim.
4. Do not estimate search volume, keyword difficulty, traffic or rankings
yourself while checking. A figure in the draft is a finding, not an
invitation to supply a better one.
End with one line and nothing after it:
NOT CHECKABLE FROM WHAT I WAS GIVEN: one item per line, or the
word "nothing"
The brief this was written from: [PASTE THE BRIEF, OR WRITE "none"]
My business: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
The draft: [PASTE THE FULL DRAFT INCLUDING EVERY HEADING]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[PASTE THE BRIEF, OR WRITE "none"]- The brief the draft was written from. It is what lets the checker tell a claim you supplied from a claim the model added, which is the difference between a fact you can stand behind and one that appeared during drafting. Writing "none" is honest and makes every specific in the draft suspect by default.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- Your business in a line. This is what the claims about my site section runs on. Without it the model cannot tell a sentence describing your product from a sentence describing the category, and those are the sentences most likely to be quietly untrue.
[PASTE THE FULL DRAFT INCLUDING EVERY HEADING]- The complete draft with headings, not an extract. Check before publication rather than after, and check the whole thing: the fabricated specific is usually in the section nobody reread, which is the one in the middle.
How to run it
- 01Check the draft before anyone reads it for style
Run this first, ahead of the editorial pass. A draft that fails on facts does not need its sentences improved, and the usual order gets this backwards, polishing prose that is about to be deleted.
- 02Use a different assistant from the one that wrote it
Paste the draft into a different model, or at minimum a fresh conversation. A model checking its own output in the same thread has the reasoning that produced the claim still in context, and it tends to defend rather than check.
- 03Read the verdict line and stop there if it says do not publish
One unsupported specific is the threshold, deliberately. It is not proportionate to the risk of a single wrong figure, it is proportionate to what a single wrong figure does to everything else on the page once a reader finds it.
- 04Open every source the draft cites
Rule 3 marks a citation unattributed unless the model actually opened it. Do that part yourself for anything you intend to keep. Confidently attached links to pages that do not contain the claim are the most common way a draft looks sourced and is not.
- 05Delete the sentences that would fit any competitor
Take the fourth section and cut, rather than rewrite. Those passages are the model producing the average of everything it has read on the topic, and the average is what every other page in the category also published.
- 06Confirm what is left can still be quoted
Cutting the generic sentences shortens a draft considerably and usually improves it. Run the result through the AI content readiness check to confirm the structure still holds and an assistant can lift a clean answer out of what remains.
Questions people ask
Is ChatGPT bad for SEO?
The tool is not. Five specific behaviours are: inventing verifiable specifics, producing search figures it has no access to, writing the category average, answering an intent it cannot see, and giving confident technical advice about a site it has never fetched. Each one is catchable before publication, and none of them is a reason to stop using it.
Will Google penalise AI written content?
The spam policies target scaled content abuse, which they define as creating large amounts of unoriginal content that provides little to no value to users, and they name using generative AI tools to generate many pages without adding value as one way of doing it. The policy is about purpose and outcome rather than the tool. That is less reassuring than it sounds, because unchecked AI drafts are exactly the thin, generic pages the definition describes.
What is the single most damaging failure?
The fabricated specific. A generic paragraph wastes an afternoon. A confident statistic with no source behind it survives into a deck, a client email and three other pages that cite yours, and it is discovered by the one reader whose trust you needed. That is why the verdict line fails on a single instance.
Why check with a different assistant?
Because a model checking its own draft in the same conversation still has the reasoning that produced the claim in front of it, and tends to justify rather than verify. A fresh session, ideally a different model, reads the draft as a stranger would, which is the reading you need.
Should I just avoid AI drafting altogether?
No. Avoid publishing AI drafting. The models are genuinely useful for structure, for outlines, for arguing with, and for the second draft of something you already know. Everything in this list is a reason to add a checking step, not a reason to give up the tool.
Unsubscribe in one click. We never pass your address on.