AI content in search: detection, ranking and ethical use

Google does not penalise AI content. It penalises content made to game the rankings

The question I get asked most about AI and search is some version of "will Google punish me if I use ChatGPT?" The answer has been public since 8 February 2023, when Google's Search Quality team published its guidance on AI generated content. The key sentence is short: "Appropriate use of AI or automation is not against our guidelines."¹ The same post reminds people that automation has produced useful pages for years, sports scores, weather forecasts, transcripts, and that "not all use of automation, including AI generation, is spam."¹

What Google does act against is content "generated for the primary purpose of manipulating Search rankings and not helping users."¹ That has been a spam rule since long before anyone had heard of a large language model. The tool is not the problem. The intent, and the result on the page, is.

Two more lines from that post are worth keeping. "Our focus on the quality of content, rather than how content is produced, is a useful guide."¹ And, drier still: "Using AI doesn't give content any special gains. It's just content."¹ That last one cuts both ways. No bonus for using AI, no bonus for abstaining. There is only the page, and whether a reader is better off for having landed on it.

So the honest question is not "can AI content rank?" but "is this page worth ranking?" The rest of this article is about how Google decides that, why the AI detectors people buy do not work, and how I use these tools myself without ending up on the wrong side of the line.

Worked example: two ways to write 300 product descriptions

An invented bike shop in Galway, Salthill Cycles, has 300 bikes online and no descriptions beyond the manufacturer's spec sheet. The owner, Dónal, tries two approaches. Nobody real, every figure made up.

The first approach takes an afternoon. He pastes each spec sheet into a chat tool with the prompt "write a 150 word product description" and publishes the output. He now has 300 pages that say, in slightly different words, exactly what the spec sheet already said. A reader learns nothing they could not have read on the manufacturer's own site, which Google already indexes.

The second approach takes a week, on and off. Dónal's mechanic, Priya, has ridden or serviced most of these bikes, a good few of them along the prom in a westerly. For each one she dictates two or three sentences into her phone: who the bike suits, what she would change on day one, which size runs small. The same chat tool tidies the grammar and fits the sentences to a template. The spec sheet sits underneath as a table.

Same tool in both cases. In the first, the tool produced the content and a human pressed publish. In the second, a human produced the content and the tool did the typing. The first set is what Google's spam policy calls scaled content, the second is a shop telling you what it knows. Nothing in Google's systems needs to detect the tool to tell them apart. The pages give the game away.

There is no AI detector inside Google Search, and the ones you can buy do not work

People imagine Google runs a classifier over every page, stamps it "AI" or "human", and ranks accordingly. Nothing Google has published says that, and the documents it gives its own quality raters point the other way. The current Search Quality Rater Guidelines, dated 11 September 2025, tell raters that pages made at scale with no original content "should be rated Lowest, no matter how they are created," and add: "Even if you are unsure of the method of creation, e.g. whether or not the page is created using generative AI tools, you should still use the Lowest rating when you strongly suspect scaled content abuse."² The rater is told not to bother working out the method. They are told to look at effort, originality and added value.

The same document is explicit that the tool itself is neutral: "the use of Generative AI tools alone does not determine the level of effort or Page Quality rating. Generative AI tools may be used for high quality and low quality content creation."² That is the clearest statement I know of Google's actual position, written for the people who audit its results.

Now to the detectors you can buy. The evidence here is not thin, it is damning, and it comes from the people with the strongest incentive to make detection work.

OpenAI launched its own AI text classifier on 31 January 2023 and said in the launch note that it was "not fully reliable": on its test set it "correctly identifies 26% of AI written text (true positives) as 'likely AI written,' while incorrectly labeling human written text as AI written 9% of the time (false positives)."³ On 20 July 2023 the company added a line at the top of the same page: "the AI classifier is no longer available due to its low rate of accuracy."³ The organisation that built the model could not reliably recognise the model's output. Fair play to them for saying so in public.

