What kinds of content still work now that AI can generate anything?

Three categories hold up. Original research, meaning surveys and insights drawn from your own systems and customer base, because a model cannot run your survey or read your platform data. How-we-built-it content, where engineers walk through the architectural decisions and trade-offs they actually made. And launch content that crisply explains what the product is, who it is for, how to start, and what it costs. Everything else is answerable by a chatbot.

Why do programmatic SEO glossary pages stop working?

Because the page has no unique data behind it. A template filled with generated definitions is a worse version of an answer a chatbot gives directly, so visitors bounce almost immediately. Programmatic SEO still works when each page carries something nobody else has, like real listings or real integration data. The technique is not the problem. Filling it with generated prose is.

What is how-we-built-it content?

An engineering account of how something was actually constructed, not a description of what shipped. It covers the architectural decisions, the trade-offs weighed, and the lessons taken from how the problem was solved before. Other engineers value it because those decisions cannot be inferred from documentation, which makes it some of the most useful material a technical marketing team can publish.

Is there any place left for AI-generated content?

Yes. Internal documentation is a reasonable use, and so are routine posts where nobody expects an original idea. The line falls at original thinking. If the piece depends on an insight that is yours, the writing has to be yours too, and the evidence for that is performance data rather than principle. Posts dictated as ideas and then written by a model did not perform, even though the ideas were genuinely original.

What is the difference between Level 1 and Level 2 content?

Level 1 is original content with original ideas, written by a human. Your blog, your homepage, your press releases, your product detail pages. It has to be structured so an AI can read it and presented so a human enjoys reading it. Level 2 is content generated by an AI and intended for an AI to absorb, index, and use in its answers, which is the modern equivalent of programmatic SEO. Humans finding Level 2 content is fine, but they are not who it is for.

How do you structure content so an AI can read it?

Put the conclusion at the top, because citations cluster heavily in the first third of a page. Frame H2 headings as the question a reader would actually type, then answer it in the first sentence underneath. Use definitive language rather than hedged phrasing. Name specific tools, companies, and people instead of gesturing at them. On the mechanical side, add JSON-LD to the page, FAQPage schema to the FAQ block, a clean heading hierarchy, and an llms.txt at the root. The FAQ block is the highest-leverage piece, because a question with a self-contained answer is the only part of a post that survives being pulled out of context.

Should startups still write how-to tutorials and comparison posts?

They are a weak investment now. The answers that kind of content delivers come straight from an LLM, so a reader asks a chatbot instead of finding your page. A smaller company gets far more out of original research, engineering content unique to what they built, and launch content with strong positioning.

The content that can't be generated

This last year I've run several experiments with AI generated content. And they all sucked.

They sucked a lot.

The first thing I did was I created 700 glossary pages. All the stuff that you would do in the old Programmatic SEO era. It cratered all my numbers in terms of engagement and visitors on the website.

Then I created a blog where I would dictate to AI. The ideas were mine. I'd record a list of things I wanted the blog post to say, several quotes that I wanted included verbatim, numerous anecdotes from my career, and the structure of the post and argument I was putting forward. And then I asked Claude to actually write the blog post. And those blog posts didn't perform. Sure, the ideas were mine, but the writing wasn't and people could probably tell.

An aside about AI content

I'm not opposed to AI generated content. It has its place. Internal docs, sure. Maybe even most perfunctory blog posts, too. But blog posts where you have an original idea, my data says that it needs to come from you and your own hands on keyboard.

I write a lot of fiction (including a novel that's coming out in September 2026). Those words come from me. Those stories come from my heart and my experiences and observations. They run the gauntlet of real developmental, copy, line, etc. editing. So in the end, the words are mostly mine, probably 60%, and a lot of the rest comes from the hands of my editors. Editing can be rough on the ego. Kill your darlings, as they say. But the stories and soul remain from me at all times.

I think the same holds true for writing in a business context.

The difficulty of writing original content

Two things make writing original content difficult today.

First, readers are inundated with blog posts, videos, tutorials, podcasts, you name it. Some of it comes from individuals, a lot of it AI-assisted, and some of it produced entirely by AI. The volume makes it difficult to stand out.

Second, a lot of this content gets absorbed by LLMs, and instead of reading content, people will ask the LLM for the answer. What then is even the point of writing content in the first place?

With this in mind, I think about writing content at two different levels.

