You have the most complete article in your field. More data, more examples, better written than your competitor's. And still, when you ask ChatGPT or Perplexity about your topic, the answer cites other sources. You don't even get a mention.

It's not bad luck, and it's not a quality problem. It's that AI engines don't choose the "best" content: they choose what they can extract and verify fastest. And unlike quality, that can be designed.

AI doesn't reward the best: it rewards the extractable and verifiable

When you ask ChatGPT, Gemini or Perplexity a question, the model doesn't read your page like a person. It doesn't enjoy your intro. It does something colder: it scans the content looking for the passage that answers the question directly, plus a signal that the passage is trustworthy (a fact, a source, a number). With that, it builds the answer and decides who to cite.

This isn't an agency opinion. The paper that coined the term Generative Engine Optimization (Aggarwal et al., Princeton / IIT Delhi, KDD 2024) tested nine content tactics across thousands of real queries: you can raise a page's visibility in AI answers by up to 40% without changing the topic. What raised it? Adding concrete statistics, citing sources, and including direct quotes. What barely moved the needle? Adding more keywords and "polishing the prose." AI isn't won by whoever writes more beautifully: it's won by whoever leaves the fact served and backed up.

That's the trap. The long article, with the answer buried in paragraph 12 and not a single citable number, loses to the weaker article that put the answer clearly up top and backed it with a figure.

The "magic format" myth (and what Google actually said)

Let's stop a common misunderstanding: "convert everything to Markdown or upload an llms.txt and AI will cite you." That's not how it works. In its official guidance on AI features, Google was explicit: you don't need to create machine-readable files or special "AI text" to appear in its answers. It clarified this even further in a recent update: keeping an llms.txt file is perfectly valid for other services or AI systems that use it, but it neither helps nor hurts your visibility on Google, because Google Search simply ignores it. It's not bad practice — it's just not the lever on Google's side.

The confusion mixes up two different things. Markdown isn't the lever; semantic structure is. Headings that preview the answer, lists, tables, short paragraphs and clean HTML don't work because they're "Markdown": they work because they tell the machine where each thing is. Format is the vehicle, not the prize. And Google adds the nuance an advanced SEO already senses: on top of that structural base, what decides is non-commodity content — information with a point of view, original data, real experience — that can't be found the same way elsewhere. That's where a genuine human voice keeps making the difference. Structure so it finds you; substance and proof so it chooses you.

What does a generative engine check before citing you?

Translated into criteria you can audit in your own content today:

  • Clear structure. Headings that preview the answer, lists, tables, short paragraphs. Format gives the machine the map of your page.
  • Self-contained answers. A block that fully answers one question without depending on the rest of the text. If understanding a paragraph requires reading the previous three, that paragraph can't be cited.
  • Verifiable data and sources. The strongest lever in the GEO study. A claim with a number and a source is citable; the same claim in generic mode is not.
  • Entity clarity. Make it explicit which brand, which product and in what context you're talking about. AI cites what it can identify without guessing.

Main idea: AI isn't looking for who knows the most, it's looking for who makes it easiest to use and back up. The fight isn't "write more": it's writing in a way a machine can use and trust.

Why does your competitor show up, even with weaker writing?

Almost always your competitor doesn't beat you on content: they beat you on packaging and proof. Picture a dental clinic answering "how long does a dental implant last?". The site AI cites has an H2 with that exact question, answers in the first line ("Between 15 and 25 years with proper care") and adds a fact with a source. Yours opens with "Oral health is fundamental to well-being…" and hides the answer four paragraphs down, with no figure.

The model lands on your competitor's page, finds the answer and the proof in three seconds, and cites it. It lands on yours, finds no clean block and no data, and moves on. Same topic, same search intent, opposite result — and not because your information was worse.

Before / after: what citable content looks like

You don't need to rewrite everything from scratch. It's structure and proof work. Three concrete fixes:

  1. The heading. Before: "The future of marketing with artificial intelligence". After: "How do I get ChatGPT to recommend my business?". The H2 has to be the user's real question; AI uses headings to map what you answer.
  2. The first paragraph of each section. Before: 400 words of intro before getting to the point. After: the answer in the first 2-3 lines, and only then the detail. Put the conclusion up top, not as a closer.
  3. The unbacked claim. Before: "AI Overviews are changing traffic for many sites". After: "Optimizing content for generative engines can raise its visibility by up to 40% (GEO study, KDD 2024)". A figure with a source turns an opinion into a citable passage.

Add the trust basics — visible author, date, explicit entities (your brand, your product, your context) — and you have the difference between a good article and one AI can actually use. And that difference, technical as it sounds, opens a deeper distinction worth being clear about: ranking on Google is not the same as being cited by AI.

Ranking on Google is not the same as being cited by AI

Here's the underlying shift. Traditional SEO fights for a spot in Google's list so the user clicks and lands on your site. GEO (Generative Engine Optimization) — optimizing so AI engines use and name you — fights for something else: being the source of the answer, often with no click at all.

A quick note to place this: when we say "generative engines" we mean both Google's AI Overviews and ChatGPT or Perplexity. They share the same underlying logic — they reward what they can extract and verify — even though they run in different ecosystems with their own rules. What you build for one pays off in the others.

They're distinct goals and they coexist. You can be #1 on Google and invisible in ChatGPT, or the other way around. To place this article within our blog: here we focus on why AI chooses you or not — the extractability and proof criteria. If what you want first is to check whether your site already shows up in these engines, we cover that in Does my website appear in ChatGPT?. So "optimizing content for AI" is no longer an extra: it's a new layer of visibility, with its own rules, built on the same technical SEO as always.

The bottom line

If your content is good but AI cites others, the problem is almost never quality: it's that you're not making it easy or verifiable for the machine. You don't need to rewrite everything or chase a magic format — Google already confirmed one doesn't exist: clear structure, direct answers up top, data with sources and explicit entities are enough. Being cited by AI is a new layer of visibility, with its own rules, and like everything in SEO it's worked with method.

Frequently asked questions

Do I need an llms.txt file for AI to cite my site?

For Google, no. Google clarified that keeping an llms.txt is valid for other services or AI systems that use it, but it neither helps nor hurts your visibility on Google Search, because it simply ignores it. So it's not bad practice — it's just not the lever. What decides on Google is well-structured, well-supported content inside your own HTML.

Does converting my content to Markdown help?

Format itself isn't the lever. What helps is semantic structure — headings that preview the answer, lists, direct answers up top — and that's achieved in clean HTML without any "magic" format.

How do I start auditing my own content?

Check three things per page: (1) is the answer in the first 2-3 lines of each section?, (2) is there at least one fact with a citable source?, (3) are the headings the real questions users ask? If you answer no to all three, that's your first fix.

Does traditional SEO stop mattering?

No. GEO is built on the same technical SEO as always; they're layers that coexist. You can be #1 on Google and still be invisible in ChatGPT, so it's worth working both in parallel.

Want help making your content citable by AI and by Google? Book a meeting with Seotronix.