A client told me last month his SEO was “in great shape.” Page one for his main keywords, agency on retainer, the works. So I asked his own question into ChatGPT — “who's a good supplier for X in our city” — the exact question his buyers ask. ChatGPT named four companies. He wasn't one of them.
His site ranked. It just didn't exist where the decision was actually being made.
That gap is what this is about. Here is the short answer, then the long one.
Short answer
ChatGPT names you when three things are true at the same time. It can read your site. It can extract a clean answer from it. And the rest of the web agrees you're a credible source. The first two are the technical “fast lever” — fixable this week. The third is the “slow lever,” reputation, earned over months. Most businesses quietly fail the first one and never find out, because nothing on their dashboard tells them.
I run this exact playbook on my own site. The page you'd land on for “GEO services” is built to every rule below. So when you ask ChatGPT about it, it can actually tell you who I am and what I do. That is the only honest proof I can give you: my own site is citable, and I can show you the before and after. More on that further down.
“Cited” is not “ranked.” Don't confuse them
Two different games. People play one and assume they've won the other.
| Ranking (SEO) | Citation (GEO) | |
|---|---|---|
| What it is | You're a blue link in a list of ten | Your brand is named inside the AI answer |
| What the user does | Still has to click to reach you | Often never clicks at all |
| Your status | A candidate, one of many | The recommendation |
When a buyer asks an assistant “who's the best X for Y,” it doesn't return ten links. It names three to seven options and moves on. If you're in that set, you exist. If you're not, there's no page two for them to scroll — just a confident paragraph that doesn't have your name in it.
And this isn't a fringe behavior anymore. ChatGPT alone passed a billion users, and Gartner forecast classic search volume dropping around a quarter as AI answers absorb the questions. In pure AI-search modes, up to 93% of queries already end with no click at all (Similarweb). The channel is moving. The only question is whether you move with it or watch it from the blue links.
How ChatGPT actually decides who to name
It pulls from two places. You need both, and they work on different clocks.
Training data — your reputation before the model was built
This is what the web said about you across other people's sites: mentions in articles, reviews, directories, “best X” listicles, forum threads. The model absorbed all of it during training. You don't edit this directly — you earn it over months. This is why a brand with no third-party footprint stays invisible even with a flawless site: there was nothing about it to learn.
Live retrieval — what the model reads right now
When ChatGPT browses, or when Perplexity and Google AI Overviews answer, they send a bot to read pages in real time. This is the part you fully control, and it's the part that's usually broken. If the bot hits an empty page or a 403, you're out of the live answer no matter how good your reputation is.
So citation = earned reputation (slow) + a machine-readable site (fast, fixable this week). Let's do the fast one first, because it's the one you can fix without anyone's permission.
The fast lever: make your site readable, then extractable
Four fixes, in order. The first two are pass/fail — get them wrong and nothing else matters.
1. Serve real HTML, not an empty shell
Most modern sites are built as JavaScript apps — React, Vue, Angular — that assemble the page in the browser after it loads. A human sees a beautiful site. An AI crawler, which doesn't run your JavaScript, sees this: <div id="root"></div> — and nothing else.
Your headlines, your copy, your case studies — all of it lives in a JS bundle the bot never executes. Eight times out of ten, when I audit a site, that's what the AI sees. An empty div.
Test it yourself in ten seconds. Open view-source:yoursite.com in your browser. That's roughly what the bot gets. If your actual text isn't sitting there in the HTML, it isn't there for ChatGPT either. The fix is server-side rendering (SSR) or static prerendering — the content gets baked into the HTML before it's served.
2. Stop blocking the AI bots
This one is almost funny, because so many sites do it without knowing. They return 403 Forbidden to the exact crawlers they want to be read by. The culprit is usually one line in robots.txt or a default Cloudflare bot-fighting rule that treats GPTBot like an attacker.
The bots you want to let in, at minimum:
- GPTBot — OpenAI / ChatGPT;
- OAI-SearchBot — ChatGPT search;
- PerplexityBot — Perplexity;
- ClaudeBot — Anthropic / Claude;
- Google-Extended — Google's AI training and AI Overviews.
Open yoursite.com/robots.txt. If you see Disallow: / under any of those, or a blanket block, that's a brand decision someone made by accident. Check your CDN/WAF too — Cloudflare's “block AI scrapers” toggle does exactly what it says. The check takes one minute. Almost nobody does it.
3. Spell out who you are in machine-readable markup
Without schema.org markup, the AI has to guess what your business is, what you sell, at what price, where. Humans infer it from your design. Machines don't infer — they read structure or they move on.
The schema types that earn their keep:
- Organization / Person — who you are, with sameAs links to your real profiles (LinkedIn, etc.) so the AI can connect you to a known entity;
- Service / Product with Offer — what you sell and the price;
- FAQPage — your Q&A, in a format AI loves to pull verbatim;
- BreadcrumbList — where the page sits in your site.
