Getting cited in AI search.

Ask ChatGPT which trading-bot alert service is fastest, or ask Perplexity for the best guide to Hyperliquid, and you get a paragraph with a few sources under it. Six things decide whether your page is one of those sources: whether the AI crawlers can reach it, whether the words are in the raw HTML, whether any single paragraph answers a question on its own, whether schema says what the page is, whether a snippet cap is throttling how much can be quoted, and whether an llms.txt exists. This is how each one works and how to check it on your own pages.

How is getting cited different from ranking?

Ranking wins a place in a list the reader chooses from. Being cited wins the passage the engine lifts into an answer the reader may never click past. Classic SEO gets you ten blue links: the searcher scans, clicks, and lands on your page. Generative Engine Optimization, GEO, plays a different game. An AI engine reads a handful of pages, writes one answer, and cites the pages it leaned on. The searcher often never clicks at all. They read the answer and move on. We wrote up the full comparison in GEO vs traditional SEO.

That changes what winning looks like. In classic search you fight for position. In AI search you fight to be quoted. Two things follow from that.

First, the top of the results still matters, more than people expect. When an AI Overview or a Perplexity answer assembles its sources, it pulls heavily from pages that already rank in the first page of Google. Google says the same thing from the other direction: its AI features documentation states there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary, and that the usual SEO fundamentals still apply. Read that as a description of where the answer gets its candidates. Weak classic SEO caps your AI-search ceiling before you start. GEO does not replace SEO. It sits on top of it.

Second, the engine has to be able to read your page cleanly and lift a passage out of it. A page that a person can read fine but a crawler cannot parse is a page that never gets cited. Most of GEO is making your content trivially easy for a machine to fetch, understand, and quote.

How do I get cited in AI search results?

You get cited by making the page trivially easy for a machine to fetch, understand, and lift a paragraph from. In practice that is six levers, run in this order: let the AI crawlers in, serve the real content in the raw HTML, write in self-contained passages, mark the page up with schema, remove the snippet cap, and publish an llms.txt. We run these on our own sites first. Every guide and product page we ship goes through the same checklist, so by the time we bring it to a client build it has been tested on something we own.

This page explains the mechanism. Which levers the work actually lands on once a site is past the basics is a different question, and we measured it: across twenty weeks on eight live sites, the first three passed on 100 percent of the sample with zero variance, while 41.0 percent of everything implemented landed in content and internal linking instead.

Are the AI crawlers allowed in?

If the bot cannot fetch the page, nothing else on this list matters. The first mistake we see is a robots file or a firewall rule that blocks GPTBot, ClaudeBot, PerplexityBot, or Google-Extended.

Read the vendors' own references before you write that rule, because the agent names do different jobs. OpenAI documents GPTBot, OAI-SearchBot and ChatGPT-User in its bots reference, and says there that ChatGPT-User is not used for crawling the web: it fetches a page because a person asked ChatGPT a question that moment. Anthropic separates ClaudeBot from Claude-User in its crawler support article, one for training and one for answering. Perplexity publishes PerplexityBot and its IP ranges in its bot guide. Google lists Google-Extended in its crawlers reference, where it governs AI training and grounding rather than Search ranking. Block the wrong one and you have shut off the answer rather than the training set.

Then read the edge logs, and be careful what you read off them. A refusal rate per user agent is close to meaningless, because most requests carrying an AI-crawler user agent are not the crawler. On this domain, across the seven days to 17 August 2026, 1,499 of the 1,972 requests presenting one of those agent strings asked for a path the site has never published: /.env, /credentials.json, /wp-config.php.bak. Those are credential scanners borrowing a crawler's name, and the firewall was right to refuse them. Group the log by user agent, path and the CDN's verified-bot category instead. Done that way the picture inverted: in that window every request Cloudflare verified as a real AI crawler was served a 200 or a 301, and every refusal on a page that exists came from an unverified impostor. These figures are ours, read from Cloudflare's httpRequestsAdaptiveGroups dataset for concept211.com over the dates given.

Does the raw HTML already contain the content?

It has to. An AI engine that fetches your URL should see the actual words in the response, not an empty shell that fills in later with JavaScript. Client-rendered pages are a real citation risk: the crawler grabs the HTML, finds nothing, and leaves. We server-render or pre-render anything meant to be found, and we verify it with a plain fetch of the URL rather than trusting what the browser shows after scripts run. If the words are not in the raw response, the page is not citable.

Can a model quote one of your paragraphs on its own?

That is the test, because AI engines quote self-contained chunks. A paragraph that answers one question completely, without needing the three paragraphs above it for context, is a paragraph an engine can drop into an answer. So we structure content in passages: a clear question as a heading, then a direct answer in the first sentence, then the detail. We put a short, standalone summary near the top of important pages, the kind of thing a model can quote word for word and be correct.

Does your schema say what the page is?

Schema is how you tell a machine what it is looking at without making it guess. Article markup on a guide, Organization and Service markup on a firm, Product or FAQ where they genuinely apply. Which types are worth writing is not a matter of taste: Google's structured data documentation lists the ones it actually reads, and anything outside that list is markup you are writing for other parsers. It does not force a citation, but it removes ambiguity about the entity and the claims on the page, and that makes a page safer for an engine to trust.