Level 1: original content with original ideas, written by a human. This is your blog, your homepage, your press releases, and your product detail pages. It has to be structured so an AI can read it, and presented so human readers enjoy it.
Level 2: content generated by an AI and intended for an AI. This is the modern equivalent of programmatic SEO. If humans find this content, that's fine, but it's there for an AI to absorb, index, and use in its answers.

In this post, I'm going to focus solely on Level 1. If there's interest, I'll write about Level 2 later.

Original research

What type of content actually works? What is the kind that matters and can't be generated?

The first kind of content that can't be generated is original research. Surveys. Information that you've gleaned from your own systems. If you have a lot of customers, what kind of information do all those customers generate? Think the Supabase State of Startups survey or the Carta data desk. What kind of trends and insights can you get from your customer base and how they use your platform? That kind of content always performs.

It's unique, it's new, and it sparks original conversations.

How we built it

The second kind of content that works is "how we built it" content. Get your engineering team to explain how they built a technology.

The architectural decisions that we made. The things that we learned from how this was done in the past and how changes in customer behavior or industry trends change the requirements. The trade-offs and the real decisions that engineering teams make are the kind of thing that other engineers love. And in developer marketing and technical marketing, that is the key to winning credibility and trust.

Launch and product content

The third kind of content that works is the classic product marketing launch content. Content that crisply explains what the product is, who it's for, how to use it, how to get started, how much it costs.

This category of content is necessary information. It positions the product in the customer's mind and helps them make a decision about adoption. As I always write, you want to help customers imagine their life with your product in it.

The other day, I was looking at the Wispr Flow website. I had heard about everyone using this product to talk to their coding agents. But I am looking for a dictation product, because increasingly I want to write by speaking instead of typing. The Wispr website did a phenomenal job of explaining what it does and how customers just like me use it to write, both code and prose. I was immediately sold on the product.

Great homepage. Great product detail pages. Great pricing page.

Structuring content for AI

Earlier, I talked about content needing to be structured so that an AI can read it. There have been a lot of recent analyses about building content in a way an LLM can absorb, understand, and reflect. The industry calls this AEO, which is the spiritual successor to SEO. Just like SEO, AEO can be gamed.

The writing is always yours, but we always want to structure the writing in a way that it's both good for the human reader as well as an agent.

My friend Kevin Indig wrote up an analysis of 1.2 million ChatGPT citations in his Growth Memo newsletter. Using that empirical backbone, he was able to identify certain characteristics of content that frequently show up in LLM responses.

The ski ramp. 44.2% of citations come from the first 30% of a page, 31.1% from the middle, and 24.7% from the last third. Which means the conclusion goes at the top and you do not save the good insight for the end.
Question-framed H2s. Content with question marks in headings gets cited twice as often, and 78.4% of those question-based citations come out of the headings themselves. Frame the header as the query somebody would type, then answer it in the first sentence underneath.
Definitive language. Citation winners use 1.8x more definitive language, the "X is defined as" and "X refers to" constructions, than the vague alternatives.
Entity density. Cited content runs 20.6% entity density, meaning proper nouns like brands and named people, against 5% to 8% for standard text. Name the tools, the companies, the versions. Vague content does not get cited.
Inside the paragraph. 53% of citations come from paragraph middles rather than opening sentences. The model is hunting for information gain wherever it sits.

The full blog post is worth reading.

There are other things that people recommend doing that lack Kevin's analytical rigor. Nevertheless, I do them on this site, and I encourage everyone to do them everywhere else as well:

JSON-LD on the page
FAQPage schema on the FAQ block
Proper heading hierarchy
An llms.txt at the root

Every post on this site carries FAQ frontmatter. A question with a self-contained answer is the easiest thing in the world for a model to lift, and it is the only part of a post that survives being pulled out of context.

While I am advocating for writing for the machine, it is more important to write well and write in a way that humans are interested in your subject matter.

What doesn't work anymore

Think about all the other kinds of content I could create. The how-to tutorials. The blog posts about a certain topic. Finding topic clusters to become an information expert.

That kind of content doesn't work anymore, because the answers that kind of content delivers just come from LLMs. You ask ChatGPT, ChatGPT tells you. Comparison content, ChatGPT figures it out for you.

There's a good argument to be made that this kind of programmatic content can be generated for LLMs to crawl and index in a way that is invisible to your site visitors. That's the Level 2 post.

But right now, it's pretty evident what does and does not work anymore.