How AI search engines decide what to cite

AI search engines cite pages that answer the exact question in the first two lines, support that answer with checkable specifics, and live on a URL a crawler can fetch and index. Google rankings predict this far less than people assume: across 15,000 long-tail queries, only 12% of URLs cited by AI assistants also ranked in Google's top 10 for the same query, with Perplexity the outlier at 28.6% (Ahrefs, 2025).

That gap is the opening for small sites. A four-page portfolio will never outrank a domain with ten thousand backlinks, and it can still supply the passage a model quotes, because models quote sentences.

How a page becomes a citation

The engine rewrites your prompt into several narrower queries, runs them, retrieves a few dozen candidate pages and pulls short passages out of them. It drafts an answer from those passages and links each one it leaned on. Two gates decide your fate: retrieval has to find the page, and the passage has to survive once it is cut away from the article around it.

The second gate is where most sites lose. A paragraph that opens with "in this section we will look at several factors" carries no information once it is lifted out of context, so the model drops it and quotes the competitor who wrote a number and a date.

The query fan-out also explains the weak overlap with Google's top 10. The engine never searched the user's original phrasing. It searched four or five sub-questions nobody typed, and your page was judged against those.

What the research says raises the odds

The GEO study from Princeton and IIT Delhi tested nine content changes against a benchmark of 10,000 queries and reported visibility gains of up to 40% in generative engine responses (Aggarwal et al., 2024). Adding relevant quotations performed best, at 41% relative improvement on their position-adjusted word count metric. Statistics and cited sources landed between 15% and 40%, depending on the domain.

Keyword stuffing produced little to no improvement and sometimes made things worse. Tactics that carried search engines for two decades do not transfer.

Volume matters for expectations. The share of US ChatGPT prompts producing a citation to an external site rose from 1.6% in June 2025 to 6.8% in May 2026, with travel and automotive far above average and professional services under 4% (Similarweb, 2026). Treat AI citations as a growing side channel and size your expectations accordingly.

Google's own guidance points the same direction and says so plainly:

From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

The same document sets one hard requirement: a page must be indexed and eligible to appear with a snippet before it can surface in any generative feature. It also tells you to skip llms.txt files and stop chopping content into fragments for machines (Google Search Central, 2026).

What I changed on my own site

My portfolio runs on TanStack Start with articles stored as markdown, one file per language. When I rebuilt the blog I went through this list on every post:

• Answer in the first two sentences. The H1 asks a question and the opening paragraph answers it with a number, a range or a decision. Everything else is support.
• One FAQ block per article. Four or five questions phrased the way people type them, each answered in a paragraph or two. These sections get quoted more than anything else I publish.
• Numbers with a link attached. Every figure carries an inline source in the sentence where it appears, so a model lifting that sentence lifts the attribution with it.
• A separate URL per language. /blog/en, /blog/ru and /blog/hr are three real pages, each with its own canonical and hreflang set. No client-side switcher.
• A visible date on the page. Published and updated dates sit in the article header and in the JSON-LD, because freshness is one of the few signals every engine uses.
• JSON-LD on every post. Article, FAQPage and BreadcrumbList, generated from the same frontmatter that renders the page, so the markup cannot drift from the text.

The sitemap builds itself from the file list, which sounds trivial until you publish in three languages and forget one. I wrote about the same trade-off between scope and finish in what a Telegram mini app costs.

Schema, hreflang and dates that earn their place

Google is direct about schema and generative features: no special markup is required, and no secret type buys you a citation. Article markup still earns its place because it pins down author, headline, datePublished and dateModified in a form every parser reads the same way.

FAQPage is the interesting case. Google's documentation now limits FAQ rich results to well-known government and health sites, so an ordinary site gets nothing visible in Search from that markup. I keep it anyway: the question-and-answer structure underneath is exactly the shape a retrieval system wants, and it costs nothing once it is generated from frontmatter.

Language handling is where small multilingual sites quietly fail. If the Croatian and Russian versions share a URL and swap text in the browser, a crawler sees one page in one language and the other two versions may as well not exist. Separate URLs, correct hreflang pairs and a canonical per page fix that in an afternoon. The same setup runs across my work and services.

How to tell whether any of it worked

Nobody can guarantee inclusion in an AI answer, since the same prompt can return different sources on two consecutive runs. Track inputs and probabilities instead. Search Console carries a generative AI performance report for Google's surfaces, and your server logs show whether GPTBot, OAI-SearchBot, PerplexityBot and ClaudeBot fetch your pages at all.

I keep a fixed list of about twenty prompts a potential client might type, run them across three engines once a month, and record which domains come back. Movement over a quarter tells you more than any single check. Referral traffic from chat domains deserves its own analytics segment, though the numbers stay small for a long time.

The work compounds slowly and survives algorithm changes, because what you are improving is whether your page holds a clear, sourced, quotable answer. My process has the same bias: ship something concrete, measure it, then decide what to change.

FAQ

Do I need an llms.txt file?

Google's optimization guide explicitly lists llms.txt among the tactics that do not help, and no major engine has committed to reading it. Spend the hour on your opening paragraphs instead.

Does schema markup get me cited?

By itself, no. Structured data helps machines parse authorship, dates and question-answer pairs without guessing, which lowers the chance of your content being misattributed or skipped. The citation still comes from the text.

How long before an AI engine picks up a new page?

Anywhere from a few days to a couple of months, depending on indexing speed and how often the retrieval index behind each engine refreshes. Pages answering narrow, low-competition questions tend to appear first, which makes long-tail topics the sane place to start.

One page per language, or one page with a switcher?

One page per language, each with its own URL, canonical and hreflang. Client-side switching hides two thirds of your content from crawlers and makes it impossible to tell which language brings you readers.

Is any of this different from normal SEO?

The technical base is identical: indexable pages, clean structure, real information. What changes is the unit of competition. You write passages that hold up when quoted alone, so effort moves to the first paragraph of every section. If you want a second opinion on a specific page, send it over.