A peer reviewed study published the same year tested 14 detection tools, 12 public ones and two commercial systems used by universities. The authors conclude that the tools "are neither accurate nor reliable and have a main bias towards classifying the output as human written rather than detecting AI generated text."⁴ Not one tool reached 80% accuracy. Machine generated text that a human had lightly edited was caught about 42% of the time, against 74% for untouched machine output. And human writing that had been machine translated into English lost about 20 points of accuracy, because translation "leaves some traces of AI in the output, even if the original was purely human written."⁴

The second study is the one I take personally. Researchers at Stanford ran seven widely used detectors over 91 English essays written by people whose first language was not English. On average 61.3% of those human essays were flagged as AI generated. At least one detector flagged 97.8% of them. Essays by American schoolchildren were classified almost perfectly.⁵ I am an Austrian who writes in English for a living. By that standard, most of what I have ever published is a robot.

Put the three together. Detectors miss most machine text, flag a worrying share of human text, and are fooled by a light edit or a non native turn of phrase. A ranking signal built on one would be wrong most of the time and biased against half the planet. Google does not use one, and says so.

Worked example: what a detector flag actually tells you

Suppose Dónal at Salthill Cycles runs a detector over his 300 pages. He knows that about 10% of them, 30 pages, were left as raw machine output after the first afternoon. The detector has the published OpenAI numbers: it catches 26% of AI text and wrongly flags 9% of human text. A page comes back flagged. How likely is it that the page is actually machine written?

This is a conditional probability, and Bayes' rule answers it:

$$P(\text{AI} \mid \text{flag}) = \frac{P(\text{flag} \mid \text{AI})\,P(\text{AI})}{P(\text{flag} \mid \text{AI})\,P(\text{AI}) + P(\text{flag} \mid \text{human})\,P(\text{human})}$$

Here $P(\text{AI} \mid \text{flag})$ is what Dónal wants, the probability a flagged page is machine written. $P(\text{flag} \mid \text{AI})$ is the detector's true positive rate, $0.26$. $P(\text{AI})$ is the share of pages that really are machine written, $0.10$. $P(\text{flag} \mid \text{human})$ is the false positive rate, $0.09$, and $P(\text{human})$ is the share of human pages, $0.90$. The rule is explained from scratch in Foundations: probability and Bayes, and the trap it guards against is in Foundations: indexing and base rates.

The arithmetic, step by step. The numerator is $0.26 \times 0.10 = 0.026$. The second term in the denominator is $0.09 \times 0.90 = 0.081$. The denominator is $0.026 + 0.081 = 0.107$. So:

$$P(\text{AI} \mid \text{flag}) = \frac{0.026}{0.107} \approx 0.243$$

Here $0.026$ is the share of all pages that are machine written and flagged, $0.107$ is the share of all pages that are flagged for any reason, and $P(\text{AI} \mid \text{flag})$ is the ratio of the two, about $0.243$. A flagged page has about a 24% chance of being machine written. Three flagged pages out of four are Priya's own words. Dónal would be deleting his mechanic's knowledge to satisfy a tool that is wrong most of the time. The base rate does the damage: when most pages are human, even a modest false positive rate swamps the true catches. Google's raters are told to ignore the method and read the page. That is the right instruction for Dónal too. If you want to see the working again slowly, it is in Foundations: probability and Bayes.

What Google ranks instead: helpful, people first content, and E E A T

If not the tool, then what? Google's own framing since 2022 has been "people first" content, which its documentation defines as content "created primarily for people, and not to manipulate search engine rankings."⁶ In August 2022 it launched a dedicated helpful content system to reward that. Then, in March 2024, that system stopped being a separate thing. Google's page on it now says that in March 2024 "it evolved and became part of" the core ranking systems.⁷ The announcement put it like this: "There's no longer one signal or system used to do this."⁸ Helpfulness became a property of ranking as a whole, not a filter bolted on the side.

The same announcement gave the one number in this story I find useful. Google expected the March 2024 update, with earlier work, to "reduce low quality, unoriginal content in search results by 40%."⁹ When the rollout finished in April it reported 45%.⁹ That is Google grading its own homework, so treat it as intent rather than audit. But the direction is not in doubt. Nearly half of the thin pages that used to rank were pushed down in one spring, and nobody needed to detect which of them a machine wrote.

