I once gave an AI coding agent a simple SEO task: choose topics, write articles, and keep going.
I did not ask for ten articles. I did not ask for one hundred. I aimed for 10,000 articles.
The agent worked for several days and actually completed the job. It produced thousands of long, well-structured pages. The files occupied hundreds of megabytes. At a glance, the result looked impressive: titles, headings, paragraphs, keywords, conclusions—the visual shape of a serious content library.
For a while, it even looked like the strategy was working. Googlebot eventually crawled all 10,000 generated article URLs. At the time, I treated that as validation. It was not. Crawling only meant Google had discovered and fetched the pages. Only about 1,000 were actually indexed, began appearing in Google Search, and generated real impressions and visits.
Then the opposite happened.
The generated pages started disappearing from the index. More followed. Eventually the problem was no longer limited to the AI articles: the entire site reached zero indexed pages, including the homepage.
That experiment permanently changed how I use AI for SEO.
Crawling is not indexing — and I learned that the hard way
This distinction sounds basic, but it is easy to blur when you are watching a large batch of URLs move through Search Console. In my case, Google crawled all 10,000 generated article URLs. That told me the pages had been discovered and fetched. It did not mean that all 10,000 had been accepted into Google's index.
Only around 1,000 pages were actually indexed and began appearing in Search. Those pages produced real impressions and traffic. The remaining thousands could be crawled without ever becoming searchable. Google's own Search Console documentation makes this distinction explicit: a URL can be crawled and still remain unindexed.
That changed how I interpret SEO signals. Crawling is not approval. It means Google is evaluating a URL. Indexing is a separate decision, and staying indexed over time is another test entirely. In my experiment, even the roughly 1,000 pages that had made it into Search eventually disappeared.
The dangerous part was not that the text looked bad
This is important because the articles were not obviously broken. They were long. They were grammatically correct. They had headings and a logical structure. If you opened one page in isolation, it could look like a perfectly acceptable SEO article.
But I had made a much deeper mistake: I had outsourced not only the writing, but also the reason for writing.
The agent chose the topics. The agent decided what to say. The agent created the structure. The agent wrote the text. I published at a scale where I could not realistically read, verify, edit, or improve every page.
There was almost no first-hand experience behind the content. No original research. No real story. Few meaningful examples. Often there were zero or one useful images. Most importantly, there was no strong answer to a basic question: why should this page exist when the web already contains thousands of pages about the same topic?
The pages had the shape of useful content without consistently containing the thing that makes content useful: original value.
I originally called it an “AI penalty.” That is too simplistic
When the site disappeared, my first reaction was to think that Google had detected AI text and punished the domain for using AI.
I no longer think that is the right way to describe it.
Google's public guidance is much more specific: using generative AI is not automatically against its guidelines. The problem is using automation—including AI—to generate many pages primarily to manipulate search rankings without adding meaningful value for users. Google calls this scaled content abuse.
That description is uncomfortably close to what I had built.
I cannot prove that a single named algorithm or a specific manual “AI penalty” caused every indexing decision on my site. Search systems are not that transparent. What I can say is exactly what happened in my own experiment: I published 10,000 largely autonomous AI articles for SEO; Google crawled all 10,000 article URLs; only around 1,000 were actually indexed, appeared in Search, and brought real traffic; those indexed pages then progressively disappeared; and eventually the entire site reached zero indexed pages.
The distinction matters. The lesson is not “Google hates AI.” The lesson is: AI does not turn mass-produced, search-first content into valuable content simply by making it long and grammatically clean.
Why 10,000 “good-looking” articles were still a bad content strategy
The biggest trap with generative AI is that production cost approaches zero while editorial responsibility does not.
Before modern AI, publishing 10,000 substantial articles required an enormous amount of money or human labor. That friction naturally forced publishers to ask which topics were worth covering. AI removes much of that friction. You can now generate an absurd amount of content before you have time to ask whether any of it deserves to be published.
That creates several problems at once:
- No real editorial selection: if an agent chooses thousands of topics because they have search potential, the site starts serving the search engine rather than an actual audience.
- Commodity information: the model often recombines information that is already widely available instead of contributing first-hand experience, original data, or a distinctive point of view.
- No practical quality control: at 10,000 pages, “I will review them later” is not a serious editorial process.
- Hidden factual risk: a fluent article can still contain invented details, outdated claims, or subtle technical mistakes.
- Weak page-level differentiation: many pages can be individually readable while the collection as a whole feels repetitive and interchangeable.
- Crawl and index waste: a huge low-value section gives search engines thousands of URLs to evaluate while the pages that actually matter compete for attention.
Length did not save me. Formatting did not save me. Having thousands of keywords did not save me. A content factory is still a content factory when every page contains 1,500 words.
What I did when the site fell to zero indexed pages
Once I accepted that the mass-generated section was the problem I needed to eliminate, I stopped trying to “optimize” it.
I deleted all 10,000 AI-generated articles.
Not a few hundred. Not only the pages that had already dropped. I removed the entire experiment. Those files represented days of generation and hundreds of megabytes of content, but keeping them because they had been expensive to create would have been a sunk-cost mistake.
For URLs that were intentionally and permanently removed, I returned 410 Gone. Then I stopped bulk publishing and changed the way I wrote.
I began creating articles around things I had actually done, broken, learned, measured, built, or thought about. Some topics may be interesting only to a narrow audience. That is fine. They are at least topics I have a reason to write about.
I still use AI heavily—but its role is different.
