Yes, ChatGPT can cite an article written with AI: OpenAI's guidance for site owners is about whether its search crawler may reach a page, and nothing in it concerns who or what wrote the text. The risk with an AI draft is that it gives the answer nothing to cite: if it only restates what other pages already say, it offers no fact, figure or source that the answer could not take from the pages it was rewritten from.
Mass-produced AI articles are often sold as "AI SEO". Below is what Google's own guidance says about AI-written content, what research found makes a page worth citing, why repeated text gives an assistant nothing to quote, and how to check an AI draft before it goes live.
Does ChatGPT check whether an article was written by AI?
ChatGPT cites web pages when it searches. OpenAI's documentation for site owners describes the crawler behind this: OAI-SearchBot is used to surface websites in ChatGPT's search features, and sites that opt out of it are not shown in ChatGPT search answers, though they can still appear as navigational links. The page is about access. It says nothing about how a page was written and sets no rule that favours or excludes AI-written text.
Not every answer involves a search. In a test we ran on 23 September 2026 with 20 German business questions, the model answered 10 without searching at all and cited nobody. Where it searched, it cited primary sources (laws, ministries, OECD, large accounting firms) on 9 of 10.
That second result is the useful one. When the model did look something up, it mostly went to the originals: the law, the ministry, the OECD. A business article competes with those originals. If it only paraphrases them, the assistant has a better page to cite, and authorship never comes into it. For the wider picture, see how ChatGPT decides which businesses to recommend.
What does Google say about AI-generated content?
Google's position is that the method of production is not the test. In its guidance on AI-generated content of 8 February 2023, Google says its ranking systems aim to reward original, high-quality content "however it is produced", that appropriate use of AI or automation is not against its guidelines, and that using AI gives content no special gains: "It's just content." The same post draws the line: using automation, AI included, to generate content with the primary purpose of manipulating search rankings violates Google's spam policies.
Those spam policies (last updated on 28 August 2026) name the pattern to watch for in any offer of mass-produced articles. Scaled content abuse is when many pages are generated mainly to manipulate search rankings rather than to help users, and the first example listed is using generative AI tools to generate many pages without adding value for users. The policy covers unoriginal content that provides little or no value "no matter how it's created", so pages churned out by people fall under it too.
Google's rules govern Google Search, not ChatGPT. Still, they point the same way as the test above, where the model went to the originals: a page that only repeats other pages adds nothing those pages lack.
What makes a page worth citing, according to research?
The paper that introduced the term, GEO: Generative Engine Optimization, by a team including researchers from Princeton University and IIT Delhi, was published in the proceedings of KDD 2024. The authors built a test engine that fetched the top 5 Google results for each query and had GPT-3.5-turbo write an answer from them. They then rewrote one source page at a time in different ways and measured how visible that page became in the answer. They also repeated the methods on Perplexity.
Three changes worked best: adding citations from credible sources, adding relevant quotations and adding statistics. These methods achieved a relative improvement of 30-40% on one of the paper's visibility measures (position-adjusted word count) and 15-30% on the other (subjective impression). Keyword stuffing, though widely used in SEO, gave little to no improvement, and the authors found no significant gain from a more persuasive, authoritative tone. It was a lab setup with an older model, so read the percentages as the paper's result, not a forecast for your site.
The pattern is the useful part. The changes that worked add substance a reader can check: a figure, a quote from a named source, a link to where a claim comes from. That is what an unattended AI draft usually lacks. A model can produce fluent prose on any topic, but it cannot produce your test results, your prices, your customer data or a quote from your engineer unless you give them to it. Asked for numbers without being given any, it may invent ones that sound right.
Why does AI text that repeats the web give ChatGPT nothing to quote?
An assistant that searches builds one answer from a limited set of pages; the GEO paper's test engine used five per query. Each statement it attributes needs a page that supports it. If ten pages make the same general point, any of them will do, and the assistant has no reason to prefer yours. If your page is the only one with the specific figure, the method from someone who did the work, or the answer to a narrow question, it is the page that answer would need.
