Eighteen months ago I started grepping access logs for GPTBot the way I used to watch for Googlebot. Back then it was a curiosity — a few dozen hits a week. Now, across the dozen-plus sites I run, AI crawlers are a steady slice of all bot traffic, and referral visits from ChatGPT and Perplexity convert at somewhere around double my organic-search average. The volume is still small. The intent is not.
Getting cited by AI search is not the same game as classic SEO, and most of what’s published about it is theory written by people who haven’t tested anything. I’ve run the experiments — same content on different hosts, with and without llms.txt, schema on and off — and the pattern is blunt: speed and structure decide who gets quoted. Most of the fashionable tactics don’t. Here’s the evidence-led version.
In this guide
How AI search actually picks its citations
Your site shows up in an AI answer through one of two doors. The first is training data — the model absorbed your content months ago, and you can’t influence that on any useful timescale. The second is retrieval, and that’s the one you can win. When someone asks ChatGPT Search, Perplexity, or Google’s AI Overviews a question, the system runs a live search, fetches a shortlist of candidate pages, extracts the passages that answer the question, and cites the survivors.
That pipeline has three gates: you have to be in the underlying search index, your page has to come back fast enough for the fetcher to bother, and your content has to be extractable — a clean, self-contained passage the model can lift and attribute. You are not optimizing for a ranking. You are optimizing to be the easiest page to quote.
One prerequisite before anything else: check robots.txt. If you — or a security plugin, or an overzealous WAF rule — blocked GPTBot, OAI-SearchBot, PerplexityBot, or Google-Extended somewhere along the way, nothing else in this guide matters. I’ve audited sites that spent months on “AI SEO” while their firewall quietly returned 403s to every AI user agent that knocked.
llms.txt — does it actually work?
llms.txt is a proposed standard: a markdown file at your site root that hands language models a curated map of your most important content. It takes ten minutes to create, and every AI-SEO thread on X treats it as step one. So I tested it — deployed on half of my sites since early 2025, deliberately left off the rest.
The result: no citation difference I can attribute to it. My logs show the major AI crawlers requesting llms.txt rarely or never, and no major provider has confirmed consuming it. The public crawl-data analyses I’ve read land in the same place. It’s a lottery ticket, not a strategy.
Ship one anyway if you like — it’s harmless, and the standard may get real adoption later. But it’s dessert, not dinner. Here’s where the effort-to-impact ratio actually sits, based on my testing:
| Tactic | Effort | Citation impact |
|---|---|---|
| Allow AI crawlers in robots.txt | Minutes | Prerequisite — nothing works without it |
| Fast origin (TTFB under ~200ms) | Depends on your host | High — the strongest pattern in the data |
| Answer-first page structure | Ongoing discipline | High |
| Article + FAQ schema | Low | Moderate — helps entity disambiguation |
| llms.txt | Minutes | Negligible today |
| Paid “GEO” tool subscriptions | $100+/mo | Mostly repackaged versions of the rows above |
Speed is the citation factor nobody budgets for
A large 2025 analysis of pages cited in ChatGPT answers found that pages with first contentful paint under 0.4 seconds picked up citations at roughly three times the rate of slower pages. That matches what I see in miniature: my fastest sites collect the citations, and my slower experiments — carrying content of the same quality — mostly don’t.
The mechanics explain it. Retrieval fetchers run on tight timeouts and pull a shortlist of candidates in parallel. If your origin takes two seconds to start responding, you drop out of the candidate set before the model reads a word. And unlike Googlebot, most AI fetchers don’t render JavaScript and won’t circle back later to give you a second chance. One slow response is a lost citation.
Here’s the part most GEO advice skips: AI crawlers disproportionately hit your pages cold. They pull deep, odd URLs that aren’t sitting warm in your page cache, so what they experience is your uncached origin performance — PHP execution, database queries, worker availability. If you’ve read my PHP workers guide, this is that same bottleneck wearing a new hat.