Underneath "helpful" sits the framework Google uses to describe quality: experience, expertise, authoritativeness and trustworthiness, written as E E A T. The first E was added in December 2022, when Google said it wanted content that shows "first hand" involvement, having used the product, visited the place, lived the thing.¹⁰ Of the four, Google is blunt about which matters: "Of these aspects, trust is most important."⁶ The rater guidelines say the same with more colour: "Trust is the most important member of the E E A T family because untrustworthy pages have low E E A T no matter how Experienced, Expert, or Authoritative they may seem."²

Two clarifications save a lot of confusion. E E A T is not a ranking factor you can switch on. Google describes it as a mix of signals its systems use to identify good content, not a single score.⁶ And the raters who grade pages against it do not touch your rankings. "Search raters have no control over how pages rank," the documentation says, and "Rater data is not used directly in our ranking algorithms."⁶ The raters audit the system. The system ranks the pages.

In practice this is almost embarrassingly simple. The pages that do well after 2024 are the ones where someone who knows the subject said something a reader could not get elsewhere, and the page makes it easy to see who that someone is. A machine can help that person type faster. It cannot supply the experience.

An example: Google publishes a list of questions to ask about your own content, and two of them do most of the work. "Does the content provide original information, reporting, research, or analysis?" and "Are you mainly summarizing what others have to say without adding much value?"⁶ Dónal's first set of 300 descriptions answers no to the first and yes to the second. Priya's set answers the other way round. No detector involved. Just two questions and an honest afternoon.

The two spam policies that actually bite: scaled content abuse and site reputation abuse

Google does maintain hard rules, and two of them matter here. Both were introduced or rewritten in March 2024, and both are written around what the pages do, not how they were made.

The first is scaled content abuse. The definition: "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." It is "typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created."¹¹ The first example on the list names the tool directly: "Using generative AI tools or other similar tools to generate many pages without adding value for users."¹¹ The other examples are older sins, scraping feeds, stitching content from other pages, creating many pages "where the content makes little or no sense to a reader but contains search keywords."¹¹ Google said the policy "builds on" its earlier rule against automatically generated content.⁸ Same rule, wider net, now covering humans who churn out thin pages by hand as well.

The second is site reputation abuse. This is the practice of publishing third party content on a respected site so that it borrows the host's ranking signals, the coupon pages and casino reviews that appeared on newspaper and university domains. Google's policy applies where third party content is published on a host site "mainly because of that host's already established ranking signals, which it has earned primarily from its first party content."¹¹ It took effect on 5 May 2024.⁸ In November 2024 Google closed the obvious loophole and said a violation stands "regardless of whether there is first party involvement or oversight of the content," adding that "no amount of first party involvement alters the fundamental third party nature of the content."¹²

Notice what is absent. Neither policy says "AI content is spam." Both say "pages made to borrow or manipulate a ranking signal are spam." If you generate 400 pages because you have 400 genuinely different things to say, each with its own data, that is programmatic publishing and it is grand. I wrote about doing it properly in Programmatic SEO and content automation. If you generate 400 pages because you want to rank for 400 town names, that is scaled content abuse whether a person or a model typed them.

An example: an invented lettings agency in Cork, Douglas Lettings, pays a content agency for "apartments to rent in [area]" pages for 60 Cork neighbourhoods, each the same three paragraphs with the area name swapped in. That is scaled content abuse. The same agency then pays the local newspaper €400 a month to host "the best letting agents in Cork," written by the agency, in the newspaper's layout. That is site reputation abuse, and since November 2024 an editor glancing at the copy does not fix it. Both were sold to the agency as SEO. Both are now the fastest way to lose the rankings it had.

Worked example: a Limerick software company, one scaled set of pages against one researched guide

Here is the whole argument in one invented company. Shannon Routing is a 12 person software business in Limerick that sells route planning software to small delivery fleets. The founder, Niamh, wants organic traffic for "delivery route planning software." Every number below is made up to show the mechanics. Nobody real, no real client.

