AI visibility
How to Read the Links Next to Perplexity Answers
A link next to a Perplexity answer is attributable only if the sentence still holds without it. Here is a three-step test you can apply sentence by sentence, plus the limits of what the visible page can tell you.
You are looking at a Perplexity answer with three or four numbered links next to the sentences. To work out how to read the links next to an answer, start with one operational rule: a link is attributable if the sentence it sits beside still holds when you take the link away. If the sentence stays true without that link, the link is context. If the sentence collapses, the link is carrying the claim. Apply this test sentence by sentence, not to the answer as a whole.
This is a reading model we propose at RankAI, not a description of how Perplexity works internally. We look at what is visible and treat it as a clue, not as proof of what the system used.
Why the link is not the source that was used
A link next to a sentence points to a page. That page is what the link shows, not necessarily what the system processed to build the answer. The distinction matters when you are checking which pages get cited: you are reading a public trace, not an internal log. That is why we say “the page the link points to” and, more cautiously, “the page that appears to have been used.” The second phrasing is a reading hypothesis, not something an outsider can verify.
Keep the terminology note for after you have run the test, not before: read first, label second.
The RankAI checklist in three steps
This is a working proposal from RankAI, not a universal standard. Each step has a criterion you can check.
- Remove the link and reread the sentence. Criterion: if the sentence loses the subject, the figure or the cause that the link introduces, the link is attributable. If the sentence reads the same, the link is context.
- Open the page and look for the sentence, not the topic. Criterion: the page is relevant if it contains the specific information in the sentence, not just the general subject. If you find only the topic, the link is weak.
- Compare several links on the same sentence. Criterion: if two links point to pages that say different things, treat the sentence as ambiguous rather than as attributed to a single page.
What to do with the checklist results
If a sentence passes all three steps, you can say the link points to a page that appears to support that claim. If it passes only one, the link is decorative or contextual. If no sentence passes the test, the links next to the answer are not doing attributive work: they are a list of pages you can consult, not a chain of evidence.
This method is for checking which pages get cited, not for reconstructing the engine's internal path. It is the same difference we draw elsewhere: what comes back from an engine through an API is the API response, not the screen a person sees. The screen with the links is a separate observation and has to be read with its own limits.
The contribution of this piece is the “remove the link” test, applied sentence by sentence. In practice, isolate the sentences that contain a number, a date, a proper name or a causal relation: these are the ones where the link tends to do attributive work. Descriptive or transitional sentences rarely stand or fall on the link, so leave them out of the count to avoid inflating the result. A second tip: record the outcome as “attributable,” “context” or “ambiguous,” without turning it into a score. Reading links is qualitative and stays that way; turning it into a number would suggest a precision that the visible observation does not have. Finally, note the date you read the answer: the response and its links can change between one visit and the next, and a result without a date cannot be compared with anything.
A limit you can check
The “remove the link” test measures how much a sentence depends on the link, not where the information came from: a sentence can hold without a link even if it was built on a page that is not cited. You can check this limit by rereading the answer without opening any link and counting how many sentences stay intact.
For the wider picture on how to read visibility traces in AI engines, see the in-depth page.