My new rule: I am the author; AI is the tool
Today I use AI for tasks where it is genuinely excellent:
- helping me turn rough notes into a clearer structure;
- challenging an argument and pointing out missing explanations;
- suggesting a better heading or introduction;
- fixing grammar, punctuation, repetition, and awkward wording;
- checking whether a technical explanation is understandable;
- helping format code examples and lists;
- translating an article I already wrote into other languages.
What I no longer want is the workflow: “choose 10,000 topics, write 10,000 SEO articles, and publish them.”
The difference is ownership. Now the idea starts with me. The experience is mine. I decide what is true, what matters, what should be removed, and what deserves emphasis. AI can help me express the idea better, but it is not responsible for inventing the entire reason the page exists.
I also read the finished article. That sounds embarrassingly obvious, but it is the most important difference between my old workflow and the new one. With 10,000 autonomous articles, I physically could not be the editor. With deliberate articles, I can.
The recovery was slow, which is another reason not to play this game
Deleting the content did not cause an instant recovery.
For a long time, the site remained effectively absent from Google. I kept the removed URLs gone, published a much smaller number of deliberate articles, and waited while Google revisited the site.
About two months later, the homepage returned to Google's index.
Compared with the roughly 1,000 pages that had previously made it into Search, getting only the homepage back may sound small, but it mattered much more to me. Google also started crawling the site again. At the time of writing, I am not claiming a complete recovery or that every new article will be indexed. I do not have evidence for that yet.
What I can confirm is narrower: the site went from zero indexed pages to having the homepage indexed again, and Google resumed crawling it after I removed the mass AI content and changed the publishing strategy.
That experience also taught me something about SEO risk: a shortcut can take days to create and months to unwind.
AI translation is a very different use case
I do not think the lesson is “never use AI to produce text.” My experience with AI-assisted translation has been almost the opposite.
If I write an article based on my own experience, reasoning, experiments, or expertise, translating that same article is not the same thing as asking an agent to invent thousands of search topics from nothing.
The underlying value already exists. AI is changing the language, not manufacturing the reason for the article to exist.
This is one of the uses of AI I am most positive about. In the past, translating every blog post into 5, 10, or 20 languages could require freelance translators or many hours of manual work. Today an AI model can create a strong first translation very quickly, which makes multilingual publishing realistic even for a solo developer.
But the SEO implementation still matters. On my sites I prefer language versions to exist as real, crawlable URLs with server-rendered or pre-rendered HTML rather than hiding all languages behind client-side state. Each version should contain genuinely translated main content, use the correct language metadata, be connected to its alternate versions with hreflang, remain internally discoverable, and be included in the site's crawl and sitemap structure where appropriate.
Google explicitly supports fully translated versions of the same page and recommends separate URLs plus hreflang annotations for multilingual sites. The important part is that the translated page is a real page for a real reader—not another low-value variation generated merely to multiply URL count.
I wrote about that workflow separately in my case study on translating a blog into many languages with AI.
Server rendering does not create quality, but it removes avoidable technical uncertainty
There is another distinction I learned to keep clear: content quality and technical SEO are different layers.
Server-side rendering or static generation will not rescue a useless article. Perfect hreflang will not make generic content original. A sitemap cannot create demand.
But once the content is worth publishing, I want search engines to receive a clean technical representation of it. For multilingual content, that means predictable language-specific URLs, crawlable server-rendered or pre-rendered HTML, reciprocal hreflang, clear internal links between language versions, and no accidental blocking of important pages.
Technical SEO should make valuable content easier to discover and understand. It should not be used as a substitute for value.
The workflow I trust now
- Start with something I actually know or experienced. A bug I fixed, a system I built, an experiment I ran, a failure, a result, or a real opinion I can defend.
- Write the substance before optimizing for keywords. I want the article to be useful even if Google never sends it a single visitor.
- Use AI as an editor and sparring partner. I let it improve structure, clarity, grammar, examples, and questions I may have missed.
- Verify every factual and technical claim. Fluent writing is not evidence.
- Add first-hand detail. Numbers, screenshots, code, mistakes, constraints, decisions, and outcomes make the article something a generic model cannot reproduce from a one-line prompt.
- Read the final version myself. If I would not put my name on it after reading it, I do not publish it.
- Translate the finished article when it makes sense. AI can help scale a valuable original into multiple languages.
- Implement the multilingual pages correctly. Separate crawlable URLs, clear language signals,
hreflang, internal links, and clean rendering. - Publish at a pace I can actually supervise. I would rather understand 30 pages than own 10,000 pages I have never read.
AI is an amplifier, not a source of purpose
The most useful way I have found to think about AI is as an amplifier.
If the starting point is real experience, a useful idea, good data, or a strong piece of original writing, AI can help amplify it: improve the prose, organize it, translate it, challenge it, and distribute the same value to more people.
If the starting point is “give me thousands of keywords so I can capture search traffic,” AI amplifies that too. It just lets you create the wrong thing at a scale that was previously too expensive to reach.
That is why I am much more careful now. I am not anti-AI. I use it every day. I am anti-autopilot.
My strongest conclusion from this experiment is simple: do not outsource editorial judgment to an agent just because the agent can generate content faster than you can read it.
Write something that comes from you. Use AI to make it clearer, stronger, better structured, and available in more languages. But keep the purpose, the expertise, the verification, and the final decision with a human.
AI made publishing easier than ever. That is exactly why deciding what not to publish has become more important than ever.