A language model writing with no input from you works from what is already widely published, so its first draft tends to restate the consensus in smooth sentences. The result is a page with nothing new in it. Google's guide to helpful, people-first content (last updated on 5 October 2026) asks the same question in its self-assessment: if the content draws on other sources, does it avoid simply copying or rewriting them and instead provide substantial additional value and originality? Its warning signs include "extensive automation to produce content on many topics" and "mainly summarizing what others have to say without adding much value".
Company blogs are not shut out of answers either. In our measurement of 21 October 2026, where we put questions about getting a business into AI answers to ChatGPT, Claude, Gemini and Perplexity, most cited sources were the blogs of GEO agencies and tool vendors themselves. A company blog is not the obstacle; an article with nothing of its own is.
How should you check an AI draft before publishing?
Treat the draft as raw material. Before it goes live, it should pass these checks:
- One thing nobody else has. Find the fact, figure, method, price or example that only your business can supply. If the draft has none, add it or do not publish.
- Every number traced. Open the source of each figure, date and quote, and link it where it is used. Delete anything you cannot trace. Google's checklist asks whether content has easily verified factual errors, and an AI draft can contain them.
- Primary sources linked. Link the law, the standard, the manufacturer's documentation or your own data, not another blog that summarised them. In the GEO study, citations, quotations and statistics were the methods that worked best.
- Restated paragraphs cut. Remove the paragraphs that say what every competing page says. What remains is the part an assistant could quote.
- The answer up front. Answer the question in the first two sentences, in the words a buyer would use.
- An honest author. Google recommends accurate bylines where readers expect them and calls fabricated creator profiles, such as AI-generated headshots or made-up names, a form of deception. It suggests AI disclosures where a reader might ask "How was this created?"
- A pace you can check. If you cannot verify ten articles a week, do not publish ten. Volume without added value is what the scaled content abuse policy describes.
- A page that can be reached. Make sure your robots.txt does not block OAI-SearchBot and other assistant crawlers; this guide to allowing GPTBot and other AI crawlers shows the lines to use.
Is "AI SEO" worth paying for?
That depends on what the articles contain, not on the label. Before you pay, ask the seller where each article's facts come from, who checks them before publication, and how they will measure whether assistants name or cite you. If the facts come from the model alone and nobody checks them, you risk buying the pattern Google's spam policy describes, whatever it is called.
Measuring means putting your buyers' questions to the assistants several times and counting the answers that name you; this guide to checking whether ChatGPT, Claude, Gemini and Perplexity name you explains how. For the field as a whole, see what generative engine optimization is.
To see where you start, the free check at ItGrows takes about 30 seconds: it generates five buyer questions from your homepage and shows two of ChatGPT's and Perplexity's answers right away, and all five once you create an account. No article guarantees that an assistant will name a business; a check tells you where you stand before you spend on articles.
Frequently Asked Questions
Does ChatGPT avoid citing pages written by AI?
OpenAI's documentation for site owners covers crawler access and sets no rule about how a page was written. What a site owner controls is whether ChatGPT's search can reach the page and whether the page holds something an answer would need. A page with original facts and linked sources has that, whoever typed it.
Will Google penalise a site for publishing AI-written articles?
Not for using AI as such: Google says appropriate use of AI or automation is not against its guidelines. Generating many pages without adding value for users is scaled content abuse under its spam policies, and that applies whether a model or a person produced them.
Should an article say that it was written with AI?
Google suggests AI or automation disclosures where a reader might reasonably ask how the content was created. It also recommends accurate bylines and says that giving AI an author byline is probably not the best way to make the role of AI clear to readers.
How many AI articles does a business need to publish to get cited?
There is no such number, and volume is not the lever. In the GEO study, what raised a page's visibility was what the page contained: added citations, quotations and statistics. One article with facts only you have gives an assistant more to quote than ten that restate the same general advice.