The New SEO — How to Get Cited by ChatGPT

Stop trying to rank a page. Start trying to be the sentence an AI quotes. When someone asks ChatGPT or Perplexity "best X for Y" or "is [person] any good," the model doesn't hand back ten blue links — it synthesizes one answer and cites a few sources. Your job in this new game is to be one of those cited sources. It's called Generative Engine Optimization (GEO), and the surprising part is that the tactics that won at Google can do nothing here.

There are four moves, and order matters: first find out where you already stand, then rebuild your pages so a model can extract them, then earn third-party consensus, and finally ship the technical files almost nobody bothers with. Each one is backed by research, and each one matters more if you're the underdog. Here's the whole thing.

The old playbook is dead — measure this first

Peer-reviewed work makes the break clean. In the GEO paper (Aggarwal et al., KDD 2024), the authors tested optimization methods across roughly 10,000 queries and found keyword stuffing — the reflex of a decade of SEO — produced no improvement in how often a page got surfaced by a generative engine. What did work: adding statistics, direct quotations, and cited sources lifted visibility by 30–40%. Hold on to that; it drives step two.

1. Audit where you actually stand

You can't fix a gap you can't see. Write down 10 to 50 real buying-intent questions — the exact things someone types right before they choose. For a dev tool that's "best open-source auth for Next.js", "Clerk alternatives", "Supabase vs your-tool"; for your own career it's "is [your name] a good staff engineer" or "who writes well about platform teams." Run every one through ChatGPT, Claude, and Perplexity — they pull from different places, so you'll get different answers. Where your name, product, or repo is missing from all three, that's your highest-priority gap — and you've just turned a vague worry into a concrete list of questions to go win.

Use a prompt like this to generate the list and grade yourself:

I'm running an AI-visibility audit.
Me / my product: [what you build + who it's for]
Closest alternatives: [2-4 names]

1. Write 20 buying-intent questions someone would type into ChatGPT, Claude,
   or Perplexity right before choosing in this space — mix "best <thing> for
   <use case>", "<me> vs <alternative>", and "<alternative> alternatives".
2. For each, answer the way an assistant would today and tell me whether I'd
   plausibly get named, and why or why not.
3. Rank the questions by how much it's worth winning each one.

2. Restructure so a model can extract you

Models pull passages, not pages — so every page has to be built for lifting, not just reading. Five concrete changes:

• Make each header the exact question someone asks, word for word.
• Answer in the first ~50 words under that header. Foundation's read of cited pages describes a three-layer shape: the direct answer first, then why it matters, then the deep detail — the answer up top is the part that gets quoted.
• Lead with data, not adjectives. Turn "our approach is faster" into "cuts a cold build from 12 min to 3." Per the GEO paper, data-backed claims and named citations were the two strongest levers; vague copy is invisible.
• Put comparisons in a table. Models copy clean tables almost word for word, and comparison content punches above its weight — Foundation found 32.5% of AI citations come from comparison-style listicles. Build the "you vs. the obvious alternative" page nobody bothers to write.
• Add an FAQ block on the pages that matter. Phrase every entry as a heading and keep each answer short, direct, and able to stand on its own.

Rewrite this page so an assistant can lift and quote it.
Page topic: [the one page that matters most]
Real facts: [features, numbers, prices, use cases]
Compare against: [main alternative + what you know about it]

1. Make the H1 the exact question a buyer would ask.
2. Open with a 50-word answer that stands on its own.
3. Add a comparison table (me vs the alternative) on the criteria buyers weigh.
4. Add 5 FAQ entries - each a question heading, each a 1-3 sentence answer.
5. Keep every claim factual and self-contained so it survives being quoted alone.

3. Build third-party consensus — the step everyone skips

This is the most important one, and the most overlooked: a model trusts what other people say about you far more than what you say about yourself. Foundation's data is blunt about it — roughly 85% of brand mentions in AI answers come from third-party sources, and only ~15% from a brand's own site. The single most-cited domain is Reddit (about 23% of top citations), with YouTube next (~13%). Models are effectively looking for the same claim to show up, independently, in several places before they'll repeat it.

So you have to exist off your own site: in the subreddits and dev communities of your space, in "best of" roundups, on YouTube, on podcasts, in conference talks. One honest third-party mention moves you more than ten posts on your own blog.

Help me earn third-party mentions so assistants recommend me.
Me + my space: [describe it]
My top buying-intent questions: [paste a few from the audit]

1. Which third-party sources do assistants cite most for my space
   (subreddits, dev communities, YouTube channels, roundups, podcasts)?
2. The exact search I'd run to find the threads/posts/roundups I should be in.
3. A genuine, non-spammy way to earn each mention - what I'd contribute, not
   "post about myself".
4. A 30-minute weekly routine to keep building it.

4. The technical files almost nobody ships

Most people stop at content. The ones who win also add the machine-readable signals — the plumbing that both identifies you to a model and makes sure it can reach your pages to begin with:

• Schema markup (JSON-LD): structured facts on your pages (Person, Organization, Product, FAQ) in a format machines parse cleanly. Foundation found 61% of cited pages use three or more schema types — it's table stakes, not a nice-to-have.
• llms.txt: a plain file at your domain root that describes what you are and lists the pages that actually matter — a site map written for models, and still rare enough to be an edge.
• Let the bots in. Check your robots.txt isn't quietly blocking GPTBot (OpenAI), ClaudeBot (Anthropic), PerplexityBot, or Google-Extended. A model that can't crawl your site will never cite it.

Set up the technical layer for AI visibility on my site.
Key pages: [list them + what each is about]
Stack: [Next.js, WordPress, custom, ...]

1. Draft an llms.txt - a short description of me plus my key pages, one line each.
2. The JSON-LD schema for my main pages (Person/Product/FAQ), filled in with my
   real details.
3. The robots.txt lines that confirm GPTBot, ClaudeBot, PerplexityBot and
   Google-Extended can crawl me.
4. Where each file goes on my stack, in plain steps.

Why this levels the field if you're unknown

Here's the non-obvious upside for a dev or EM without a big domain. Traditional SEO rewards incumbents — backlinks and domain age. GEO doesn't. The same study found the cite-sources method gave a 115% visibility increase for content ranked #5 in normal search, versus a drop for what was already #1. A precise, well-cited answer from a nobody can outrank a vague one from a giant.

My rule of thumb: don't write for the crawler, write to be quoted by a skeptical colleague. If a sentence about your work, your product, or your profile can't stand alone as a fact someone would repeat, an AI won't repeat it either.

Today: pick the one page that matters most — your product, your portfolio, your GitHub README — and rewrite its opening paragraph with one hard number and one cited source. That's your first citation.

Sources

• GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024)
• Columbia Tow Center — AI Search Engines Are Bad at Citing News
• Foundation — Generative Engine Optimization: The Definitive Guide