In short
An AI assistant picks its sources in stages: your page must be indexed, then retrieved for the query, then chosen over the other candidates, then actually quoted. Selection favours pages that answer directly, state facts a model can lift cleanly, and come from entities it recognises. Each stage filters out most of the web.
Key points
- Citation is a pipeline, not a lottery. Indexed, retrieved, selected, quoted, in that order.
- Most sites fail at the boring early stages, access and retrieval, not at the glamorous last one.
- Selection rewards direct answers, verifiable claims, recognisable authors and agreement with other sources.
- The Princeton GEO study found quotations, statistics and cited sources lifted visibility by up to about 40 percent.
- Citations are probabilistic. The same question can cite different sources tomorrow, so measure trends, not runs.
The short version
Between your page and an AI answer stand four gates: indexed, retrieved, selected, quoted. Every assistant that cites live sources, ChatGPT search, Perplexity, Google's AI Overviews and AI Mode, Copilot, runs some version of this pipeline. The engineering differs, the shape does not.
Understanding the shape matters because each gate fails differently and is fixed differently. Most advice about "getting cited by AI" is advice about the last gate, offered to sites stuck at the first two.
Gate 1: indexed
Assistants do not search the live web. They search an index: Google's for Gemini and AI Overviews, Bing's plus its own for OpenAI, Perplexity's own crawl for Perplexity. A page missing from the relevant index does not exist for that assistant, whatever its quality.
This is why the Getting found by AI section precedes everything else. Crawler access, JavaScript rendering, indexation, the unglamorous layer, decides eligibility before any content quality is examined.
Gate 2: retrieved
The user asks a question. The system turns it into searches, often several, fetches the top candidates, and splits them into passages. Two things decide whether you are in that candidate set.
You must rank somewhere for something. Retrieval leans on classic search relevance. A page that never appears in the top handful of results for any phrasing of the question is rarely fetched. AI visibility inherits SEO here, directly.
The match happens at passage level. Systems retrieve and rank chunks of pages, not whole pages. A focused page whose passages each say one thing cleanly gives the retriever something to grab. A sprawling page where every paragraph does three jobs blurs into nothing at chunk level, which is the technical reason behind the one page, one question rule this course keeps repeating.
Gate 3: selected
The retriever hands the model a stack of candidate passages. The model, guided by its instructions to answer well and support claims, effectively chooses which to lean on. This is the gate where content quality finally operates, and the influences are consistent across platforms:
- Directness. A passage that answers the question in its first sentence beats one that builds to the answer through background. Models are choosing supporting material, and the passage that is the answer is easiest to use.
- Liftable facts. The Princeton GEO study, the first controlled experiment here, found that adding quotations, statistics and cited sources raised a page's visibility in generated answers by up to about 40 percent in their benchmark. A concrete claim gives the model something to quote and a reason to attribute.
- A resolvable source. Named authors, consistent identity, an entity the model has seen elsewhere. When candidates are close, being someone beats being anyone.
- Agreement. Answers are usually assembled from several sources. A page in line with the consensus, or precise about why it dissents, is safer to cite than an outlier making the same claim loudly.
- Freshness, for queries where it matters. Dated pages, visible updates and current facts win on anything time-sensitive, and matter little on evergreen definitions.
Gate 4: quoted
Selection still is not attribution. Models compress several sources into one answer and name only some of them. What tips a used source into a named one, as far as anyone can tell from the outside: being the origin of a specific number or claim rather than background texture, and phrasing tight enough to quote in one line. A sentence a person could read aloud in eight seconds is a sentence a model can cite cleanly.
There is luck at this gate. There is much less at the first three, which is why the work concentrates there.
Why citations change between runs
Ask the same question twice and the sources can differ. Sampling in generation, slightly different fan-out queries, index updates between runs, all of it moves the outcome. This is not a bug you can optimise away; it is the medium. It also means single observations are noise. Measure by asking a fixed set of questions on a schedule and counting mentions over months, which is the entire respectable version of AI rank tracking.
Common mistakes
- Polishing gate 3 while failing gate 1. The commonest audit finding: citation-optimised content behind a firewall rule that 403s the crawlers.
- Publishing summaries of other people's material. The model can already see the original. A paraphrase adds no fact worth attributing, so it has nothing to win with.
- Burying the answer. Six hundred words of context before the point loses the retriever at chunk level and the model at selection.
- Anonymous authorship. No author, no about page, no consistent identity anywhere else on the web leaves nothing for the entity check to resolve.
- Reading one prompt as a verdict. One run tells you almost nothing in either direction.
When this does not apply
For answers drawn purely from training data, older general knowledge with no retrieval step, none of these gates run, and no page tweak changes what a frozen model already believes. You influence those answers slowly, by being written about across the sources models train on. The pipeline on this page governs the growing share of answers grounded in live search, which is where near-term effort pays.
FAQ
How do AI assistants choose which sources to cite? Through a pipeline: the page must be in the assistant's search index, retrieved as a candidate for the query, selected by the model over other candidates, and finally quoted by name. Selection favours direct answers, concrete verifiable claims, recognisable authors and agreement with other retrieved sources.
What is retrieval in AI search? The step where the system turns a user's question into searches, fetches the top candidate pages from an index, and splits them into passages for the model to work from. If your page is not retrieved, nothing about its quality is ever evaluated.
Why does my site never get cited by AI? Usually a failure at the early gates: crawlers blocked at the firewall, content invisible without JavaScript, pages that rank for nothing, or answers buried mid-page. Check access and retrieval before touching content, because no writing fix repairs an eligibility problem.
What makes content citable? A direct answer near the top, at least one concrete fact worth attributing, a number, a named source or an original observation, phrasing tight enough to lift in one line, and a named author that resolves to a real identity. Pages of paraphrased common knowledge give a model nothing to cite.
Do AI assistants prefer certain websites? They inherit the preferences of their indexes, so authoritative, well-linked sites are retrieved more, and models lean toward sources they can recognise and resolve. But selection happens per passage, which is why a small specialist site with the best direct answer beats a large one with a vague page often enough to be worth playing for.
Does freshness matter for AI citations? For time-sensitive queries, clearly yes: current facts, visible dates and maintained pages win. For evergreen definitions it matters much less. A visible last-verified date is cheap and helps both the reader and the model's preference for maintained sources.
Why are AI citations different every time? Because generation samples, background queries vary, and indexes update between runs. Variance is inherent to the medium, so judge visibility by asking a fixed question set on a schedule and tracking the trend in how often you are named, never by a single run.
Which search index does each assistant use? Google's AI Overviews, AI Mode and Gemini use Google's index. ChatGPT search uses Bing together with OpenAI's own crawling. Perplexity maintains its own index. Copilot uses Bing. Being indexable everywhere is therefore the safe default, which mostly means blocking no major crawler.
Sources
- GEO: Generative Engine Optimization, Aggarwal et al., KDD 2024, the controlled study behind the 40 percent figure
- Google's documentation of AI Mode's query fan-out mechanism, 2025
- Platform documentation and crawler pages for OpenAI, Anthropic and Perplexity, which describe the index and crawler arrangements above
- First published.