Will ChatGPT and Google quote content written with AI or for SEO?

Somewhere in your content folder there are now three kinds of article. The first was written to a keyword brief: a target phrase, a word count, headings that repeat the phrase and a paragraph or two under each. The second came out of ChatGPT or a similar tool, lightly edited, and reads smoothly without saying anything you would recognise as yours. The third was written by one of your partners after a difficult job, explaining what went wrong and what they would do differently. The first two are most of what many firms publish. The third is what a client forwards to a colleague.

Will ChatGPT and Google quote the first two? Keyword engineering does not help, and it can hurt. In a Princeton and IIT Delhi study presented at KDD 2024, keyword stuffing lowered a source's visibility in AI-generated answers by about 8%, while adding quotations from relevant people raised it by about 41% and adding statistics by about 34%. The test used a research engine, so treat the numbers as directional, but the direction is clear.

AI drafting is a different question. Google does not penalise a page for being written with AI. It penalises pages made mainly to manipulate rankings, however they were produced, and a generic draft gives an assistant nothing it cannot find on a hundred other pages. Below: what Google has said, how assistants choose what to quote, the three kinds of page side by side, and what to do with old keyword pages.

#Can Google or ChatGPT tell if our content was written by AI?

Google says it does not need to. Its guidance on AI-generated content (February 2023) describes a "focus on the quality of content, rather than how content is produced", and says: "Using AI doesn't give content any special gains. It's just content." Unhelpful pages are demoted whether a person or a model wrote them.

OpenAI, Anthropic and Perplexity have not said their assistants detect or penalise AI-written pages, and we know of no published study showing that they do. What the evidence suggests they reward is what an unedited AI draft usually lacks: quotations, figures, named sources and first-hand detail. Google's guide to its generative AI features (updated July 2026) calls the alternative commodity content, which "is often based on common knowledge, which could originate from anyone, and typically adds little unique insight for readers". An unedited draft is commodity content almost by definition, because it is assembled from what is already published.

#Will Google penalise AI-generated content on our website?

Not for being AI-generated. Google acts against pages made mainly to manipulate rankings. In March 2024 it announced a scaled content abuse policy that applies "no matter whether content is produced through automation, human efforts, or some combination of human and automated processes", and said sites that break its spam policies "may rank lower in results or not appear in results at all". Its current spam policies (updated August 2026) list "Using generative AI tools or other similar tools to generate many pages without adding value for users" as an example. The AI features guide adds that creating separate content for every possible variation of a query breaks the same policy.

Should you say you used AI? Google says disclosures "are useful for content where someone might think 'How was this created?'" For an article drafted from an interview and checked by the named expert, a line saying so is usually enough.

#Does keyword stuffing still work for Google in 2026?

No. Google's spam policies define keyword stuffing as "filling a web page with keywords or numbers in an attempt to manipulate rankings in Google Search results", and it is one of the named practices that can push a page down or out.

The old idea that Google only counts words is years out of date. Google's guide to its ranking systems describes BERT as an AI system that lets it "understand how combinations of words express different meanings and intent", alongside RankBrain and neural matching, which relate words to concepts rather than matching exact strings. Google's AI features guide says plainly that "AI systems can understand synonyms and general meanings". Repeating a phrase adds nothing a language model needs, and it makes the page worse to read.

#How do ChatGPT and Google's AI answers choose which pages to quote?

They search first, then choose. ChatGPT search, Copilot and Google's AI features start from a search index, retrieve a set of pages, and lift the passages that answer the question best. So ranking and substance both count.

Ranking helps: AirOps found in April 2026 that pages ranking first were cited by ChatGPT 58.4% of the time, against 14.2% for pages ranking tenth. It does not guarantee a mention: Ahrefs found in March 2026 that only 38% of AI Overview citations came from top-ten pages, down from 76% in July 2025, because Google now runs several related searches for one question (query fan-out). The Princeton and IIT Delhi study also found that lower-ranked pages gained most from changes such as citing sources.

A page that ranks but says nothing specific can lose the citation to one that explains the mechanism. Our guide on whether you still need SEO now that buyers ask ChatGPT covers the wider picture.

#What do the three kinds of page look like side by side?

For example, say a mid-sized accountancy practice wants a page on helping clients move to a new payroll provider.

Version one, written to a keyword brief: "Payroll migration is a key challenge for businesses considering payroll migration. Our payroll migration service supports payroll migration for companies of all sizes. A successful payroll migration requires payroll migration planning. Contact us for payroll migration support." The phrase appears seven times and nothing is explained.

Version two, an unedited AI draft (our own illustration of the pattern, not a real output): "Switching payroll providers can be a complex process, but with the right planning it can be smooth and stress-free. First, assess your current payroll needs. Next, research potential providers and compare their features and pricing. Communication is key: keep employees informed throughout the transition. Finally, test the new system thoroughly before going live." Every sentence is reasonable. None could only have been written by this firm, and the advice matches what many other pages already say.

Version three, written from experience: "Most payroll moves go wrong for the same reason: the old provider's data is exported before anyone has checked what is missing. Pension opt-outs, salary sacrifice arrangements and year-to-date figures for people who joined mid-year are the usual gaps, and each takes a week to reconstruct once the old system is switched off. We now insist on a parallel run for one full pay period, which clients resist because it costs a month, and which has prevented every serious error we have seen since we introduced it. If your provider will not offer a parallel run, that is the first sign to look elsewhere."

