How do AI citations work?
When an assistant answers a question it retrieves candidate sources, keeps the ones it treats as authoritative for that specific question, writes an answer from them and credits what it used. Citations are generated fresh each time rather than stored, so the same question can return different sources on different days. Across 4,291 answers we measured, the number of sources credited ranged from an average of 1.8 in ChatGPT to 9.9 in Perplexity.
By Liron Segev. Last updated August 2026.
Figures on this page are from our own research, not industry estimates: 4,291 answers collected between 28 April and 10 August 2026 across ChatGPT, Claude, Gemini and Perplexity, and 60 businesses across 41 niches tested between 3 June and 14 August 2026. Method, sample and limitations.
What makes a source worth citing
It addresses the actual question
Not the topic, the question. A page about your services is about a topic. A page that answers whether hail damage is worth an insurance claim addresses a question. Only the second is a candidate.
It can be lifted cleanly
A self-contained passage that still makes sense with the surrounding page removed. Anything that depends on the three paragraphs above it cannot be quoted, because a quote arrives without them.
The source is resolvable
The engine can tell who published it and treat that publisher as a consistent entity. An unattributable page is a risk to cite and gets skipped in favour of one that is not.
Something else agrees
Corroboration from sources that are not you. A claim that appears only on the site of the business making it is the weakest form of evidence available.
It is current enough to be safe
Freshness matters more for some questions than others. Prices, regulations and availability decay fast. A definition does not.
Notice what is absent from that list. Domain age, page count, how long you have been in business, and how much you spend on marketing. None of them are selection criteria, which is why small businesses routinely take citations off large ones on specific questions.
Why source counts differ so much
The five-to-one spread between engines is the single most useful fact about citations, because it means the same page faces two completely different competitive situations depending on where the question is asked.
In an engine citing ten sources, being credible and relevant is enough. You do not have to be the best answer, you have to be one of the acceptable ones. That is a wide door and it is why businesses tend to appear in Perplexity first.
In an engine citing under two, there is no consolation position. Either your answer is the one it uses or it is not. This is why ChatGPT feels so much harder, and it is also why being in it is worth more: when ChatGPT names a business it does so in the first 200 characters 88% of the time, so you lead the answer rather than appearing in a list at the bottom.
The mistake this causes is judging the whole effort by the strictest engine. Appearing in Perplexity months before ChatGPT is the normal sequence, not a sign of failure.
Being cited and being mentioned are different
A citation credits a source, usually with a link. A mention names your business inside the text of the answer. They often happen together and they are not the same thing, and the second is the one that actually moves a buyer.
Someone asking which company to use is reading the sentence that names a business. They are not clicking the reference markers. So an answer that names you without linking you has done the job, while a citation that credits your page without naming your business in the text has done considerably less.
This matters because it is measured backwards almost everywhere. Traffic dashboards can only see the click. The mention, which is worth more, leaves no trace in your analytics at all. If you judge this work by sessions you will systematically undervalue the result you are paying for, and you will optimise toward the wrong thing.
How to find out whether you are being cited
You can do this yourself in about twenty minutes, and it is worth doing before you pay anyone to do it for you.
Write down ten real buyer questions. Not keywords. The sentences a customer would actually type, including the ones about price and the ones that imply distrust. If you cannot produce ten, that is your first finding.
Ask all four engines, without your business name. ChatGPT, Claude, Gemini and Perplexity. Leaving your name out is the entire point: you are testing whether you get surfaced, not whether the engine can look you up. Asking "what do you know about my company" tells you nothing about visibility.
Record two things separately. Whether your business is named in the text, and whether your site appears in the credited sources. They are different results and mixing them hides the more valuable one.
Then look at who did get named. Open their page on that question. Nine times out of ten you will see immediately why they were chosen and you will not need any further diagnosis.
Two cautions. Answers vary between runs and between accounts, so a single result is noise and a pattern across ten questions is signal. And do it from a logged-out or private session where you can, since personalisation from your own history will flatter you.
The five reasons, in order of frequency
Nothing addresses the question. By a distance the most common. Not a quality problem, an existence problem.
The answer cannot be extracted. It is on the page, wrapped inside an argument, with no heading and no boundary.
The publisher cannot be resolved. The engine cannot confidently establish who stands behind the page.
Nothing corroborates it. The only source for the claim is the business making it.
Something contradicts it. Conflicting facts about your own business across your site and your listings make every claim riskier to repeat.
Four of those five are fixable in weeks. The fourth is the slow one, and it is the one that decides whether the other four hold.
Related questions
What is an AI citation?
A citation is the source an assistant credits when it writes an answer, usually shown as a link or a numbered reference beside the text it informed. It is the machine version of showing your working: the assistant is indicating which sources it drew on so the answer can be checked. Being cited means your content was one of the sources used to build the answer.
Is an AI citation the same as a backlink?
No. A backlink is a permanent link on someone else's page that you can go and look at. A citation is generated at the moment the answer is written and may differ the next time the same question is asked, by the same engine, from the same account. Backlinks are an asset you accumulate; citations are an outcome you earn repeatedly.
How many sources does an AI answer cite?
It varies enormously by engine. Across 4,291 answers we measured, Perplexity cited an average of 9.9 sources per answer and ChatGPT cited 1.8, a spread of more than five to one. That difference changes the strategy: being one of several acceptable sources can be enough in Perplexity, while ChatGPT behaves much closer to winner takes all.
Do AI citations send traffic to my site?
Some, and much less than the equivalent search position would. Many people read the answer and never click, which is the point of the format rather than a defect in it. The value of a citation is mostly that your business was named as credible at the moment someone was deciding, not that a session appeared in your analytics.
Why does my competitor get cited and I do not?
Almost always because they published something that addresses the specific question and you did not, or because what you published cannot be lifted out of its page cleanly. It is worth checking directly rather than assuming: ask the question, look at what the assistant cites, and open those pages. The reason is usually visible within a minute.
Can I get cited without being linked?
Yes, and it happens constantly. An assistant can name your business inside an answer without attaching a link to it, which counts as being recommended even though it registers nowhere in your analytics. This is the most commonly missed form of the result, and it is one reason judging this work by traffic understates it.
Other questions buyers ask
Find out where you actually show up
The free scan runs your business against ChatGPT, Claude, Gemini and Perplexity and shows you which buyer questions you appear in and which ones a competitor owns.