Chapter 19: Producing Content with AI

Last updated: November 1, 2026.

With a language model, you can now write hundreds of pages in a day. Some teams see this as a huge opportunity: a page for every city, a page for every product variation, a page for every question. It’s also the most debated use of AI in SEO.

To see where that opportunity is real and where it’s a trap, we first need to look at how Google views AI-generated content and at its Scaled Content Abuse policy. After that I’ll cover how the policy has been enforced since 2024, what the data shows, where human input needs to go, and a quality control process. Since this is the last chapter of the book, there’s a short closing at the end too.

My short answer: the problem isn’t AI itself, it’s scale and intent. A good page written with AI isn’t a problem. Thousands of empty pages written by people are.

1. Google’s position

Google’s view of AI content has been consistent since 2023. The title of its February 2023 statement sums it up: rewarding high-quality content, however it is produced (Google Search Central). According to the same statement, using automation, including AI, to generate content with the primary purpose of manipulating search rankings is a violation of its spam policies. But not every use of automation is spam; Google gives examples like sports scores, weather forecasts and transcripts.

Google’s current helpful content guidance says the same thing and treats “why” the content was created as the most important question. Was it created mainly to help people, or to attract visits from search engines? If automation or AI is used to produce content whose main purpose is manipulating search rankings, that’s a violation of the spam policies (Google Search Central).

Google’s separate guidance on AI-generated content takes the same line. It says generative AI can be particularly useful when researching a topic and adding structure to original content. But using it to generate many pages without adding value for users may violate the scaled content abuse policy (Google Search Central).

Read the three documents side by side and the picture doesn’t change. Google isn’t asking whether content was written with AI. It’s asking who it was written for.

2. What is Scaled Content Abuse?

On March 5, 2024, Google announced three new spam policies: expired domain abuse, site reputation abuse and scaled content abuse (Google Search Central).

Scaled content abuse is generating many pages whose main purpose is manipulating search rankings rather than helping users. Google writes that this practice is typically focused on creating large amounts of unoriginal content that provides little or no value to users, no matter how it’s created.

The old policy was about “automatically generated content”; the new one doesn’t care about the production method. According to Google’s announcement, the policy can apply whether content is produced through automation, human effort or a mix of the two. Having a hundred freelance writers produce a hundred city pages from the same template can fall under it too.

Google’s spam policies page gives these examples of scaled content abuse (Google Search Central):

• Using generative AI or similar tools to generate many pages without adding value for users.
• Scraping feeds, search results or other content to generate many pages, including through automated transformations like synonymizing or translating.
• Stitching or combining content from different pages without adding value.
• Creating multiple sites with the intent of hiding the scaled nature of the content.
• Creating many pages that make little or no sense to a reader but contain search keywords.

One more item from Google’s AI guidance belongs on that list. As we discussed in Chapter 3, Google states plainly that creating separate content for every variation in the query fan-out violates this policy if the main purpose is to manipulate rankings or AI answers (Google Search Central). Mass page production done in the name of “content for AI” is judged within the same framework.

3. Enforcement from 2024 to 2026

On the day it announced the policy, Google also started the March 2024 core update. The update finished on April 19, 2024, and Google said that with it we’d see 45% less low-quality, unoriginal content in search results (Search Engine Land).

Since then, these are the spam updates announced on the Search Status Dashboard (Google Search Status Dashboard):

Google doesn’t explain in detail what each of these updates targeted, so I can’t say for sure how much any of them targeted scaled content. But 2026 is the year with the most announced spam updates since 2022.

There’s movement on manual actions as well. In October 2026, Search Engine Roundtable reported that some SEOs had seen an increase in manual action notices in recent weeks, with the reported types including thin content, pure spam and scaled content abuse. The same report notes that Google hadn’t commented and that the increase wasn’t confirmed by data (Search Engine Roundtable). So I read it as a signal to watch, not as a trend.

If your site gets a manual action, Google notifies you through Search Console, and once you’ve fixed the problem you can request a reconsideration. Google’s March 2024 announcement also reminds site owners that sites violating spam policies may rank lower or not appear in results at all (Google Search Central). Cleaning up scaled content is hard work. You have to evaluate hundreds or thousands of pages one by one, remove or deindex the ones with no value, and strengthen the rest. The cheapest fix is never to produce those pages in the first place.

