When an AI assistant answers with web search, it works a bit like a researcher in a hurry. It searches, reads a handful of pages, writes a summary and links to some of what it read. No engine publishes its full method, and each one works differently. But the broad steps are documented by the companies themselves and can be seen in the answers. They tell you a lot about what makes a page easy to cite.
This guide sticks to that documented behavior and says where engines differ. Anyone who claims to know the exact weight each engine gives to each factor is guessing.
Step 1: deciding whether to search
Not every question sends an assistant to the web. For general or timeless questions, it may answer from what its model learned in training. For questions about products, prices, places, recent events or "the best" of something, it is more likely to search. ChatGPT, for example, decides for itself when to search, and users can also ask it to.
Google's AI Overviews add another filter. Google shows them only when its systems judge that they add something to the regular results, so many searches show none at all.
This matters for your strategy. A brand named without any search comes from the model's training data, which you can only influence slowly, through what is published about you across the web. A brand named in an answer that used search comes, at least in part, from the pages the assistant just read. Those you can influence much faster.
Step 2: fan-out, one question becomes several searches
Assistants rarely search for your exact question. They rewrite it into several narrower searches, often called fan-out queries, and combine what they find. Google says both AI Overviews and AI Mode may use this technique, running related searches across subtopics and data sources. Other assistants sometimes show the searches they ran next to the answer, and those lists follow the same pattern.
An example. The prompt "best family hotel in Lisbon" might fan out into searches like:
- family friendly hotels in Lisbon with a pool
- Lisbon hotels with family rooms
- best area to stay in Lisbon with kids
- Lisbon hotel reviews from families
Two things follow. Your page can be found through a sub-question you never targeted, such as which neighborhood suits families. And a page that only repeats the headline question, without the questions behind it, gives the assistant less to work with.
Step 3: reading a handful of pages
From the search results, the assistant reads a small set of pages, or passages from them. What it can read depends on where each engine searches:
- Google AI Overviews and AI Mode draw on Google's own index and search results. That is why classic Google rankings still matter there.
- ChatGPT surfaces sites in its search answers with the help of OpenAI's crawler, OAI-SearchBot. OpenAI says sites that opt out of it won't be shown in ChatGPT search answers.
- Perplexity runs its own crawler, PerplexityBot, to find and link sites in its answers.
- Gemini can ground its answers in Google Search.
- Claude can search the web when web search is turned on, and cites the pages it used.
Pages that are hard to fetch can be missed: a page blocked in robots.txt or by a firewall, a page that times out, or a page whose main content only appears after scripts run. Not every crawler runs JavaScript.
Step 4: writing the answer and choosing citations
The model writes its answer from what it read and links some of those pages, inline or as a numbered list of sources. A few things are worth knowing:
- Not every page the assistant read ends up cited.
- The brands named in the answer aren't always the ones on the cited pages. Some come from the model's training data.
- The same prompt can produce different sources on different days, because the searches, the results and the writing all vary.
- Citations point to specific pages, not whole sites. A site that is strong in general still needs a page that answers this particular question.
Google is a special case
Google's AI features are built into Google Search. Its guidance for site owners is direct. To be eligible as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown with a snippet. There are no additional technical requirements, and the usual SEO fundamentals apply. Snippet controls work here too: nosnippet, data-nosnippet, max-snippet and noindex limit what Google can show from a page, including in its AI features.
So for Google, rankings are the first gate. A page that doesn't rank for the question or its sub-questions is rarely in a position to be cited. How to get cited in Google AI Overviews covers the practical steps.
What makes a page easy to cite
None of the engines publishes a list of citation factors. What follows isn't a secret formula. It is what the engines' own guidance points to, and what you can see in the pages they cite. None of it is a guarantee.
A clear, direct answer
Put the answer in the first lines of the page or section, under a heading that matches how people ask the question. Then add the detail. A passage that makes sense on its own can be quoted. One that depends on three paragraphs of setup is harder to use.
Specific facts
Prices, sizes, opening hours, dates, steps, named options. "Open daily from 7 a.m. to 6 p.m." gives an assistant something to use; "convenient opening hours" doesn't. Vague claims give the model nothing to cite and nothing to check against other sources.
Comparisons, lists and tables
Many prompts are comparisons in disguise: "best X for Y", "X or Y", "alternatives to X". The answers to them are usually lists of options, each with a reason. Pages that compare options against clear criteria, including options that aren't yours, are natural sources for those answers. Lists and tables are also easy to extract.
Current information
For anything that changes, such as prices, product ranges, rules or yearly "best of" lists, an old page is a weak source. It can also lead to an answer that is wrong about you. Update facts when they change, remove claims that are no longer true, and show when the page was last updated.
Presence where assistants already look
Assistants cite a mix of site types: editorial reviews and buying guides, community threads on sites such as Reddit, Quora and YouTube, reference pages, institutional sites and company sites. If the pages cited for your questions are mostly reviews and forum threads, improving your own site may not change the answer. Being mentioned, accurately, on the pages assistants already trust is part of the work. A mention on a cited page can get you named even when your own page isn't cited.
Pages crawlers can reach
Check that robots.txt, your CDN and your firewall allow the search crawlers you care about: Googlebot for Google's AI features, OAI-SearchBot for ChatGPT search, PerplexityBot for Perplexity. Blocking training crawlers such as GPTBot is a separate decision, and OpenAI treats the two independently. Keep important content in the page's text, not only in images or behind scripts.
What you can't control
Some things stay out of your hands:
- Engines change their models and methods without notice.
- Answers vary with the exact wording, the market, the language and the day.
- Part of what a model says comes from its training data, which changes only when the model is updated.
That is why it pays to measure instead of guess. Ask the same prompts every day, in each market, and look at which sources are cited across many answers.
How to see this for your own questions
Take one important prompt and read the answers from several engines. List the cited sources and sort them by type. Mark the ones that name a competitor and not you. Over a few weeks, the pages that keep coming back form your shortlist. Some are your own pages to improve. Others are third-party pages where you want to be mentioned.
How GoGetCited helps
GoGetCited records every answer in full with the sources it cites, typed as editorial, UGC or community, reference, institutional, corporate, competitor or your own site, and shows the fan-out queries when an engine reports them. Sources ranks the domains and pages cited for your prompts and shows the gap: where competitors are cited and you aren't. AI traffic shows which AI crawlers visit your site and how many people arrive from AI assistants.