Skip to content

AEO and AI search

How do AI assistants choose which sources to cite?

IN SHORT

AI assistants usually answer by retrieving pages that match the question, picking passages that are clear, relevant and trustworthy, and writing an answer from them. How each assistant weighs these signals is not public and changes often, so the reliable strategy is to be the clearest, best-evidenced answer to the question.

By Adnan Gourija, Head of Search Marketing at Webtec (SEO and AEO)Reviewed October 11, 20269 min read

  • Many assistants combine what a model learned in training with pages retrieved live.
  • Pages that crawlers cannot reach, or that are not indexed, usually cannot be retrieved.
  • Clear, self-contained passages with facts, quotations and sources are easier to use.
  • Nobody outside the providers knows the exact weighting, and each provider documents only part of it.

TRY IT

Toggle page qualities and see how quotable a page becomes

how quotable is this page?

30%

Retrievable, with little to make it the page an assistant picks.

What adds up

  • Answer-first opening
  • Cited sources and specific facts
  • Named author and review date
  • Structured data
  • Facts match elsewhere on the web

An illustration of how the qualities add up. Real assistants do not publish their weighting, and no percentage predicts a citation.

Two sources of knowledge

A model carries general knowledge from training, including what it picked up about businesses and brands. Many assistants also retrieve live pages, a method called retrieval-augmented generation, then write from what they found.

Training memory favors well-known, widely mentioned businesses, and it is slow to change because models are updated in releases. Retrieval is where a clear, relevant page can win even if the business is small, because the page is fetched at the time of the question.

Which mode an assistant uses can differ by question. A question about a local plumber or a software comparison usually triggers a search. A question about a general concept may be answered from memory with no sources at all.

What the providers document

Provider documentation tells us how pages get into the retrieval pool, which is the part you can control. Google says a page must be indexed and eligible to appear in Search with a snippet to be shown as a supporting link in its AI features (Google Search Central, December 2025).

OpenAI documents separate crawlers: OAI-SearchBot surfaces sites in ChatGPT’s search features, GPTBot collects content that may be used for training, and ChatGPT-User visits pages when a user asks. Sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers, though they may appear as plain links. Perplexity says PerplexityBot surfaces and links sites in its results and is not used for training, and Anthropic documents ClaudeBot for training, Claude-SearchBot for indexing search results and Claude-User for user requests.

The pattern is the same everywhere: training crawlers and search crawlers are separate, and you can allow one and block the other. The full detail is in AI crawlers and robots.txt.

What gets retrieved

  • Pages an AI crawler is allowed to fetch, by robots.txt and by your firewall or CDN.
  • Pages that match a sub-question, not just the main keyword (see query fan-out).
  • Pages that load and render cleanly for machines, with key content present as text.
  • Pages the provider’s index already knows about, which is why internal links and sitemaps still matter.

What gets chosen

From the retrieved set, assistants favor passages they can use with confidence. No provider publishes a formula, so the list below is what we treat as likely to help, drawn from documentation and from research, not a ranking factor list.

  • A direct answer near the top of the section.
  • Specific, checkable facts with cited sources. In the GEO research paper, adding quotations, statistics and cited sources raised visibility in generated answers by the largest margins among the nine methods tested, while keyword stuffing scored below the unedited baseline.
  • A named author and signs of E-E-A-T.
  • Facts that match across your site and trusted third-party sources, as covered in brand, entity and mentions.
  • Freshness, where the question needs it.

How to write a passage an assistant can use

Write each section so it works alone. Put the answer in the first two sentences, then the reasoning, then the evidence. A passage that depends on three earlier paragraphs to make sense is hard to quote.

Use exact numbers with a source and year: "According to the tax authority’s guidance, the filing deadline is..." beats "usually around the end of the quarter". If you cannot source a number, leave it out and explain in words.

Do not chop pages into artificial fragments. Google’s guide says you do not need to split content into small chunks for its AI features and that one page can cover several topics. Clear headings and self-contained sections are enough.

What we can and cannot know

Providers do not publish their source selection, and it shifts. Anyone who claims a precise formula is guessing. Google also says that third-party tools do not have access to its internal systems.

We rely on what is consistently observable, on provider documentation, and on measurement of real answers over time. Test the assistants your buyers actually use. The difference between this work and classic SEO is covered in AEO vs SEO vs GEO. Measuring AI visibility explains how to do that without fooling yourself.

Examples by industry

Illustrative scenarios, not client results. Pick the one closest to your business.

  • ACCOUNTING AND FINANCIAL FIRMS

    A firm whose blog posts are long and slow to get to the point. Each post opens with three paragraphs of background before it answers anything.

    1. Open every post with a two-sentence answer.
    2. Cite the tax authority’s own guidance for each rule mentioned.
    3. Show the author’s name and credentials on the page.
    4. Make each heading a self-contained question a client would ask.

    Takeaway: A direct answer with a named source is far easier to quote than a long preamble.

  • LAW FIRMS

    A firm whose facts differ from directory to directory. The firm’s name, phone number and practice areas are written three different ways across its site and listings.

    1. Standardize the name, address, phone and practice areas everywhere.
    2. Make the site and the profiles match exactly.
    3. Add Organization schema that states the same facts.

    Takeaway: Assistants trust facts that match across sources, so consistency is part of being chosen.

  • TRADES, CONTRACTORS AND HOME SERVICES

    A plumbing company that blocked AI crawlers by accident. A security plugin set rules that stopped several crawlers from reaching the site, so none of its pages could be retrieved.

    1. Review robots.txt and any firewall rules.
    2. Allow the search crawlers the business wants, and decide on training crawlers separately.
    3. Re-test that key pages can be fetched.

    Takeaway: Nothing else matters if the pages cannot be reached.

