How Do You Build a Site AI Actually Cites? (Part 1: Building It)

Is GEO ultimately a technical problem?

No. It’s mostly strategy. The tech gets you flagged if you do it wrong, but doing it well won’t earn you a citation.

As AI search eats into classic search, “GEO” (generative engine optimization) comes up more and more. The pitch is that a few tags and a few lines of schema get you cited in ChatGPT or Perplexity. After running this, I think the opposite. What earns citations is mostly strategy: what you answer with authority, and how far that answer has spread through your channels and outside mentions. The tech only keeps that answer from being unreadable or misread by a machine. It doesn’t produce the citation.

So when I rebuilt this site I flipped the order. I didn’t start by picking a pretty template or a framework. I decided first who reads the site. On top of people, the primary readers were AI crawlers and answer engines. What follows is a record of what I actually did on that footing, and what turned out to be wasted effort.

So what do you do on a page to actually get cited more by AI?

What research confirmed is adding sources, statistics, and quotations to the body. Visibility in generative engines rose by up to 40%.

The GEO paper from Princeton and Georgia Tech (KDD 2024) tested nine methods on real generative engines. The three that helped most were citing sources, adding quotations, and adding statistics. Those three lifted visibility metrics by 30 to 40%. Improving readability helped next.

The answer-first format I use, opening each section with a question and answering in one line right below, isn’t on that list. It’s not a method proven with a specific number. I use it because putting the answer up front makes it easy for an engine to lift a piece out. Every post on this site, from the first one to the one you’re reading, is in this format. I keep what’s verified separate from what merely seems reasonable. The 40% belongs to sources, statistics, and quotations, not to answer-first.

Does traditional SEO keyword optimization still work?

No. Pages stuffed with keywords did worse than the baseline in generative engines.

Classic SEO’s reflex is keywords. You repeat the words you’re targeting in the title and body to climb the search rankings. In generative engines this backfired. In the same paper, keyword stuffing performed below the baseline, and on Perplexity it was 10% worse. The engine looks at the grounds and the wording of an answer, not word frequency. Optimizing for search and optimizing for citation are not the same job.

Do machine-readable metas like JSON-LD, FAQ schema, and llms.txt help?

You should have them, but they won’t get you cited. FAQ rich results are already gone.

This site emits JSON-LD (post, author, organization), llms.txt, per-locale RSS, and hreflang. I have them because they help bots understand the page. But I don’t believe they produce citations.

FAQ schema is a good example. It used to show folded questions and answers in search results. Google restricted it in August 2023 to well-known government and health sites, and as of May 7, 2026 removed it from Search entirely. The days of getting a visible perk from a few lines of schema are over. I keep the machine metas minimal and correct, and earn citations with content.

To get AI crawlers to read you, and is optimizing for one engine enough?

Open them explicitly in robots. And don’t optimize for one engine, because the sources they cite differ from engine to engine.

To be cited you first have to be collected. This site’s robots.txt explicitly allows AI crawlers like GPTBot, ClaudeBot, and PerplexityBot. Block them and you drop out of the citation pool entirely.

But latching onto one engine and optimizing for it is a trap. Across several industry analyses, the domains ChatGPT and Perplexity cite in common came to only about 11%, from datasets of 100,000 and 110,000 items. The rest came from completely different sources per engine. Optimize for one and you miss the rest. So I keep the formats and channels spread wide.

Once it was built, the machine signals contradicted each other?

They did. As the code and the live site drifted apart, they were sending bots opposite signals.

The theory is all above, but once it was actually built, four machine signals were at odds with each other. First, visitors and crawlers coming to the home page were being redirected to a page in the wrong language. I’d set the default language to Korean, but the root was sending them to English. Second, a dead payment API endpoint inherited from the previous theme was being advertised, on every response, by a meta file that agents read. An agent’s first probe ended in a 404. Third, a site brand tagline was automatically appended after each post’s meta description, polluting the snippet. Fourth, on posts with no translation, the language switch link was a 404.

None of the four are easy to catch with the eye. The pages rendered fine. What was off wasn’t the human layer but the layer bots read. GEO’s 20% of tech turned out to be less about adding new features and more about keeping the signals you already send from contradicting each other.

So did it work?

Whether it worked, I don’t know yet. Honestly, that’s the truth. The site just settled in, and it takes weeks for AI engines to collect a new page and reflect it in citations. All I can say now is that the signals are in place. Not that anything got cited.

That’s why this is Part 1. What I built and what I’m watching is what’s here. Which engines actually start citing this site, I’ll write up with numbers in Part 2, after crawling and indexing have run. If GEO compounds, the compounding starts not from the first post but from the first citation.