You put this in the page as JSON-LD. It's not visible to users. It's a clean, unambiguous statement of fact for the machine. Stop making it guess.
4. Write so a passage can be lifted out
AI doesn't quote your whole page. It lifts passages — a sentence, a list, a definition. Most copy is written to be read top to bottom, which makes it useless for extraction. Rewrite for the lift:
- Answer first, explain second. Lead the section with the direct answer, then justify it — the AI grabs the first clean sentence;
- Question-shaped headings. “How much does X cost?” beats “Pricing” — it matches how people ask;
- Facts in lists and tables, not buried mid-paragraph. Numbers, prices, steps — pull them out;
- A real FAQ block with self-contained answers. Each answer should make sense on its own, ripped out of context — because that's exactly how it'll be used.
The receipts: my own site went from 25 to 64
I'm not going to hand you a checklist I haven't run myself. When I first ran a deep GEO audit on my own site, it scored 25 out of 100. Critical. And I build this for a living.
Same site. Same audit method. Two weeks of real work between the two readings.
Why so bad? Every mistake above, at once:
- it was a client-side React app — AI crawlers saw the empty div, my content lived only in the JS bundle;
- the AI bots were blocked twice — once in robots.txt, once by a default Cloudflare rule. Belt and suspenders, both wrong;
- zero schema — nothing told the machine who I was.
I rebuilt it to the four rules above. Prerendered HTML, bots unblocked, full JSON-LD. Two weeks. The score went from 25 to 64 — and kept climbing as the site aged. I'm not smarter than the agencies that “do AI visibility.” I just did the work and then I measured it, instead of running “some activities” and calling it a strategy.
The full teardown of my own site — what was broken and what moved the numberThat's the honest version. Not “easy and fast.” It was my own site, I knew exactly what to do, and it still took two weeks of real work.
The slow lever: become the consensus answer
Here's the part the agencies don't lead with, because it's the hard part: a perfect site, on its own, will not get you cited.
The model trusts entities the rest of the web already trusts. It looks for your name in places that aren't your own domain — industry articles, directories, real reviews with real names, the “best X” comparison lists, the forum thread where someone asked your exact question. If every mention of you lives on your own site, the AI reads that as a brand talking about itself. Which is worth roughly nothing — a University of Toronto study put a number on it: AI engines cite third-party mentions of a brand roughly 5× more often than content on the brand's own site.
I hear the objection every week: “we already work with an SEO agency, a couple of months now.” Fine. Ask the agency for one number: how many times your brand is named, by name, in a third-party source an AI would trust. Most can't answer, because backlinks and rankings were never the same thing as being the consensus answer. So the real off-site work: get mentioned where your category is actually discussed, earn reviews with real attribution, get into the comparison lists. The site and the reputation have to move together — that's the whole job.
How to measure it without fooling yourself
Do not track an “audit score” as your KPI. It's an input, not the result. Track citation rate — the share of buyer questions where the AI names your brand. The method takes an afternoon to set up:
Write down 15–20 questions a real buyer would type into an AI in your category. Mix the stages: informational (“what is X”), comparative (“best X for Y”), and decision (“who should I hire for X”).
Ask those exact questions to ChatGPT, Perplexity and Gemini. Each one, every engine.
Count how often your brand gets named.
Re-run it monthly. Same questions, so it's comparable.
The number that matters is the percentage of the time the machine recommends you when it counts. Not impressions, not a vanity score.
The mistake almost everyone makes
Treating this as SEO with a fresh coat of paint. Good SEO does not guarantee AI citation. Different crawlers. Different signals. Different technical base.
BrightEdge found the overlap between AI-answer sources and Google's own top-10 fell from 75% to 17–38% in under a year — a strong rank no longer predicts a citation. You can sit at #1 on Google and be completely absent from ChatGPT — I see it constantly, including from the client who told me his SEO was “in great shape.” They're separate games played on the same field. You have to play both, on purpose.
A starting checklist for this week
If you do nothing else, do these five, in order:
Run view-source: on your three most important pages. If the text isn't in the HTML, fix rendering first. Nothing else matters until it's there.
Open robots.txt and your CDN settings. Unblock GPTBot, PerplexityBot, ClaudeBot, OAI-SearchBot, Google-Extended.
Add JSON-LD: Organization/Person, Service, FAQPage.
Rewrite your top page answer-first, with a real FAQ block.
Build the prompt matrix — 15–20 buyer questions — and take your first citation-rate reading. That's your baseline.
The first four you can finish in a week. The fifth you'll repeat for as long as you're in business.
Want to check your own site?
I'll look at your site through the eyes of an AI crawler and send you a short read: where the exact holes from this playbook are, and what to fix first. No pitch — just where you're invisible to AI.
