An AI citation is a source an engine attaches to a claim it just made. Unlike a backlink it passes no authority, expires with the next crawl, and can appear without the reader ever visiting the page.
Every measurement habit borrowed from link building breaks on those three properties, which is why teams that count citations like links end up with a number that cannot be compared to itself.
The word is borrowed from academia and the borrowing is doing quiet damage. In a paper, a citation is a permanent, countable, author-chosen reference. In an answer engine it is none of those things, and treating it as though it were produces reporting that looks rigorous and means nothing.
What exactly is a citation here?
An AI citation is a source reference an answer engine attaches to a statement in a generated answer, identifying where the material behind that statement came from. It exists only within that answer, and a second run of the same prompt may cite something else entirely.
The mechanics are documented. Anthropic’s web search tool states that “citations are always enabled for web search”, and each one carries a URL, a title, and up to 150 characters of the cited text. That last field is the giveaway: a citation is anchored to a specific passage the model used, not to a page it admired.
How is it different from a backlink?
It passes no authority
Nothing accrues to your domain.
- Backlink
- A durable signal another site chose to give you.
- Citation
- A byproduct of retrieval. Nobody decided you deserved it.
It expires
There is no stock, only a rate.
- Backlink
- Sits in the HTML until somebody edits it. Auditable years later.
- Citation
- Exists for one answer at one moment. Nothing to audit afterwards.
It can be invisible
Influence without a visit.
- Backlink
- Its value is partly the clicks it sends, which you can count.
- Citation
- A reader can be persuaded and never click. Nothing reaches analytics.
Why does it expire?
Because it was never stored. Most answer engines follow the retrieval-augmented pattern set out in Lewis et al., where documents are fetched at query time and the answer is generated from whatever came back. The citation is a record of that fetch, so it lasts exactly as long as the answer does.
Does everyone show citations the same way?
No, and the differences are large enough to change what you can measure.
Some engines attach numbered sources to specific sentences, some list them at the end of the answer, and some name a brand in prose while citing nothing at all. Perplexity tends to cite densely, which makes it the most informative engine to audit even if it is not the one your buyers use most. Assistants that summarise without linking produce mentions that no link-based tool will ever see.
That variance is a measurement problem rather than a quality ranking. A tool reporting one citation number across several engines is averaging over products that disagree about what a citation even looks like, which is one more reason to keep the per-engine figures visible rather than rolled up.
Does a citation even point at you?
Often it does not, and this is the part that surprises people most. The engine cites the document it drew the claim from, which is frequently somebody else’s page describing you.
The answer links to you. This is what citation rate measures, and it is the rarer of the two outcomes for comparison questions.
You are named in the answer, and the citation goes to the review platform or roundup that said it. Real visibility, no link, nothing to your domain.
If it passes no authority, is it worth anything?
Yes, and the value is in a different currency, which is why link-shaped reporting cannot see it.
A backlink is worth something to an algorithm. A citation is worth something to a person, and to the answer they are reading right now. Being the source behind a claim in the answer a buyer is acting on is not a ranking signal. It is the point of the exercise.
Nothing accrues. A thousand citations leave your link profile exactly where it was, and no SEO metric moves at all.
You are the source the buyer’s answer was built from, at the moment they are deciding. That has never been available for purchase.
There is a second-order effect worth knowing. Because engines cite the passage they used rather than the page they admired, an article that answers one question cleanly is more citable than one covering ten topics well. Google’s snippet documentation describes the same mechanic from the search side: snippets are generated automatically from page content, and self-contained content produces better ones. Structure is doing more work here than word count.
So how should you count them?
As a rate over a held-constant experiment, which is the only form that survives the three properties above.
The denominator is the prompt set. Change it and the rate is no longer comparable to last month’s.
Runs per prompt per engine per day. Without it, “citation rate” is a number with no denominator.
The citation is gone the moment the answer is. Store the full text or the record is unrecoverable.
The practical consequence for reporting is a sentence worth rehearsing before somebody asks: we are cited in 34% of runs across a fixed set of thirty prompts on three engines, sampled five times a day. That is defensible. “We have 200 AI citations” is not, because there is no such stock to have.
Both terms are defined precisely in the glossary, alongside the other 58 the field actually uses, and the wider argument for why the whole scoreboard has to be rebuilt sits in AEO vs SEO.
Can you influence what gets cited?
Indirectly, and only by making a passage worth quoting rather than by making a page worth ranking.
Because the citation attaches to the passage a model actually used, the practical lever is passage-level: a heading that states a question, an answer directly underneath it that makes sense to somebody who arrived at that paragraph and nothing else, and a claim specific enough to be worth attributing. Vague prose is retrievable and unquotable, which is the worst combination available.
What you cannot do is earn citations the way you earn links. There is nobody to ask, no outreach to run, and no relationship to build with the retriever. You make the material citable and then you sample the answers to find out whether it worked, which is a less satisfying loop than link building and a considerably more honest one.
One last framing that saves arguments. A backlink is something you earned once. A citation is something you are currently earning, continuously, and stop earning the moment the corpus moves. That is less comfortable than a link profile, and it is a fairer description of how visibility actually works.