THE WEBTEC ANSWER METHOD · STEP 3 OF 6

Answer: how we do it

We write pages answer-first, so each section can stand alone as a quotable passage.

  • Open with the answer in two or three sentences
  • Back claims with sources you can name
  • Add a named author and a reviewed date
See all six steps

Common mistakes

  • Burying the answer under a long introduction, so the passage an assistant needs is never self-contained.
  • Making claims with no source or number behind them, which gives nothing checkable to quote.
  • Blocking AI crawlers by accident, through a plugin, firewall or copied robots.txt rules, so pages cannot be retrieved at all.
  • Contradicting yourself across pages and profiles, which lowers confidence in any one version.
  • Blocking the search crawler while trying to block training, because the two are separate user agents.
  • Assuming one assistant’s behavior applies to all. Each provider runs its own crawlers and index.

Checklist

  • Each key page opens with a direct answer.
  • Claims have sources, with the year where it matters.
  • A named author and review date are visible.
  • robots.txt and your firewall allow the search crawlers you want.
  • Key content is present as text, not only inside images or scripts.
  • Your business facts match everywhere.
  • You have asked each major assistant your top five buyer questions and noted the sources it used.

Questions

Does ranking on Google mean an assistant will cite me?

Not automatically, but it helps. For Google’s own AI features, a page must be indexed and eligible to appear in Search with a snippet to be shown as a supporting link. Other assistants use their own crawlers and indexes. Ranking well is a good sign the page is findable, but test the questions in each assistant your buyers use rather than assuming. A practical check is to ask the assistant the same question three times and note whether your page appears as a source, rather than inferring from your rankings.

Do assistants prefer big brands?

Well-known brands benefit from training memory, because they appear more often in the text a model learned from. Retrieval, however, can surface a smaller business with the clearest, best-evidenced answer to a specific question. The narrower the question, the more room a small business has. Broad questions such as "best CRM" favor large names, while local or niche questions favor whoever answers them most clearly. Build the evidence a smaller brand can control: named experts, specific data, genuine reviews and consistent facts, so the passage is easy to trust even when the name is unfamiliar.

Can I tell which sources an assistant will use?

Not in advance, and not with certainty. Some assistants show their sources, so you can see which pages were used for a given answer, and you can repeat that for your key questions. Treat that as a sample, because answers vary from run to run. Providers do not publish the weighting, and Google warns that third-party tools cannot see its internal systems. A useful habit is to keep a log of the sources shown for your top questions each month. Over time it shows which pages and competitors the assistants rely on.

Should I block AI crawlers?

Only if you have a reason, and then block precisely. Training crawlers and search crawlers are separate. OpenAI says disallowing GPTBot signals that content should not be used for training and does not affect search appearance, while opting out of OAI-SearchBot removes you from ChatGPT search answers. If you want to be recommended, allow the search crawlers, and decide on training separately. Whatever you choose, write it down and check it after every site change, since plugins, firewalls and CDN settings often block crawlers without anyone deciding to.

How fast does a change to my page show up in answers?

It depends on the provider. OpenAI says robots.txt changes can take about 24 hours to affect search results, and Perplexity gives a similar figure. Google says recrawling can take from several days to several months. Content changes follow the crawl, so measure monthly, and do not judge a rewrite after one week. Submitting an updated sitemap and keeping internal links to the changed page helps crawlers find it sooner, but no provider promises a timeline for answers themselves. Check the page in Search Console’s URL Inspection tool to confirm Google has seen the new version.

Is this different for a local business?

Slightly. Google’s guide points local businesses to keep Business Profile details up to date, and the same discipline helps everywhere, so keep your name, address, phone number, hours and services identical on your site and profiles. Then add pages that answer the questions locals ask, such as pricing ranges and service areas. Reviews also matter, so make sure genuine ones mention the specific services you want to be known for. Respond to reviews and keep photos and services current on your profile, since those are the facts people and assistants see first.

Sources

Webtec reviews this page regularly. Search and platform features change, so check provider documentation for the latest. How we research and verify our guides.

Adnan Gourija

Adnan Gourija

Head of Search Marketing at Webtec (SEO and AEO). Leads search marketing at Webtec: the SEO and AEO strategy that gets a business found in Google and named in AI answers.

Related guides

Pick up where this one leaves off.

All AEO and AI search guides
  1. 01Learn guide · AEO and AI searchQuery fan-outQuery fan-out is when an AI system splits one question into several searches. How it works, why it changes content planning, and how to map the sub-questions.
  2. 02Learn guide · AEO and AI searchSchema markup for AISchema markup labels what a page is about so machines read it without guessing. What Google says about AI features, which types to use and what to avoid.
  3. 03Learn guide · AEO and AI searchllms.txtllms.txt is a proposed Markdown file that points AI models to your best pages. What it holds, what Google says, what the data shows and whether to publish one.
  4. 04Learn guide · AEO and AI searchMeasuring AI visibilityMeasure AI visibility with a fixed prompt set, repeated runs and share of answers, tied to referrals and enquiries. The method, the data sources and the limits.
  5. 05Learn guide · AEO and AI searchAI crawlers and robots.txtLearn which AI crawlers to allow in robots.txt, from OAI-SearchBot to ClaudeBot and Google-Extended, what each one does, and how to check that it works.
PreviousAEO vs SEO vs GEONextQuery fan-out
Book a consultation

Explore our services

Start with the one closest to what you just read.

All services

Rated by our clients

Don't take our word for it.

Some of the businesses we have worked with

Your cookies, your call

We use only the cookies this site needs to work. If you say yes, we also turn on Google Analytics to see what works, and our live chat, which records how you use the site so we can help you faster. Nothing optional runs until you choose. See our cookie policy.