Perplexity is built around showing its sources inline rather than hiding them, which makes it the engine where citation behaviour is easiest to observe and the weakest proxy for how engines that cite sparsely behave.
Nobody outside Perplexity knows how it ranks sources, and this page does not pretend otherwise. What it covers is what the company documents, what you can observe, and why auditing on Perplexity alone will mislead you.
What does Perplexity actually document?
Two crawlers, and the split matters more than it first appears.
Perplexity’s bot documentation describes PerplexityBot as “designed to surface and link websites in search results on Perplexity”, and states explicitly that it is not used for training AI models. Separately, Perplexity-User supports user-initiated actions, visiting pages when someone asks a question, and generally ignores robots.txt on the reasoning that a person asked for that specific page.
PerplexityBot
Surfaces and links sites in Perplexity’s results. This is the one that decides whether you can be cited at all, and it needs allowing.
Allow this one
Perplexity-User
Fetches a page because a person named it. Robots rules generally do not apply to this class of visit.
Ignores robots.txt
The one thing this does tell you is unambiguous and actionable: a blanket rule blocking “the AI bots” catches PerplexityBot, and blocking it removes you from Perplexity’s answers entirely. That is a configuration decision rather than a ranking one, and it is the single most consequential thing on this page.
What is observably different about it?
The product is built around its sources rather than around a conversation that happens to have some. Citations appear inline, attached to statements, and they are the interface rather than a footnote.
That is a product design choice with a measurement consequence, and it is the honest version of “ranks sources differently”. Whether its retrieval is genuinely different from another engine’s is not observable from outside. Whether it shows you more of what it used plainly is.
Inline attribution makes the retrieved set unusually legible. You can read the sources behind a claim about your category without any tooling.
The scoring, the ordering and the cut-off are not published. Observing the output tells you the result, never the function.
Why is it the best engine to audit with?
Because visible sources turn a vague question into a concrete one.
When you ask a buying question in your category and get four vendors and a source list, that source list is a map of where your category’s comparative evidence lives. Sorting those domains tells you which review platforms, which comparison articles and which communities are actually being read on your behalf. That is a research output you can act on regardless of which engine produced it.
A list of the sources your category’s answers are built from, which is otherwise guesswork.
Whether those domains mention you, mention a competitor, or mention both differently.
Read from the sources themselves, which is where an inconsistent description becomes visible.
Why is it the worst engine to generalise from?
Because engines cite at very different densities, and a workflow built around abundant visible sources breaks on engines that provide few or none.
The families genuinely differ here. Anthropic documents that for its web search tool “citations are always enabled”, so sources come back reliably. Google’s AI surfaces cite links drawn from the Search index, and a page qualifies as a supporting link when it is “indexed and eligible to be shown in Google Search with a snippet”. And an engine answering from stored knowledge rather than a live lookup may name you with no source to point at, because none was fetched.
The domains carrying comparative material are largely the same corpus whichever engine reads them. That finding travels.
Measured on an engine that cites heavily, it will overstate what you would see on one that cites sparingly or not at all.
Why does the crawler split matter more than the ranking?
Because one of them is a decision you control and the other is not.
The ranking function is unpublished and would be unactionable even if it were published, since you cannot change how a retriever scores. The crawler configuration is a line in a file, it is documented, and getting it wrong removes you from the product entirely rather than lowering your position in it.
The failure is common because the two Perplexity agents are easy to conflate with the training crawlers people meant to block. A rule written to opt out of model training, if written broadly, catches PerplexityBot as well. The resolution rules are also specified rather than conventional: RFC 9309 defines that the most specific matching rule wins and that an allow beats an equivalent disallow, which is the reverse of what most people assume when they write these files.
Allowing the crawler. Binary, checkable in two minutes, and decisive.
Self-contained passages and third-party evidence. Helps here and everywhere else equally.
The ranking function. Unpublished, and nothing you could do with it if it were.
How should it sit in a prompt set?
As one engine among at least three, from different families, with its figures reported separately rather than blended.
Blending is where the damage happens. Averaging a citation rate across engines that cite abundantly and engines that cite rarely produces a number describing no engine at all, and it moves when the mix changes rather than when your visibility does. Per-engine rows, always.
The wider point is that engines disagreeing is a finding rather than a measurement problem. If you are cited on Perplexity and absent on a grounded surface, that is a real difference in where your evidence lives, and it is worth more than a tidy average would be.
What can you actually do about Perplexity specifically?
Less than the amount written about it would suggest, and the honest list is short.
Allow PerplexityBot. Make your pages retrievable as self-contained passages, since retrieval scores chunks rather than documents. Get into the third-party sources that show up in its citation lists, because those are what it is summarising. That is the whole intervention, and every item on it is the same work that helps on every other engine.
There is no Perplexity-specific tactic worth having. If you find advice claiming otherwise, check whether it cites Perplexity’s own documentation, because that documentation covers crawlers and stops there.
The full engine breakdown and how the retrieval families differ is in the engine guide, the architecture underneath all of it is in the RAG explainer, and if you want to run this properly across several engines rather than just this one, the no-budget audit has the method.