ContextBolt SEO Free for 7 days. Keywords, SERPs, backlinks and AI visibility, inside Claude. SEO data inside Claude. Start free trial
Guide · How AI Assistants Choose Citations

How AI Assistants Choose Citations (2026)

An AI assistant doesn’t rank pages and hand you the winner. It breaks the question into several searches, pulls candidate pages for each one, then writes an answer and attaches a citation to the sentences it drew from.

That means the usual “rank well and you’ll get cited” advice only covers the first step. Most pages that go uncited fail at a different step than their owner assumes.

Quick answer
  • Three steps decide it: retrieval, extraction, attribution. Ranking only helps with the first.
  • You are never competing for the question that was typed. Google’s AI Mode splits it into a “multitude of queries” before anything is retrieved.
  • Only 37.9% of AI Overview citations rank in the top 10 for the query, per Ahrefs in March 2026. 31% rank nowhere in the top 100.
  • The unit that gets cited is a passage, not a page. Claude’s web search returns up to 150 characters of cited text per citation.

How do AI assistants choose which sources to cite?

A citation in an AI answer is a link attached to a specific claim in generated text, marking the source the model drew that claim from. It isn’t a ranking position and it isn’t a reward. It’s a receipt.

A ranking answers “which page is best for this query”. A citation answers “where did this sentence come from”. Those two questions often have different winners. The pipeline behind the receipt has the same shape in ChatGPT, Claude, Perplexity and AI Overviews.

StepWhat the assistant doesWhat decides it
1. RetrievalTurns the question into searches and pulls back candidate pagesWhether your page is findable and fetchable. This is the step traditional SEO touches
2. ExtractionReads the candidates and pulls the fragments that answer each partLength, structure and clarity of the passage. Authority barely enters
3. AttributionWrites the answer and links the fragments it usedWhether anything on your page made it into the final text

A page can be retrieved and read and still never appear. Most “why am I not cited” questions are a step 2 problem being answered with step 1 advice.

How many searches does an assistant run before it answers?

More than one, and the question you typed is rarely among them. Google says so directly. Its May 20, 2025 announcement states that AI Mode uses a query fan-out technique, “breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.”

Anthropic documents the same behavior. Its web search tool documentation says the search process “can repeat multiple times throughout a single request”, and gives numbers: “Simple factual queries typically use 1-3 searches; comparative or multientity research can use 10 or more.”

So when someone asks “what is the best bookmark manager for researchers”, nobody searches that string. The assistant searches things like what tools exist, what researchers need, what each costs, what people complain about, and which are still maintained. Each one runs its own retrieval and pulls its own candidates.

You’re competing for one sub-question inside the query. A page that is the tenth-best overview of a topic but the clearest answer to “what does it cost” can win the citation while the definitive guide gets nothing, because the guide never separated that answer out cleanly enough to lift. A fan-out has no keyword, only a question shape.

Does ranking first still get you cited?

Less than it used to. Ahrefs analyzed 863,000 keyword SERPs and 4 million AI Overview URLs and published the breakdown on March 2, 2026. In July 2025, the same team measured 76.10% overlap with the top 10.

Where the cited page ranksShare of AI Overview citationsWhat it means for you
Top 1037.9%Ranking still helps. It is no longer the whole game
Position 11 to 10031.2%Page two content gets quoted constantly
Outside the top 10031.0%Retrieved by a sub-query you never targeted

Ahrefs attributes part of the shift to better parsing on their side and part to the fan-out itself, so some of the change is measurement and plenty of it is real.

The bottom row is the most useful number here. Roughly a third of citations go to pages that don’t rank for the query at all, retrieved by a sub-query nobody targeted. A keyword tool can’t chase that. You get it by answering narrow questions completely, on their own pages, in their own words.

Which crawler has to reach your page?

Retrieval has a prerequisite: someone’s crawler has to have your page, and it’s often not the one you were thinking about. The AI companies run separate bots for separate jobs. OpenAI’s bot documentation states that “OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features”, while GPTBot “is used to crawl content that may be used in training our generative AI foundation models”.

Perplexity splits it the same way. Its crawler documentation describes PerplexityBot as “designed to surface and link websites in search results on Perplexity”, adding that “it is not used to crawl content for AI foundation models”. Perplexity-User fetches a page live because a person just asked, and “since a user requested the fetch, this fetcher generally ignores robots.txt rules”.

AssistantCrawler that feeds citationsSeparate training crawler
ChatGPT searchOAI-SearchBotGPTBot
ChatGPT, live fetchChatGPT-Usern/a, user-triggered
PerplexityPerplexityBotNone. Says it does not train on crawls
Perplexity, live fetchPerplexity-Usern/a, ignores robots.txt by design

The failure is common and silent. A site adds a blanket AI block in robots.txt to keep its writing out of training data, catches the search bots in the same rule, and drops out of a citation pool without a single ranking moving. Nothing in Search Console reports it. The full user-agent list is in our AI crawlers list, and the file-level detail is in llms.txt vs robots.txt.

Does an AI assistant cite a page or a passage?

