AI Citations: How Do AI Models Choose Which Websites to Cite?
Updated: October 2026 | ORBA Insights
AI citations are becoming an increasingly important part of online discovery. When platforms such as ChatGPT, Google AI search experiences and Microsoft Copilot generate an answer using information from the web, they can reference the websites that helped support that response.
For businesses, being one of those sources creates a new type of visibility.
But how does an AI system decide which websites to retrieve? Why does one source appear in an AI-generated answer while another doesn’t? And can businesses improve their chances of being understood and selected?
The answer is more complicated than simply “ranking higher.”
Current research suggests that AI citation visibility can involve several stages — from discovering and retrieving a page to deciding whether information from that page is useful enough to support the final answer.
Understanding those stages helps explain why AI citations are becoming an important measurement alongside rankings, impressions and website traffic.
What are AI citations?
AI citations are references or links to external sources that an AI system provides to support information contained within a generated response.
For example, someone might ask:
“What should a small business consider when choosing accounting software?”
An AI search system could retrieve information from several websites, synthesise that information into an answer and provide links to some of those websites as supporting sources.
Different platforms display sources differently.
ChatGPT Search can provide inline citations and a Sources panel containing links to relevant web pages.
Google’s AI-powered Search experiences can provide clickable links to supporting websites.
Microsoft’s AI search experiences can similarly ground generated answers in web content and surface sources.
The underlying idea is important:
The user may receive an answer before deciding which website to visit.
That creates a different discovery journey from conventional search.
AI citations don’t work exactly like traditional Google rankings
Traditional search is usually understood as a ranked list.
A user enters a query and the search engine determines which pages should appear and in what order.
Generative AI introduces another layer.
An AI system may retrieve several potentially relevant pages and then use information from some of those pages to construct an answer.
That means several different events can occur:
1. Your page is discovered.
2. Your page is retrieved as potentially relevant.
3. Your page is selected as a source.
4. Your website receives an AI citation.
5. Information from your page actually influences the generated answer.
These are related outcomes — but they aren’t necessarily the same thing.
A website can potentially be retrieved without being cited.
And even being cited doesn’t necessarily tell us how significantly the information influenced the final answer.
That distinction is becoming increasingly important in Generative Engine Optimisation (GEO).
Citation selection and citation influence are different
One of the more interesting developments in 2026 GEO research has been the distinction between citation selection and what researchers describe as citation absorption.
A 2026 study analysing citation behaviour across ChatGPT, Google AI Overview/Gemini and Perplexity examined more than 21,000 search-layer citations.
The researchers proposed two separate stages:
Citation selection — whether the AI platform chooses a webpage as a source.
Citation absorption — whether information from that source actually contributes language, evidence, structure or factual support to the generated answer.
This distinction matters.
Simply counting how many times a website appears as a citation may not tell businesses how influential that content actually is within AI-generated responses.
The researchers found that pages with stronger citation influence tended to be more structured, semantically aligned with the query and richer in extractable evidence, including definitions, numerical facts, comparisons and procedural information.
How do AI models find websites to cite?
There isn’t one universal process used by every AI platform.
Different systems can use different search providers, indexes, retrieval technologies and ranking mechanisms.
However, one important concept is retrieval-augmented generation, usually abbreviated to RAG.
RAG allows an AI system to retrieve information from external sources before generating an answer.
Google explains that its generative Search experiences use retrieval-augmented generation to retrieve relevant, current pages from Google’s Search index before using information from those pages to help generate responses.
Google can also use query fan-out, where a complex question triggers multiple related searches.
This means a user’s single question can potentially cause an AI system to investigate several related concepts.
The result is that a page doesn’t necessarily need to match only the exact wording of the original question.
It needs to contain information that is relevant to the underlying information need.
Relevance appears to matter — a lot
So what actually increases the likelihood of receiving AI citations?
Research is still developing, and there is no guaranteed formula.
However, relevance repeatedly appears as an important factor.
A large 2026 controlled study investigated citation selection across six large language models using 252,000 trials.
The researchers tested 18 different content factors while controlling which two sources the models could choose between.
Their results found that topical relevance was one of the strongest factors associated with being cited first.
Other factors — including recent timestamps, explicit pricing information in relevant contexts, completeness and trust-related signals — also produced effects within the experimental environment.