This is where hosting stops being a commodity decision. On cheap shared hosting I routinely measure uncached TTFB between 600ms and 1.5 seconds — disqualifying. The sites I run on Kinsta (Google Cloud C2 machines behind Cloudflare’s enterprise edge) hold uncached TTFB around the 200ms mark, and WP Engine‘s EverCache stack does similar work at a lower entry price — ~$20–25/mo at the time of writing. My full numbers are in the managed WordPress hosting comparison and the AI web hosting pillar. For AI visibility specifically, treat origin speed as a ranking factor you buy, not one you tune into existence on a $4 plan.
Write so a machine can lift the answer
Extraction rewards a specific shape. When I rewrote one underperforming guide into that shape — no new information, same topic — it started appearing in Perplexity answers within a few weeks. The shape:
- Answer first. The first sentence under every heading should be a complete, standalone answer in roughly 40–60 words. Context and caveats come after, never before.
- Headings as questions. Phrase h2s the way a person actually asks. Retrieval matches queries to passages, and a matching heading is the strongest signal that a passage answers the query.
- Lists and tables. Structured fragments get lifted disproportionately — they arrive pre-chunked for the model.
- One idea per section. A passage that mixes three topics survives extraction as zero citations.
- Visible dates and a named author. Assistants lean toward fresh, attributable sources — anonymous, undated pages get passed over for the same facts published with a byline.
None of this fights classic SEO. It’s the featured-snippet playbook with the screws tightened.
Schema that actually gets read
Structured data won’t rescue a slow, muddled page, but it does two jobs well. It disambiguates who you are — Organization and Person markup with sameAs links ties your site to an entity the model already knows — and it labels what your content is: Article with author and dates, FAQPage where you genuinely have Q&As. In my testing that’s the moderate-impact tier. Measurable, not magical.
On WordPress, Rank Math or Yoast covers the basics; validate with Google’s Rich Results test rather than trusting the plugin’s green light. And if you’re on WP Engine, their Smart Search add-on builds vector indexes of your own content for on-site AI search — a different job than external citations, but the same structured-content discipline feeds both. Setup walkthrough in my Smart Search guide.
How to know it’s working
There’s no Search Console for AI citations yet, so measurement is manual. Three habits cover it:
- Grep your access logs monthly for GPTBot, OAI-SearchBot, PerplexityBot, and ClaudeBot. Rising crawl frequency reliably shows up before rising citations do.
- Segment AI referrals in analytics. Filter for chatgpt.com and perplexity.ai referrers. These visitors arrive pre-sold — the assistant already vouched for you — which is why their conversion rates run high.
- Ask the money questions yourself. Once a month I put my ten most valuable queries to ChatGPT, Perplexity, and Google AI Overviews and record who gets cited. It’s tedious, and it’s the single most honest metric available.
Expect lag. Retrieval systems pick up fast pages in days to weeks, but citation share builds over months as your crawl history deepens. This is a compounding channel, not a campaign.
Frequently asked questions
How do I get my website cited by ChatGPT?
Allow GPTBot and OAI-SearchBot in robots.txt, get your uncached TTFB under roughly 200ms, and structure pages so every heading is followed by a standalone 40–60 word answer. Citations go to pages that are indexable, fast, and extractable — in that order.
Does llms.txt actually improve AI visibility?
Not measurably, as of mid-2026. No major AI provider has confirmed using it, and in my own split testing across a dozen sites it produced no attributable citation difference. It’s harmless to add — just spend your real effort on speed and structure first.
How long does it take to show up in AI search results?
Retrieval-based systems like ChatGPT Search and Perplexity can cite new or updated pages within days to a few weeks of crawling them. Getting absorbed into a model’s training data takes many months and isn’t worth planning around.
Should I block AI crawlers instead of courting them?
If your business is selling the content itself, maybe. For most site owners, AI answers are becoming a discovery channel, and blocking GPTBot trades long-term visibility for short-term principle. I allow retrieval bots on every commercial site I run.
Want the guided version? Our free courses walk you through building a fast, citation-ready site start to finish — including “Launch Your First Website with Claude.”