An AI tool can help produce version three. Record the practitioner explaining the parallel run, have the tool draft from the transcript, and have the practitioner check it. The substance still comes from the person who did the job.

#What gives keyword-led and generic AI writing away?

In our experience, a handful of patterns. None is a rule any assistant has published; they are simply what thin text looks like, and they leave a reader or a model nothing to use.

• Repetition nobody would say aloud. The same phrase five times in four sentences.
• Sentences that contain the phrase and say nothing. "A data strategy is important for firms implementing a data strategy."
• Even, shallow depth. Five subsections of the same length. Real explanation is uneven: the crucial point gets six paragraphs and the minor one a line.
• No reasons. It says what to do, never why, or why the obvious approach fails.
• No exceptions. Nobody who has done a job for twenty years thinks their method always works.
• Steps that fit any firm. Generic AI drafts lean on "first", "next" and "finally" where reasons should be.

Why ChatGPT ignores sites with hundreds of pages shows what happens when a whole site reads this way, and what makes content read as expert rather than marketing covers the signals to aim for.

#How do we write for ChatGPT without losing our Google rankings?

Write one page for both. Google's AI features guide says optimising for its generative AI features is still SEO, and "You don't need to write in a specific way just for generative AI search." The habits that get a page quoted are the ones Google's quality guidance already asks for.

• Build each page around a buyer's question, using the buyer's words once in the title and the first paragraph.
• Put the direct answer in the first paragraph and the key figure near the top.
• Keep pages that rank. Add the missing specifics, examples and limits rather than rewriting from scratch.
• Cover the follow-up questions, because Google's AI features run related searches for each question.

#Should we delete old SEO blog posts written around keywords?

Not all at once. Check each page's traffic, its links and whether it sits on a question buyers ask, then keep, rewrite, merge or delete.

We know of no published evidence that ChatGPT or Perplexity score a whole site when choosing a citation, so the case for pruning is practical: near-duplicates compete with each other, and a buyer who lands on a thin page judges the firm by it. How to audit existing content for AI assistants sets out the scoring method.

#How do we shift a content programme without losing what works?

Keep the pages that still earn traffic and change how new work is produced.

#Step 1: Choose topics by difficulty, not volume

Ask what buyers struggle to understand, where the received wisdom is wrong, and what you know that is not written down anywhere. These questions may have low search volume, and they are often the ones with no good source yet.

#Step 2: Put a practitioner in the room

A forty-minute recorded conversation with the person who has done the job, and a writer, with or without an AI tool, who turns it into prose without sanding off the specifics.

#Step 3: Write to the length the subject needs

Hard subjects need the mechanism, the exceptions and an example; narrow questions need a short, exact answer. Whether ChatGPT prefers long or short articles sets out what the studies say.

#Step 4: Put a name on it

A byline such as "by the head of payroll, sixteen years in practice" beats "by the team", and the same name and role should appear on LinkedIn.

#Step 5: Bring a number when you have one

A finding from your own records is something an assistant can only attribute to you; see whether publishing your own research helps.

#Step 6: Link the pieces to each other

Link related pieces so readers and crawlers can follow the subject; organising content so AI assistants see you as the expert explains how.

#What to do this month

1. Read your ten most visited articles aloud for thirty seconds each and mark any that repeat a phrase or say nothing a competitor could not.
2. Write down your rules for AI drafting: which steps a tool may do, who adds the substance and who checks every figure.
3. Pick the three questions your clients struggle with most and check whether any page answers them with reasons and exceptions.
4. Record a forty-minute conversation with one practitioner on one of those questions and publish it as one thorough piece under their name.
5. Change the brief you give your writer or agency: no target phrase, no word count, one question to answer completely and an expert to interview.

#Frequently asked questions

#Will writing this way hurt our Google rankings?

In our experience it helps them, because pages people read, share and link to tend to be the expert ones. Leave old keyword pages that still bring traffic in place and let the new work overtake them.

#How long before we see a difference?

Assistants that search the web live, such as ChatGPT search, Perplexity, Copilot and Google's AI features, can reflect a new page within weeks. Answers drawn from a model's own memory change only when it is retrained, which takes months. ID-Pal's case study gives a dated example: "UK buyer questions where AI names ID-Pal: none in May, two in August, four in September 2026." See the ID-Pal case study.

#Do we need a famous author?

No, a credible one. A name, a role, a sentence of background and a body of work on one subject is enough. What matters is that the same identifiable person keeps explaining the same field well.

#Can an outside writer produce expert content?

Yes, provided a real expert is the source. The writer interviews, structures and makes it readable; the expert supplies the mechanism, the examples and the exceptions. How to outsource content without losing your voice sets out the briefing method.

#We are in a commodity sector. What is there to be expert about?

The use of the product rather than the product itself: where implementations go wrong, what the hidden costs are and which situations suit which option. A freight forwarder cannot be a thought leader on containers, but can be on why customs delays happen and how to avoid them.

If you want a second opinion on how your content reads to buyers and assistants, book a free Growth Review: 30 minutes with David Adams and three fixes in writing. Turning your partners' expertise into content that gets quoted is part of Reputation That Wins.

#Sources