Formatting changes alone were considerably less influential.
This reinforces an important principle:
Trying to make a page “look optimised for AI” is unlikely to compensate for information that isn’t genuinely relevant to the question being answered.
What makes a website useful as an AI source?
No checklist can guarantee an AI citation.
However, the available research and guidance point towards several characteristics that make web content more useful for both people and retrieval systems.
1. Answer the actual question
This sounds simple, but many commercial websites avoid giving direct answers.
Consider someone searching:
“How long does commercial solar panel installation take?”
A page could spend several paragraphs explaining why solar energy is important before eventually discussing installation times.
Or it could say:
“Commercial solar installation timelines vary by system size, site complexity, approvals and grid requirements. Smaller projects may take days to install on-site, while larger projects can require considerably longer planning and installation periods.”
The second version establishes the answer immediately and can then provide the necessary detail.
The objective isn’t to write for machines.
It’s to make useful information easy to locate.
⸻
2. Make factual information extractable
AI systems often need specific pieces of evidence.
That could include:
-definitions
-statistics
-prices
-dates
-comparisons
-processes
-technical specifications
-research findings
-step-by-step instructions
Clear information is easier to identify than important facts hidden inside vague marketing language.
If you have original evidence — such as proprietary research, survey results, case-study outcomes or first-party data — present it clearly and explain how it was produced.
⸻
3. Stay tightly relevant to the topic
More words don’t automatically create better AI visibility.
A 3,000-word article filled with loosely related information may be less useful for a particular question than a focused 1,200-word resource containing the information needed to answer it.
Create depth where depth genuinely helps.
Avoid adding sections solely to increase word count.
⸻
4. Support claims with credible sources
If you make a factual claim, particularly one involving statistics, research or another organisation’s technology, link to the evidence.
Where possible, favour primary sources.
If you’re explaining a Google Search feature, Google’s documentation should generally be the starting point.
If you’re reporting government statistics, use the relevant government dataset.
If you’re discussing academic research, reference the research itself.
External references aren’t magic GEO signals.
They make your content more transparent, verifiable and useful.
⸻
5. Keep important information current
Outdated information can reduce the usefulness of a source.
This is particularly important for topics that change quickly, such as:
-software pricing,
-regulations,
-technology,
-statistics,
-product specifications,
-AI platforms themselves.
Review important evergreen pages periodically and clearly update information when circumstances change.
⸻
6. Give pages clear structure
Use descriptive headings.
Keep related information together.
Use tables when genuine comparisons are involved.
Use lists when presenting steps or criteria.
Define technical terms before relying on them.
And avoid burying the most important answer beneath unnecessary introductory copy.
Research into citation influence has found an association between structured, evidence-rich pages and stronger contribution to generated answers.
That doesn’t mean adding headings will magically generate citations.
It means good information architecture makes useful information easier to identify.
Can ChatGPT cite any website?
Public websites can potentially appear in ChatGPT Search, but technical accessibility matters.
OpenAI’s current publisher guidance states that public websites can appear in ChatGPT search results.
For website content to be discoverable and included in summaries or snippets, OpenAI says publishers should ensure they aren’t blocking OAI-SearchBot.
This makes technical accessibility part of AI discoverability.
A business can publish excellent information, but if relevant systems cannot access the page, that information may not be available for retrieval through that route.
OpenAI also notes that search citations can occasionally be incomplete, outdated or incorrect and recommends checking the original sources when accuracy matters.
Source: OpenAI — Publishers and Developers FAQ
Google says strong SEO foundations still matter
Businesses shouldn’t interpret the growth of AI citations as a reason to abandon SEO.
Google’s official 2026 guidance says that its generative AI features are rooted in its existing Search ranking and quality systems.
Google recommends continuing to focus on fundamentals such as:
-crawlability and indexability
-useful, unique content
-good page experience
-clear internal linking
-important information available as text
-accurate structured data
-high-quality supporting images and video where appropriate
Google also warns businesses to be cautious about third-party tools or services claiming that they can guarantee performance within AI experiences.
No external GEO platform has access to Google’s internal ranking systems.
That includes ORBA.
It’s to help businesses identify and improve the things they can control.
Microsoft is beginning to measure AI citations directly
Perhaps one of the clearest signs that citations are becoming their own visibility metric came from Microsoft in 2026.