Then there are the Search Quality Rater Guidelines. Google’s AI content guidance suggests looking at them to see how scaled content abuse (section 4.6.5) and main content created with little effort, little originality and little added value (section 4.6.6) are evaluated. In the same place, Google reminds readers that these ratings don’t directly affect rankings and are used to measure how well its ranking systems perform (Google Search Central). Still, I think it’s reasonable to assume that what raters are taught to mark as lowest quality is also what the systems are trying to identify.

4. What the data says

Policy is one thing; what shows up in the results is another. Two studies help here.

In July 2025, Ahrefs analyzed about 600,000 pages ranking in the top 20 for 100,000 random keywords using its own AI content detector (Ahrefs). 4.6% of the pages were classified as pure AI, 13.5% as pure human and 81.9% as mixed. 86.5% of the top-ranking pages contained some amount of AI content. The correlation between the share of AI content and ranking was 0.011, which is practically zero. On the other hand, pages written entirely by AI rarely reach the number one spot; the very top positions have slightly less AI content.

According to a Graphite study published in Axios in October 2025, among 65,000 random URLs from Common Crawl, the share of AI-generated articles briefly overtook human-written ones in November 2024 and the two have been roughly equal since. A second report from the same group found that 86% of articles ranking in Google were written by humans, and 82% of the articles cited by ChatGPT and Perplexity (Axios).

Both studies share the same weakness. AI content detectors aren’t precise, and Ahrefs and Graphite both say so themselves. The numbers should be read as directional, not as exact measurements.

What I take from these two studies is that using AI isn’t penalized on its own, but content left entirely to AI doesn’t stand out either. The top positions and the sources in AI answers come from content with a clear human contribution.

5. The commodity content problem

Besides the spam risk, mass content production has a quieter problem. What it produces is commodity content.

Google’s AI guidance describes this as the difference between “commodity content” and “non-commodity content.” Commodity content is based on general knowledge anyone could have and adds little that’s new for the reader; Google’s example is a piece like “7 tips for first-time home buyers.” Non-commodity content draws on expertise or experience and goes beyond general knowledge. The guidance says plainly not to reproduce what others have already said or what a generative AI model could easily produce (Google Search Central).

I think that’s the key sentence of this chapter. Imagine publishing a page an AI model could write in a few seconds. When a user asks an AI the same question, the model can already write that answer itself; it has no need to read your page, cite it or recommend it. Content an AI could write on its own gives an AI no reason to cite you.

Compare two titles from the filling-machine example. “7 things to consider when choosing a filling machine” is commodity content; an AI can write it in seconds from general knowledge everyone has. “We compared a bottom-up nozzle with top filling on foaming liquid soap: bottles per minute and overflow rate” isn’t, because it contains numbers only someone who ran the test could know. The title is hypothetical, but the difference is clear. The second page tells the AI something it couldn’t know on its own, and that’s why it’s more likely to be cited.

The pages that get cited and recommended in AI search are the ones that say something the model couldn’t know on its own: original data, real tests, case studies, expert opinion, current prices, concrete capacity figures. We covered this kind of content in detail in Chapters 11 and 12.

6. Where human input belongs

“Let AI write it and a human check it” is a common line, but it’s incomplete. Checking means finding the mistakes in a text, and making a commodity text error-free doesn’t make it any less of a commodity. Human input has to go into the content itself, before the checking stage.

The “Who, How, Why” framework in Google’s helpful content guidance is a good guide here (Google Search Central).

“Who” asks who wrote the content and whether readers can see that. Google strongly encourages accurate author information, but it also warns against deceptive bylines: making it look as if an expert wrote the content with AI-generated profile photos, made-up names or fake credentials is a form of deception and a sign of a low-quality page.

“How” asks how the content was produced. In product reviews, showing how many products were tested and how the tests were done builds trust. The guidance says that if automation or AI substantially generated the content, it can help to disclose that to readers and explain why it was considered useful.

“Why” is, according to the guidance, the most important question. Was the content created mainly to help people?

In practice, I see human input belonging in four places:

1. Choosing the topic. A person should decide which page answers a real need, and that decision should come from the prompt list, the sales team and customer questions.
2. Original information. Test results, numbers, case studies, photos, expert opinion; everything the model can’t know.
3. Judgment. Giving a clear answer to a question like “Is this machine suitable for foaming products?” and taking responsibility for it.
4. The final check. Accuracy, freshness and fit with the brand’s voice.

AI is a valuable helper between these steps: building structure, drafting, editing and checking consistency. But it doesn’t replace any of the four.