A passage. Anthropic’s web search documentation describes what a citation object contains. Citations are always on for web search, and each one carries a url, a title, and a cited_text field holding “up to 150 characters of the cited content”. The page and the section never come into it.

A citation is roughly one sentence long. The model isn’t deciding your page is good. It’s deciding that one sentence on your page is the cleanest way to state one fact, out of every sentence it retrieved. A 4,000-word guide is a few hundred candidate sentences, and most are useless for citation because they only make sense with the paragraph above them. A model can’t use “as we saw in the previous section, this drops to about half”. It can use “Claude’s web search is priced at $10 per 1,000 searches”.

The habits that get you cited are different from the ones that get you shared.

  • Self-contained sentences beat flowing prose. Each claim should survive being lifted out of the page with no context around it.
  • Named numbers, dates and versions beat adjectives. “38% of citations” is quotable. “A significant share” is not.
  • One question per section. If a heading covers three questions, the answer to each is diluted into the other two.
  • Front-load the answer. The sentence that answers the heading should be the first sentence under it, not the conclusion of the argument.

It’s also what both sides of the content chunking for SEO argument end up recommending.

Why do assistants cite pages that are obviously worse than yours?

Extraction and quality are different tests, and only one runs at that moment. At step 2 the model has eight candidate pages open and a sub-question to answer. It isn’t judging which site knows the most. It’s scanning for the passage that answers this narrow thing most directly. A thin page with a clear one-line answer beats a thorough page that reaches the same answer in paragraph nine.

There is a second reason, and it is less flattering to all of us. Assistants favor sources whose claims are checkable, because a checkable claim is a safer thing to repeat. A page with a specific figure, a date, and a named source is low-risk to quote. A page with confident prose and nothing to verify is a liability the model routes around, however good the prose is. When a draft started with a chatbot, running it through an AI detector is a fast way to find the smooth but empty passages that need a real figure or source added.

The old authority work still pays, indirectly. Being named on other sites, having a real author and citing your own sources all raise the odds a model treats you as safe to repeat. Answer engine optimization covers that layer, and the assistant-specific tactics are in how to rank in ChatGPT.

What changes your citation odds?

Ordered by how much they move and how fast:

  1. Be in the candidate set at all. Allow the search crawlers, keep pages fetchable without JavaScript, and rank somewhere respectable for the sub-questions. Position 15 is in the game now. Position 90 mostly is not.
  2. Split questions onto their own pages. One narrow question answered completely beats the same question buried in a broader guide.
  3. Write liftable sentences. Answer in the first line under each heading and assume every paragraph will be read alone.
  4. Put checkable specifics in. Numbers, dates, versions, prices, user-agent strings, named sources.
  5. Keep it current. Claude’s search results include a page_age field showing when a page was last updated, so the age of a page is data the model sees.

Publishing an llms.txt file won’t do it, and neither will schema markup on its own. Google’s AI features guidance says you don’t need to create new machine-readable files or AI text files to appear in AI Overviews or AI Mode. The work is in the pages.

How do you know whether any of it worked?

Asking an assistant about yourself once tells you nothing. Answers vary between runs and they’re personalized. What you need is the same set of prompts run repeatedly across assistants, with your brand’s mention rate tracked against named competitors over time. Doing that by hand collapses within a week.

ContextBolt SEO does it from inside the agent you already work in. You connect one MCP URL to Claude Code, Claude Desktop or Cursor, then ask in plain language whether AI tools are naming you, and it reports your AI visibility and share of voice against competitors alongside normal keyword, rank, backlink and SERP research. Every lookup lands on your SEO Dashboard so the research is still there next week. It is $35 a month with a 7-day free trial. For a single reading with no signup, the free AI visibility checker will give you one, and best AI visibility tools compares the other trackers.

For the numbers behind source selection, see the current citation studies.

Where next

Frequently asked questions

Something else on your mind? Get in touch.

How do AI assistants choose which sources to cite?

In three steps. The assistant rewrites your question into several searches, retrieves candidate pages for each one, then generates an answer and attaches a citation to the specific sentences it drew from. Being retrievable, quotable and checkable matters at a different step each time.

Does ranking number one get you cited by AI?

It helps, but far less than it used to. Ahrefs analyzed 863,000 keyword SERPs and 4 million AI Overview URLs in March 2026 and found only 37.9% of cited pages also appeared in the top 10. Another 31% ranked nowhere in the top 100.

Do AI assistants cite whole pages or single passages?

Passages. Anthropic's web search documentation shows each citation carrying up to 150 characters of cited text from the source. The unit that gets quoted is a fragment, so a page that answers one question cleanly in one paragraph beats a page that answers it across four.

Which crawler has to reach my site for AI citations?

It depends on the assistant. OpenAI uses OAI-SearchBot to surface sites in ChatGPT search, separately from GPTBot for training. Perplexity uses PerplexityBot for its index. Blocking the search crawler removes you from that assistant's citation pool entirely. See our AI crawlers list.

Why does an AI assistant cite a worse page than mine?

Because the model is picking the source it can quote most cleanly for one sub-question, not ranking the best page on the topic. A thin page with a direct, self-contained answer beats a thorough page that buries the same answer in paragraph nine.