Bing introduced AI Performance within Bing Webmaster Tools.
The reporting shows website owners how their content appears across supported AI experiences, including Microsoft Copilot and AI-generated Bing experiences.
Metrics include:
Total Citations — how many citations from the website appeared as sources in supported AI-generated answers.
Average Cited Pages — the average number of unique pages from the site cited each day.
Microsoft explicitly cautions that these metrics don’t indicate ranking or authority.
That’s an important distinction.
A citation is a visibility signal.
It isn’t automatically an endorsement, recommendation or ranking.
Should businesses optimise specifically for AI citations?
Businesses should optimise for usefulness and machine clarity, not chase citations as if they were another keyword-ranking trick.
A sensible strategy is to ask:
Can search and AI systems access the page?
Is it immediately clear what the page is about?
Does it answer a real question?
Are important claims supported?
Does the page contain original or useful information?
Can individual facts be understood without reading the entire website?
Is the information current?
Is the organisation behind the information clear?
Those improvements benefit conventional search, AI retrieval and — most importantly — the people using the website.
Why doesn’t AI always cite the same websites?
AI citations can change because generative search systems are dynamic.
Different wording can change the information retrieved.
A follow-up question can change the context.
Sources can be updated.
Indexes change.
Models and retrieval systems are updated.
Platforms may use different search providers and retrieval methods.
Even similar questions can therefore produce different source selections.
This is one reason businesses should be cautious about running one AI query and concluding:
“We’re visible in AI.”
Or:
“AI doesn’t know our business.”
A single response is an observation.
Repeated testing across relevant questions provides a much more useful picture.
AI citations shouldn’t be measured alone
Citation counts are useful, but they’re only one part of AI visibility.
Businesses should consider several signals together:
AI mentions — Is the business mentioned when relevant questions are asked?
AI citations — Is the website provided as a source?
Citation frequency — Does this happen repeatedly across relevant questions?
Interpretation accuracy — Does the AI correctly understand what the business does?
Source pages — Which pages are actually being selected?
AI referral traffic — Are users visiting after interacting with an AI platform?
Citation influence — Where measurable, does information from the page meaningfully contribute to the answer?
This creates a much more complete picture than simply counting links.
Frequently asked questions
What are AI citations?
AI citations are links or references to external sources that support information contained within an AI-generated answer. They can help users identify where information originated and explore the source in more detail.
How can ORBA help improve my website for AI search?
ORBA scans your website and identifies opportunities to make its content easier for AI systems to understand. It provides an AI visibility assessment, identifies gaps and recommends practical improvements businesses can make to strengthen their presence across AI-powered discovery.
Can ORBA tell me if AI understands my website?
ORBA analyses your website to show how clearly its content communicates your business to AI systems. It identifies what AI can understand from your website, highlights information that may be unclear or missing, and provides practical recommendations to improve your website’s AI visibility.
Why do AI models cite websites?
AI search systems can use external web sources to provide current information, ground generated responses and allow users to examine supporting material.
How do I get AI citations for my website?
There is no guaranteed method. Businesses can improve accessibility, topical relevance, content clarity, evidence, structure and freshness, but individual AI platforms ultimately determine which sources they retrieve and cite.
Does ranking first on Google guarantee an AI citation?
No. Traditional search ranking and AI citation selection are related but distinct processes. AI systems can retrieve multiple sources and decide which information to use based on the particular question and system.
Can ChatGPT cite my website?
Yes. OpenAI says public websites can appear within ChatGPT Search. Website owners should ensure that relevant content isn’t unintentionally blocked from OAI-SearchBot if they want it to be discoverable through ChatGPT Search.
From being found to being useful
Traditional search taught businesses to ask:
“Can people find our website?”
Generative search adds another question:
“Is our information useful enough to become part of the answer?”
That is why AI citations matter.
They provide one indication that information from a website is participating in an AI-driven discovery experience.
But citations shouldn’t become another metric businesses chase without context.
The bigger objective is making sure your website clearly communicates:
who you are,
what you do,
who you serve,
what you know,
and why the information you publish can be useful.
Because before an AI system can cite your business, it first needs to find and understand the information you’ve given it.
ORBA is being built to help businesses examine that machine-readable picture — identifying what’s clear, what’s missing and where website content can be strengthened for the changing world of AI-powered discovery.