A practical way to get expert input is to ask the expert questions instead of handing them a blank page. In the filling-machine example, you could do a half-hour interview with a production engineer and ask: What’s the most common mistake with foaming products? How do customers get capacity wrong? When would you not recommend this machine? Give the recording or notes to AI to produce a draft, then have the same expert read the draft. The expert’s time goes into passing on what they know and doing the final check, not into writing. And the result is a page the model couldn’t have written on its own.

7. Which pages should be published?

With mass content, the real decision is the publishing decision, and it comes before the writing decision. Here’s an example from our own work. Filmler.com had about 450,000 cast and crew records. We opened only the pages that had content to indexing.

Being able to generate a page doesn’t mean you should publish it. AI has made that question even more pressing, because now you could have a few paragraphs written for every one of those 450,000 pages. If those paragraphs carry no new information about the record, what you end up with isn’t content; it’s a textbook case of scaled content abuse.

When you’re thinking about producing pages at scale, ask these questions for each page type:

• Does this page carry real information that’s different from the other pages on the same template? The same text with the city name swapped doesn’t pass.
• Where does that information come from? Real data like your own database, your stock or your service areas, or generic text written by the model?
• If a user came to this page directly rather than from search, would it be useful to them? Google’s helpful content guidance describes the “why” question in exactly these terms: content should be useful even when a visitor comes to your site directly.
• Does this page answer a user’s question better than another page already on your site? If not, it’s better to strengthen the existing page.

The helpful content guidance also lists among its warning signs producing lots of content on many different topics in the hope that some of it performs well, and relying heavily on automation to produce content on many topics.

Take an example for the filling-machine manufacturer. Opening 81 pages, one per Turkish province, like “filling machines in Bursa” and “filling machines in Konya,” runs into scaled content risk if the company has no separate dealer, service or references in those provinces. But if the company has three regional service centers, opening three pages, each with its address, team, the provinces it covers and average response time, means three pages that give users real information.

8. Programmatic pages

Producing lots of pages from a template isn’t scaled content abuse in itself. An e-commerce site’s thousands of product pages, a real estate site’s thousands of listings and a movie site’s thousands of film pages are all built from templates. We call these programmatic pages, and much of the web works this way.

What makes a programmatic page legitimate is that there’s real, distinct data behind each one. A product page has that product’s price, stock, specs, images and reviews. A listing page has that home’s location, square footage and photos. The user gets information on that page they can’t find on any other.

The trouble starts when you try to fill the template with text instead of data. Here are a few familiar patterns from Turkey, with hypothetical examples:

• Province and district pages: pages like “leak detection in Kadıköy” and “leak detection in Üsküdar,” multiplied by swapping the district name, with no separate team, address or references in that district.
• Duplicated manufacturer descriptions: the same product description rewritten with synonyms or by AI and spread across hundreds of product pages. Same product, same information, just different words.
• “Best of” lists: pages like “10 best robot vacuums” where no product was tested and the list is stitched together from other sites.
• Question pages: pages opened for every variation of a question, repeating the same answer in different sentences.

In all of these the number of pages grows, but the amount of information on the site doesn’t. If it’s text and not information that’s driving up your page count, you’ve walked into scaled content risk.

9. Where AI is safe to use in content

With a proper review process, AI can be used safely for these jobs:

• Drafting and structure: producing a first draft from an expert’s notes, an interview recording or the sales team’s answers.
• Editing: making an existing page easier to understand, reworking paragraphs so they carry their own context, writing clearer headings.
• Data to text: drafting product descriptions from real product data. The data comes from you; the model just turns it into sentences.
• Translation drafts: producing a first draft of a page in another language. As we discussed in Chapter 16, translation isn’t localization; the translated page still has to answer the questions people in that market ask.
• Summaries and repurposing: turning a long report into a short summary, a webinar into an article, a case study into a social media post.

What these jobs have in common is that you are the source of the content. The model shapes the information you have; it doesn’t make up information.

One more note for e-commerce sites: Google Merchant Center has its own rules for AI content. According to Google’s guidance, AI-generated images must include IPTC DigitalSourceType TrainedAlgorithmicMedia metadata, and AI-generated product data such as titles and descriptions must be specified separately and labeled as AI-generated (Google Search Central). If you’re enriching product feeds with AI, check these rules alongside your feed work from Chapter 15.