Is a max-snippet cap throttling your own quotes?

Check, because most templates set one by default. There is a robots meta tag, max-snippet, that caps how much of your text a search engine may show; Google specifies it alongside nosnippet and data-nosnippet in its robots meta tag reference. The same three appear on Google's AI features page as the controls that limit what its AI features may show from a page, so this is not a classic-search-only setting. Set it low and you have told Google, and the AI Overviews that draw on Google, to use only a sliver of your page. We set max-snippet to -1, unlimited, on content we want quoted. It is a one-line fix that a lot of templates get wrong by default.

Should you publish an llms.txt?

Yes, and almost nobody has. An llms.txt file is a plain-text summary of your site written for models: who you are, what you offer, where the important pages live. Be clear about its status. It is a proposal by Jeremy Howard, published in September 2024 and revised to a second version in August 2026, and no search engine has committed to reading it. Google's AI features documentation does not mention it at all. So write it for the assistants that fetch pages on request rather than expecting it to move a ranking. It is low effort and still uncommon, which is exactly why we put one on every site we run, including this one.

How do you measure whether you are being cited?

You measure it with citation tracking, because AI citations do not show up in a normal rank tracker. This is the problem our own product, ConceptSEO, was built to solve. It tracks GEO citations and the keywords that trigger them: which of your pages get cited, for which prompts, on which engines, and how that moves week to week. Classic rank tracking tells you where you sit in Google. GEO tracking tells you whether the machines that write the answers are quoting you at all.

We watch two things together. Classic positions, because they set the ceiling for what can be cited, and citation share, because that is the actual outcome. When a page climbs into the top ten and starts getting quoted in AI answers for its target prompts, the levers worked. When it ranks but never gets cited, the passage structure or the crawlability usually needs another pass.

What does the fix look like on a real page?

Unglamorous, and mostly engineering. The one example we can show end to end is our own, because a client site would have to be anonymized first: the Hyperliquid Guide, an independent 18-chapter reference to a fully on-chain perpetual futures exchange.

We ran the six levers on it in the order above. Fetch the homepage as GPTBot, ClaudeBot, PerplexityBot or Googlebot and all four get HTTP 200 and 162,663 bytes of rendered HTML, so nothing waits on JavaScript and there was no firewall rule to unpick. We rewrote the key sections as passages that answer one question without needing the paragraphs above them, put Article schema on, took the snippet cap off, and confirmed each step with a plain fetch rather than a browser. The guide joined the weekly cycle on 8 March 2026, and 614 of the 629 recommendations raised against it have been implemented since, each one a row in the platform rather than a number someone remembers.

Where it sits now: 519 of its pages earned Search Console impressions in the 28 days to 25 July 2026, taking 35,723 clicks against 588,999 impressions. Read those as the guide's current footprint, not as the yield of this checklist. Crediting them to the six levers would need a control we do not have. The crawler response above is the honest thing to point at, and you can check that one yourself in a minute.

What is GEO, in one paragraph?

GEO is the practice of getting your pages cited by AI engines like ChatGPT, Perplexity, and Google's AI Overviews, rather than only ranked in classic search. It works by keeping AI crawlers unblocked, serving real content in the raw HTML, writing self-contained passages an engine can quote, marking pages up with schema, removing snippet caps, and publishing an llms.txt. Because AI answers draw heavily on pages that already rank, strong classic SEO is the foundation rather than a substitute. The result is measured as citation share: how often the engines quote you for the prompts that matter.

Methodology, and reusing this checklist

The six levers are not a survey of other people's advice. Each one is a check we run on our own sites first and then on client work, verified the same way every time: a plain fetch of the URL per AI user-agent with the response code recorded, the raw HTML read for the body text rather than the rendered DOM, the robots meta read for a max-snippet value, and the schema validated against the page it describes. Where this page cites measured outcomes rather than mechanism, the figures come from the twenty-week benchmark and the window and sample size are stated there. Nothing on this page is modeled or projected.

Two kinds of claim sit on this page and they are sourced differently. Anything about what an engine or a crawler does is linked to the vendor's own documentation at the point it is claimed: Google for max-snippet, AI features and structured data, OpenAI for GPTBot and ChatGPT-User, Anthropic for ClaudeBot, Perplexity for PerplexityBot, and llmstxt.org for the llms.txt proposal. Anything about what we observed is ours, with the site, the dataset and the date window named in the sentence that makes it, so you can tell first-hand measurement from documentation without following a link. If you want the longer comparison against classic search, it is in how GEO differs from traditional SEO, and the practice that runs this checklist on a site is our technical SEO and GEO service.

You may republish or adapt this checklist with attribution and a link to https://concept211.com/articles/geo-ai-seo-getting-cited-in-ai-search/. Every heading above carries a stable id, so you can link to one lever rather than the whole essay.

Photo by Stephen Dawson on Unsplash