In January she tries the scaled route first, because it is quick. A contractor uses a chat tool to produce 400 pages of the form "delivery route planning software in [town]" for 400 towns across Ireland, every place big enough to have a main street. Each page is 600 words that could apply to any town. The whole job costs €1,200 in tooling and a day of the contractor's time.

In the same month she also writes one guide: "How many stops can one van actually do in a day? What 2,000 real routes told us." It uses anonymised data from her own customers, three customer interviews, a drag and drop calculator, and a sources list. She and a colleague spend three days on it, and I spend one. Call it €3,600 of time.

Here is how the two performed over the year, in monthly organic visits. Invented, but the shape is one I have watched more than once.

The horizontal axis is the month, January to December. The vertical axis is organic visits from search in that month, for the 400 town pages added together and for the single guide on its own. The scaled set wins the first quarter: 1,800 visits in January, 2,400 in March, mostly from towns where nobody else had written a page. Then a core update lands in April and the set loses two thirds of its traffic by May. By December it is down to 150 visits across all 400 pages. The guide starts at 40 visits, because nobody has found it yet, and climbs every month as it earns links and as Google learns that people stay on it. It crosses the scaled set in June and ends the year at 1,300 a month.

Add up the year. The scaled pages collected 10,060 visits. The guide collected 8,230. On raw traffic the scaled set still wins, narrowly, and if you stop reading here you will draw the wrong conclusion. So do not stop.

What Niamh sells is a software trial, and the two kinds of visitor behave very differently. A visitor who lands on a generic town page, sees nothing specific, and leaves, rarely signs up. In this example 0.2% of them do. A visitor who has just spent six minutes with a calculator built from real route data is a different person. In this example 2.5% of them sign up. Leads are visits times the sign up rate:

$$L = V \times r$$

$L$ is the number of trial sign ups, $V$ is the visits over the year, and $r$ is the sign up rate, the share of visitors who start a trial. This is a plain rate, the same shape as a conversion rate, and it is explained in Foundations: ratios and rates.

For the scaled pages: $L = 10{,}060 \times 0.002 = 20.1$, so about 20 trials. For the guide: $L = 8{,}230 \times 0.025 = 205.75$, so about 206 trials. Cost per trial is the cost divided by the trials. The scaled pages cost €1,200 for 20 trials, which is €60 a trial. The guide cost €3,600 for 206 trials, which is €17.48 a trial. The researched page produced ten times the leads at under a third of the cost per lead, and the scaled pages are still falling in January while the guide is still rising.

Three things keep this example honest. The scaled set really does win early, which is why people keep doing it. The crossover takes months, longer than most budgets are patient for. And the guide's advantage comes from the data and the interviews, not the prose. A chat tool helped Niamh structure the guide and tidy the transcripts. It could not have produced the 2,000 routes or the three customers.

How I use AI for content, and where I stop

This is the ethics section, and I would rather describe what I actually do than preach. I use AI tools for the simple, repetitive parts of writing, and I finish every piece with knowledge the tool does not have. The same principle runs through the rest of my work, including the code, which I wrote about in AI writes the code. Who is responsible for it?. The tool is fast and tireless and knows nothing about my clients. I am slow and tire easily and know a great deal about them. The process is built to use each for what it is good at.

The diagram runs top to bottom through six steps. The brief is always mine: who is reading, what question they arrived with, what the one point is. In research the tool is a useful assistant, fetching documents and pulling the passages, telling me what a 160 page PDF says about generative AI on page 112. But every source it finds, I open. Third, the tool drafts: scaffolding, a structure from a pile of notes, the sentence that explains a conversion rate for the fortieth time. Fourth, the fact check, the decision point: every claim is read against its source and anything I cannot trace is cut, not softened. Fifth, the rewrite, where the article becomes mine. The examples are mine, the shop in Galway and the fleet in Limerick, and so are the judgements and the jokes, such as they are. Sixth, publication with footnotes, so a reader can check my work the way I checked the tool's.

The order matters. If the rewrite comes before the fact check, the errors get polished rather than removed. If the brief comes last, the tool decides what the article is about, and it will pick the average of everything ever written on the subject, which is precisely the content Google spent 2024 pushing down.

