How do AI answer engines choose what to cite? A reading list of primary sources only

None of the nine vendor documents on this list discloses a formula for choosing what to cite. They describe crawler access, eligibility, stated ranking principles, and reporting. None of it tells you why an engine picked a particular page for a particular answer.

An answer engine, such as ChatGPT search, Perplexity, Microsoft Copilot, or Google’s AI Overviews and AI Mode, is an AI system that answers in its own words and links to the sources it used. The three research papers below cover one retrieval design, one audit of citation support, and one test of what changed measured visibility. Each supports a bounded finding. None predicts whether your page will be cited.

Why read primary sources only?

A primary source is the document closest to the claim: a company’s own documentation of its product, or the paper that ran the experiment. A summary can help you find the original, but it cannot replace it. Each retelling can drop a condition, a date, or a sample size, and any of those can change what a finding means.

There is a second boundary. Company documentation is primary evidence of what the company says. It is not evidence of what the product does on your page; that needs observations of its own. A research paper has the same limit: its results belong to the setup it tested.

This list gives each original, what it supports, and where that support ends. The companion guide on how answer engines choose sources turns the same material into a playbook.

The reading list

Every document was opened on 24 September 2026, when the quotations were checked and Google’s “Last updated” stamps recorded. Vendor pages change without notice, so reopen the passage you rely on before you act on it.

Research papers

1. Lewis and colleagues, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (NeurIPS 2020). arXiv 2005.11401. Tells you: one design for retrieval-augmented generation (RAG), in which a retriever finds passages and a generator writes the answer from them. Its retrieval index was Wikipedia. The authors listed “providing provenance for their decisions” as an open research problem. Does not tell you: how a commercial answer engine retrieves from the open web or chooses what to cite.

2. Liu, Zhang, and Liang, “Evaluating Verifiability in Generative Search Engines” (Findings of EMNLP 2023). arXiv 2304.09848. Tells you: a human evaluation of Bing Chat, NeevaAI, perplexity.ai, and YouChat found that, on average, “a mere 51.5% of generated sentences are fully supported by citations and only 74.5% of citations support their associated sentence.” Does not tell you: how today’s products perform. The figures describe the systems and queries in that study.

3. Aggarwal and colleagues, “GEO: Generative Engine Optimization” (KDD 2024). arXiv 2311.09735. Tells you: the authors built a benchmark of 10,000 queries from nine sources and a test engine that answers from the top five Google results with gpt-3.5-turbo. In that setup, adding citations, quotations, or statistics to a source raised its measured visibility by up to 40%, a relative gain on the paper’s own metric. Keyword stuffing did not perform well. A separate test on Perplexity used 200 test-set samples and uploaded files. Does not tell you: how ChatGPT, Google, or Perplexity choose sources from the live web.

Google

4. “AI features and your website” (last updated 10 December 2025). Google Search Central. Tells you: to be eligible as a supporting link in AI Overviews or AI Mode, “a page must be indexed and eligible to be shown in Google Search with a snippet”, and there are no additional technical requirements. Both features may use “query fan-out”: several related searches run for one question. Does not tell you: which eligible page will be selected. Google also states: “Indexing and serving isn’t guaranteed.”

5. “Optimizing your website for generative AI features on Google Search” (last updated 10 July 2026). Google Search Central. Tells you: Google says these features are “rooted in our core Search ranking and quality systems”. By its account, Google Search does not use llms.txt or other special AI files, and content does not need to be split into small pieces. On length: “There’s no ideal page length.” For measurement, the guide names a Generative AI performance report in Search Console; the December 2025 AI features page puts AI feature traffic in the overall Web report. Does not tell you: how other engines work, or which page Google will cite.

6. “Creating helpful, reliable, people-first content” (last updated 10 December 2025). Google Search Central. Tells you: Google’s suggested self-assessment asks who made the content, how, and why. The page states that “E-E-A-T itself isn’t a specific ranking factor.” E-E-A-T stands for experience, expertise, authoritativeness, and trustworthiness. Does not tell you: a checklist that guarantees selection.

OpenAI

7. “Overview of OpenAI crawlers.” OpenAI. Tells you: OpenAI uses OAI-SearchBot to surface sites in ChatGPT search. Sites that block it “will not be shown in ChatGPT search answers, though can still appear as navigational links.” GPTBot is a separate crawler for model training. OpenAI says its search systems can take about 24 hours to reflect a robots.txt change. The site’s GPTBot vs OAI-SearchBot guide separates the three OpenAI bots. Does not tell you: how ChatGPT ranks what it retrieves.

8. “Searching the web with ChatGPT.” OpenAI Help Center. Tells you: the nearest thing to a ranking statement among OpenAI’s documents here: “ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.” The page also warns that citations “can be incomplete, outdated, or incorrect.” Does not tell you: what those factors are, or how they decide any particular citation.

9. “Publishers and Developers - FAQ.” OpenAI Help Center. Tells you: at the 24 September 2026 check, the FAQ advised publishers who want to appear in ChatGPT’s summaries and snippets not to block OAI-SearchBot. It also said ChatGPT adds utm_source=chatgpt.com to the links it sends people through, so those visits can be identified in your analytics. Does not tell you: how often your page was cited when nobody clicked.

