How AI Assistants Choose Which Sources To Cite
An assistant answering a question does not rank a hundred results and show you ten. It retrieves a handful of documents, usually somewhere between three and ten, and writes an answer from them. Everything else on the topic is invisible.
That changes the shape of the problem. Ranking eleventh used to mean a trickle of traffic. Being the eleventh most retrievable document on a topic now means nothing at all.
Mention And Citation Are Different Things
Worth separating before anything else, because people conflate them and then measure the wrong one.
A mention is your name appearing in the generated text. It can come from the model’s training data, from a retrieved third-party page, or from a source that never named you as a link.
A citation is your specific URL surfaced as a source, usually as a numbered reference or an inline link.
You can be mentioned constantly and cited never, which is common for well-known products, and cited without being mentioned in the prose, which is common for reference pages. They are influenced by different things, so track them separately.
Most Citations Are Not Your Website
This is the finding that reorders priorities. Across large samples of AI citations analysed through 2025 and 2026, the overwhelming majority come from third-party sources rather than the brand’s own domain. Encyclopaedic and community sources take a disproportionate share: Wikipedia is consistently the single largest source in ChatGPT citation studies, and community discussion sites account for a large slice across platforms.
The practical implication is uncomfortable for anyone whose content plan is entirely on-domain. If assistants answer “best X for Y” mostly from forums, review sites, and reference works, then publishing your twentieth comparison post on your own blog is not where the leverage is.
Where the leverage is: being accurately described in the places that do get cited, and being the source those places cite when they describe you.
What You Can Actually Influence
Be describable in one sentence. If an assistant has to compress you to a clause, decide what that clause says. “A Windows text rewriter with an offline local mode” survives compression. “An AI-powered productivity platform” does not, and gets replaced by whatever the model assumes.
Publish the specific facts nobody else has. Model sizes, memory requirements, exact pricing, what leaves the device. Third parties writing about you will lift those, and those lifts are what get cited later.
Make your claims checkable. Content with concrete numbers and attributed sources is picked up more often than content that asserts. That is the most robust finding in the GEO research literature, and it survives every change in retrieval plumbing.
Fix the third-party record. Out-of-date entries on reference sites, wrong pricing in a roundup, an abandoned comparison page from 2024. Those are load-bearing now in a way they were not when they simply ranked below you.
Write The Sentence You Want Quoted
The single highest-value edit is turning a positioning paragraph into an extractable sentence.
Before:
Our platform leverages cutting-edge language technology to help modern professionals communicate more effectively across a wide range of contexts, with a strong emphasis on privacy and user control.
After:
Wrivio is a Windows text rewriter that opens over any application with
Ctrl+Shift+Space. Its Local mode runs an open-weights model on the device and makes no network calls during a rewrite; its Preview mode sends text to Wrivio’s own backend.
The first version is unquotable, so a model summarising you will produce its own version, and its version will be wrong in whichever direction the training data leans. The second version is specific enough that quoting it is easier than paraphrasing it.
The general technique is covered in how to write so AI search quotes you correctly.
A Context For Positioning Text
A Wrivio Context for this could say:
Rewrite this into two or three self-contained sentences that a stranger could quote without context. Replace every abstract capability phrase with the concrete mechanism or figure given in the original. Remove marketing intensifiers. Keep every product name, keyboard shortcut, price, and technical figure exactly as written. Do not add capabilities, integrations, or claims that are not in the original.
Press Ctrl+Shift+Space, paste the paragraph, and read the diff closely. The clause about not adding capabilities is doing real work, because a rewrite that promotes “privacy focused” into “end-to-end encrypted” has manufactured a claim you will have to defend.
What Not To Bother With
Buying placements in listicles that exist only to be scraped. Those pages get discounted or removed, and the ones that survive tend to describe you in the vendor’s template rather than yours.
Chasing a vendor “visibility score” with no published methodology. No assistant publishes a ranking signal to optimise against, so any score is a proxy someone invented.
Adding files and tags in the hope that they change retrieval. They do not, as the llms.txt story shows in detail.
Check It By Hand, Monthly
Ask each major assistant the five questions your buyers actually ask. Record which sources it cites, how it describes you, and what it gets wrong. Twenty minutes a month produces a better picture than any tool currently sells, because you are measuring the thing itself instead of a proxy for it. The rest of the measurement set is in AI search metrics worth tracking.
Common Questions
How many sources does an assistant use per answer?
Typically three to ten when browsing is enabled. That small number is why marginal visibility gains matter far less than they did in ranked search.
Why does Wikipedia get cited so often?
It is broad, structured, heavily linked, and consistently formatted, which makes it cheap to retrieve and easy to chunk. Citation studies through 2026 consistently show it as the largest single source for ChatGPT.
Can I pay to be cited?
Advertising placements and generated citations appear to operate as separate layers, and no vendor publicly documents ad signals influencing retrieval. Treat any offer to guarantee citations as unfounded.
Does being cited actually send traffic?
Very little directly, since users rarely click source links. The value is in being described accurately at the moment someone is deciding, which is why the framing matters more than the referral.
Download Wrivio for Windows to tighten your positioning text into sentences worth quoting, with a diff that shows exactly what changed.
Read Next
Search Console AI Performance Reports: What They Do And Do Not Tell You
Google added generative AI reporting to Search Console in 2026. What is counted, what is bundled, and which questions the data still cannot answer.
Zero Click Search: What To Do When The Answer Replaces The Link
Users click far less when an AI summary appears. What the measured numbers actually say, and how to write pages that still earn something without the click.
AI Search Metrics Worth Tracking, And The Ones That Are Vanity
Rank tracking stopped answering the question. A short set of measurements that survive contact with generative search, including one you do by hand.
llms.txt in 2026: A Reality Check Before You Ship One
Adoption jumped, then Google said it ignores the file entirely. What llms.txt was proposed for, what it does not do, and when it is still worth publishing.
This article is filed underContent & SEO, which has 30 articles.