10. A quality control process

The quality control process I suggest for AI-assisted content has five steps:

1. Accuracy. Google’s own guidance says generative models don’t retrieve information; they predict a likely sequence of words based on their training data and can therefore produce inaccurate information. It considers it critical that all AI-generated content be checked by hand for accuracy and reliability before publishing (Google Search Central). Check every number, date, technical spec and quote against its source.
2. Originality. Is there at least one thing on the page the model couldn’t have written on its own? A test result, a number, a photo, an expert comment, a case? If not, the page isn’t ready.
3. Metadata. Titles, meta descriptions, structured data and image alt text are content too, and Google’s guidance asks that they go through the same check.
4. Author and disclosure. Is there a real author or responsible expert on the page? If it’s the kind of content where a reader might ask “how was this made?”, is there a short note about the process?
5. Post-publication monitoring. After the page goes live, how does it perform in Search Console and in the prompt measurement from Chapter 17? Is it described correctly in AI answers?

This process takes back some of the time AI saves. It does. But I think the real gain from producing content with AI isn’t writing more pages in the same time; it’s writing the same number of pages better. Save time on drafting, not on quality control.

The process also needs to be written down. Put on one page how your team or agency uses AI for content: which jobs it’s used for, which it isn’t, which checks are mandatory, and who signs off. That page becomes a starting point for new team members, and when a client or management asks “how do you use AI?”, you already have the answer.

Don’t leave quality control to AI content detectors. Both studies above say themselves that these tools aren’t precise. Besides, what we should be looking at isn’t whether a text was written with AI but whether it adds value for the reader. A commodity text that a detector labels “human-written” is still a commodity text.

11. Before you publish

AI made producing content easier, but Google’s question hasn’t changed: who was this content made for? Before producing a page type at scale, open a few examples and put them side by side. If the pages differ only by city name or word choice, don’t publish them.

If each one has real, distinct information behind it, things are different. Then AI is a good helper for getting that information onto the page, and there’s no reason to hold back on publishing a page that’s passed quality control.

Checklist

A final word

We started this book with a question: if AI is giving the answer, what do we need to do for our brand to show up correctly in that answer?

Nineteen chapters later, I don’t have a magic technique. A large share of AI answers is still fed by search engines’ indexes and ranking systems; a page that isn’t crawled or indexed still doesn’t get read. What’s changed is how we define success. Ranking alone isn’t enough anymore; you need to make the shortlist, be described correctly and be recommended. And we track that with rates measured regularly, not with a single screenshot.

Throughout the book we kept coming back to the same place. AI looks for sources that say things it couldn’t know on its own: real prices, real capacity, real tests, real customer experience. Most companies already have that information. But it isn’t on the website; it’s in the sales team’s heads, in an Excel file, or in an engineer’s experience. I think the real work of the next few years is getting that information onto the web in a form that’s readable, verifiable and current.

In Turkey, AI Overviews and AI Mode started rolling out on February 18, 2026. In many industries the race has only just begun, and the pool of Turkish-language sources is still relatively small. Brands producing concrete, current Turkish content today can earn a place in AI’s Turkish answers that will last a long time.

This book is a “2026 edition.” In this year alone, Search Console got AI reports, Google added a separate control for its AI features, and spam updates became more frequent. A lot more will change in the coming months, and I’ll try to keep the fast-moving details current in the web version.

The part I don’t expect to change is short. Know your users’ questions, give them an answer no one else can, and measure it regularly. Search will keep changing. The value of a good answer won’t.

Sources

• Google Search Central Blog: Google Search’s guidance about AI-generated content (February 8, 2023)
• Google Search Central Blog: What web creators should know about our March 2024 core update and new spam policies (March 5, 2024)
• Search Engine Land: Google March 2024 core update rollout is now complete (April 2024)
• Google Search Central: Spam policies for Google web search (last updated August 28, 2026)
• Google Search Central: Google Search’s guidance on using generative AI content on your website (last updated October 1, 2026)
• Google Search Central: Creating helpful, reliable, people-first content (last updated October 5, 2026)
• Google Search Central: Optimizing your website for generative AI features on Google Search (last updated July 10, 2026)
• Google Search Status Dashboard: Ranking incidents history (checked November 1, 2026)
• Search Engine Roundtable: Google search manual action spike (October 8, 2026)
• Ahrefs: AI-Generated Content Does Not Hurt Your Google Rankings (600,000 Pages Analyzed) (July 7, 2025)
• Axios: AI writing hasn’t overwhelmed the web yet (October 14, 2025)

Previous: Chapter 18, AI in SEO Work · Table of contents