Perplexity, Microsoft, Anthropic

10. “Perplexity Crawlers.” Perplexity. Tells you: PerplexityBot surfaces and links sites in Perplexity’s search results and “is not used to crawl content for AI foundation models.” Perplexity-User fetches pages when a person asks Perplexity a question, and “generally ignores robots.txt rules.” Perplexity says changes may take up to 24 hours to register. Does not tell you: how Perplexity ranks or chooses its sources.

11. “Introducing AI Performance in Bing Webmaster Tools Public Preview” (10 February 2026). Microsoft Bing. Tells you: Microsoft describes a report that counts citations of your pages in Microsoft Copilot, AI summaries in Bing, and some partner products. It includes a sample of “grounding queries”: the phrases the AI used to retrieve the content it cited. Does not tell you: why a page was chosen. Microsoft says its cited-pages figure “does not indicate ranking, authority, or the role of any page within an individual answer.”

12. “Does Anthropic crawl data from the web, and how can site owners block the crawler?” Claude Help Center. Tells you: Anthropic separates three bots. ClaudeBot collects content that could contribute to model training, Claude-User fetches pages for people using Claude, and Claude-SearchBot is for search. Anthropic says blocking Claude-SearchBot “may reduce your site’s visibility and accuracy in user search results.” Does not tell you: how Claude chooses among sources.

Assess the sources with TRACE

TRACE is the framework from my book The Verified Research Loop, which is available now in English and Dutch. It turns a research question into a decision record that someone else can inspect. You do not need the book to apply it here. The five steps:

1. Target. Write the research brief first: the decision, the bounded question, scope, exclusions, deadline, and stop rule, before a single search runs.
2. Retrieve. Obtain the documents themselves, and log what stayed inaccessible, instead of trusting one confident summary.
3. Assess. Give each source a role (Primary, Secondary, Advocacy, or Grey), then appraise what it can bear for this specific claim.
4. Cross-check. For each claim the decision rests on, record one relation per source and map shared origins. Five retellings of the same Google page are one origin, not five confirmations.
5. Export. Hand over a versioned decision memo, traceable to the claim ledger, that a co-founder or investor can act on without redoing the work.

Here is Assess applied to this list and to the commentary around it:

What you are reading	Role in TRACE	What it can support
A company’s own documentation	Primary, for what the company says	Its stated controls, eligibility rules, and promises. Behaviour on your page needs evidence of its own.
A research paper	Primary, for its own experiment	Results within that setup and those dates. Transfer to a live product needs checking.
A news story or blog post about either	Secondary	Reporting or interpretation. Check it against the original before you repeat it.
A study by a firm that sells AEO or GEO services	Advocacy	Possibly useful data, read with its method and its commercial incentive in view.
A screenshot thread or anonymous test	Grey	A lead. Check its provenance and context before it supports anything.

A role routes your attention. It is not a credibility score, and what a source can support is settled one claim at a time. A vendor document can prove that the vendor made a promise without proving that your page benefits from it. The TRACE framework page has the full routine.

How do you read a vendor study?

The same distinction applies to research from firms that sell AEO or GEO services. Having original data does not exempt a study from appraisal. Start with six questions:

1. Who paid, and do they sell the fix? If so, the study has an Advocacy role. That does not make it wrong. Appraise the data and the method with the incentive in view.
2. What was counted? A mention (your name in the answer), a citation (a link to your page), and a click are three different events.
3. Which engine, mode, location, and dates? Was search switched on? Which country and language?
4. Which prompts? How many, who wrote them, and are they published?
5. What could someone else check? Look for raw data and enough method detail to rerun the analysis. Record what is missing.
6. Does the headline claim a cause? “Pages with X get cited more” shows that two things occur together. On its own, it does not show that X caused the citations.

Google’s optimization guide adds a limit of its own: “No third-party tool has access to our internal ranking or AI systems.”

Try this today (20 minutes)

Pick one claim you are about to act on, such as “FAQ schema gets you cited” or “llms.txt helps with Google.” Spend the time on a single claim record:

1. Write the claim and the action it would justify. Name the engine and the change you are considering for your page.
2. Open the original. Use the list above to find it. Record the URL, the passage, the document’s date, and the date you checked.
3. Label the source and its reach. Give it a role. Does the passage support an eligibility rule, a stated policy, or a measured result? Does it say anything about your page at all?
4. Record the gap and the decision. If the source supports only a narrower claim, rewrite the claim to match. If you find no support, mark it “unsupported by the sources checked” and hold the change until evidence turns up.

For the llms.txt example, entry 5 gives you Google’s own statement that Google Search does not use it. Keep the record with the task. A missing source leaves a claim open; it does not prove it false.

Cite this:How do AI answer engines choose what to cite? A reading list of primary sources only.Len P. van der Hof. https://lenvanderhof.com/en/blog/answer-engine-reading-list/ · Published 17 October 2026.