On disclosure, I follow Google's own suggestion. Its guidance says to answer the question a reader would ask: "Who wrote this?" with a real name, and "How was this created?" with a plain explanation where a reader would reasonably want one.¹ It also says that "Giving AI an author byline is probably not the best way" to disclose it.¹ So my name is on my articles because I am responsible for every claim in them, and this section is the explanation of how they are made. Google's documentation also warns against "fabricating creator profiles," which it calls deception.⁶ Inventing an expert to sign a machine's work is the one thing in this whole area I would call unethical without hedging.

Where I stop is easy to state. I do not publish a page whose only reason to exist is a keyword. I do not let a tool state a figure I have not seen in its source. I do not use AI to pretend to an experience I have not had, because experience is the first E and the one a model cannot fake. And I do not flood. Google says it pushed down nearly half the unoriginal content in its results in 2024, and every signal it publishes points at effort, originality and trust. Flooding is not just ethically thin. It is a bet against the one direction the ranking systems have moved in for three years.

One drift I will resist: how to appear inside AI answers, in ChatGPT or Google's own overviews, is a different question with its own article, How to rank in ChatGPT and AI search engines. This one is about the ordinary results page.

Where this leaves you

Three sentences to pin above the desk. Google judges the page, not the typist, and says so in writing. The detectors do not work, not even OpenAI's, so neither you nor Google should make decisions with one. And the content that ranks after 2024 is the content where someone who knows the subject said something new and signed it.

Use the tools. I do, daily, for the parts of writing that do not need me. Then do the part that does.

Footnotes

1. Danny Sullivan and Chris Nelson, Google Search's guidance about AI generated content, Google Search Central Blog, 8 February 2023, https://developers.google.com/search/blog/2023/02/google-search-and-ai-content

2. Google, Search Quality Rater Guidelines: General Guidelines, 11 September 2025 edition, sections 3.4, 4.6.5 and 4.6.6, https://static.googleusercontent.com/media/guidelines.raterhub.com/en//searchqualityevaluatorguidelines.pdf

3. Jan Hendrik Kirchner, Lama Ahmad, Scott Aaronson and Jan Leike, New AI classifier for indicating AI written text, OpenAI, 31 January 2023, with the 20 July 2023 retirement note, archived copy, https://web.archive.org/web/20230831225148/https://openai.com/blog/new-ai-classifier-for-indicating-ai-written-text

4. Debora Weber-Wulff, Alla Anohina-Naumeca, Sonja Bjelobaba, Tomáš Foltýnek, Jean Guerrero-Dib, Olumide Popoola, Petr Šigut and Lorna Waddington, Testing of detection tools for AI generated text, International Journal for Educational Integrity, volume 19, article 26, 2023, https://arxiv.org/abs/2306.15666

5. Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, GPT detectors are biased against non native English writers, Patterns, volume 4, issue 7, 2023, https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/

6. Google, Creating helpful, reliable, people first content, Google Search Central documentation, https://developers.google.com/search/docs/fundamentals/creating-helpful-content

7. Google, Google Search's helpful content system, Google Search Central documentation, https://developers.google.com/search/docs/appearance/helpful-content-system

8. Google Search Central, What web creators should know about our March 2024 core update and new spam policies, 5 March 2024, https://developers.google.com/search/blog/2024/03/core-update-spam-policies

9. Elizabeth Tucker, New ways we're tackling spammy, low quality content on Search, The Keyword, Google, 5 March 2024, updated 26 April 2024, https://blog.google/products/search/google-search-update-march-2024/

10. Elizabeth Tucker, Our latest update to the quality rater guidelines: E E A T gets an extra E for Experience, Google Search Central Blog, 15 December 2022, https://developers.google.com/search/blog/2022/12/google-raters-guidelines-e-e-a-t

11. Google, Spam policies for Google web search, Google Search Central documentation, https://developers.google.com/search/docs/essentials/spam-policies

12. Google Search Central, Evolving the site reputation abuse policy, 19 November 2024, https://developers.google.com/search/blog/2024/11/site-reputation-abuse