How to read AI citation and source data

Mention, citation, retrieved source: what each one counts

Three different things get called sources in this field. A mention is the answer naming your brand, a citation is a link the engine attaches to a claim, and a retrieved source is a page the engine pulled in while composing. They count different events, so a chart that merges them gives you nothing to act on.

The distinction cuts both ways. An answer can cite a page of yours without naming you, which means your content did the work while a competitor got the credit. It can also name you with no link at all, usually because the model already knew the brand and never searched. Separate the three and each one answers a different question.

What citation share measures, and what it predicts

Citation share is your domains’ share of all cited domains in the same answers. It measures the raw material behind answers about your category rather than the answers themselves, which is why it can move before Visibility does.

Work through it concretely. Take the answers you tracked this period, collect every source link in them, group by domain, and see what fraction sits on your domains. If that share climbs while Visibility stays flat, more of the material is yours and the answers have not changed yet. Treat the ordering as a hypothesis worth testing, not a rule: whether citation share leads in your category is something your own history has to show.

Two things distort it. A single publication that gets cited across half your answers can move the number on its own, so always read the percentage next to the domain list. And engines contribute unevenly, because some of them publish no source list at all. Say which engines are in the count, or a gap in one engine’s coverage will read like a fall in yours.

How engines differ in what they show you

Perplexity attaches numbered citations to every answer by design, and its help center says so in as many words. Other engines give you no source list in some modes, so an empty column for one engine usually means the mode did not search, not that nobody cited you.

Perplexity’s own description of the product: “Each answer includes numbered citations linking to the original sources” (How does Perplexity work?). That makes it the easy case, because the sources are part of the interface rather than a hidden step.

Elsewhere, an engine’s default chat mode may answer straight from the model with no search step, in which case there is no list to read. When a mode does search, links usually show up in or under the answer. I would not quote a percentage for how often each engine cites: it moves with the prompt, the mode and the day, and a fixed figure you read in a vendor post is one sample presented as a law of nature.

How to read a sources list

Start with domain counts instead of prestige. Rank every domain by how many answers cite it, then split that list into domains that feed answers about you and domains that feed answers about your competitors. The second list is where the work is.

• Count by domain, not by URL. One publication shows up under many URLs, and it is the domain the engine tends to trust.
• Split the list by outcome. Which domains appear in answers that mention you, which in answers that mention only a competitor, which in answers that mention neither brand.
• Look at the overlap. A domain cited in both your answers and your competitor’s is the fairest ground. A domain only your competitor gets is the gap worth naming.
• Check the frequency. A domain that appears in a couple of answers is noise until it repeats next period.

That is the mistake to avoid: celebrating a citation you would put in a pitch deck. The engine’s next answer about your category draws from whatever sits in the pool that day, so a prestigious domain that never appears in answers about you has no influence on them.

Where the source data lives

The Sources view in advisible is the working surface for this loop: domains ranked across your answers, the sources behind each individual answer, and a CSV export for the part you want to slice in a sheet.

The per-answer detail is what makes a domain list defensible. When someone asks why a domain is on the list, open the answer and read it. Every original answer is kept, so when the analysis improves later the history behind an older number can still be checked against the text it came from.

For the other half of the job, deciding which sources to pursue rather than which to read, see which sources AI engines cite and get cited by ChatGPT and Perplexity.