How to rank in Perplexity
Perplexity is the only major assistant that shows its working on every answer. That makes it the cheapest feedback loop in generative engine optimisation - and the place to run your experiments before you spend anything on the surfaces you can't see into.
The short answer
Perplexity runs a live search, then writes an answer from what it found and shows you the sources. So "ranking" in Perplexity is mostly two jobs stacked: be retrievable for the search step, then be the most quotable passage among what came back. Neither is mysterious. Both are checkable, today, for free.
That second part is why this page exists. On every other major assistant you are guessing about why you were or weren't named. On Perplexity the sources are printed on the answer. It is the only surface in generative engine optimisation with a visible scoreboard, which makes it the correct place to run experiments before you commit budget to surfaces you can't see into.
Treat Perplexity as your test rig, not your goal. It probably isn't where most of your buyers are - but it's where you find out whether your GEO work is doing anything at all, within days instead of never.
How it differs from ChatGPT
Both are assistants; they are not the same kind of machine. ChatGPT blends what the model already knows with a search step it may or may not take, and cites inconsistently. Perplexity is search-first by design: retrieval happens nearly every time, and the citations are the product, not a footnote.
| Question | Perplexity | ChatGPT |
|---|---|---|
| Does it search live? | Nearly always | Sometimes - depends on the question and mode |
| Are sources shown? | Yes, by default, on the answer | Inconsistently |
| Can you see why you lost? | Partly - you can see who won instead | Rarely |
| Feedback loop | Days | Weeks, and often ambiguous |
| Best used for | Testing whether a change worked | Measuring where the audience actually is |
The practical consequence: run your experiments on Perplexity and your reporting across everything. Optimising exclusively for the surface with the best feedback is a classic streetlight error, and it's how people end up winning a channel their buyers don't use.
What its citation behaviour appears to reward
Stated plainly: none of this is documented. What follows is inferred from repeatedly watching what gets cited and what doesn't, by people testing from the outside. Read it as a set of reasonable bets, not a specification.
Being in the search results it draws from
Because retrieval happens live, conventional search visibility for the underlying query is the strongest lever most people have. If your page isn't findable for the question at all, the synthesis step never sees it. This is the least fashionable insight in GEO and the most reliable one: a lot of "ranking in Perplexity" is ordinary SEO happening upstream of the answer.
Answering the specific question, close to the top
Answers get assembled from passages. A page that states its answer in the first self-contained paragraph under a matching heading is trivially usable; a page that builds to its point over eight paragraphs is not. Put the answer first, then elaborate for the humans who stayed.
Freshness and visible dates
A live-search system leans toward recent, clearly dated documents, particularly for anything that changes. An undated page asks the machine to guess whether you're current, and the guess tends not to go your way. Show published and updated dates, and make them true.
Specificity it can attach to a source
Synthesis needs something concrete to cite. Numbers, named comparisons, and original data give it a reason to point at you rather than paraphrase the consensus. Generic advice has no owner, so it gets no citation.
Third-party corroboration
Retrieval returns whatever documents best match the query, and for buying questions those are often roundups, forums and comparison pages you don't control. Being present and accurately described on those pages puts you inside the candidate set before your own site is ever considered.
How to check whether you're cited
This takes an hour and it is the only part of this page that produces evidence rather than opinion.
- Write twenty to forty prompts your buyers would actually type. "Best X for Y", "alternatives to Z", "how do I do W" - the questions asked before a purchase, not the ones you wish were asked. Then freeze the list. Changing prompts between rounds destroys the comparison.
- Run them and read the source list, not the prose. The answer text flatters everyone. The citation list is the actual result.
- Record who gets named instead of you. Competitor share of voice across your prompt set is a far more useful number than your own presence, because it tells you what the machine currently thinks the category consists of.
- Run each prompt more than once. Results vary between runs. Treat it as sampling a distribution, not reading a rank - one lucky check is worth nothing.
- Check your server logs for PerplexityBot. If the crawler has never been to a page, that page has no chance of appearing, and you've found a technical problem rather than a content one.
- Segment assistant referrals in analytics. The volume is small and the intent is high. If you don't split it out, it disappears into direct traffic and you'll conclude, wrongly, that nothing happened.
What to actually do
In order, cheapest and most gating first. There is no point writing quotable content for a crawler that isn't allowed in.
- Allow PerplexityBot in
robots.txt, deliberately. Blocking is a legitimate choice; blocking by accident and then wondering why you're never cited is not. - Serve content in raw HTML. If your page needs JavaScript to show its words, assume it's empty to retrieval. This is the single most common silent failure on AI-built sites.
- Rebuild your key pages around one question each, with the answer in the first paragraph beneath the heading. One page trying to answer six questions loses to six pages that each answer one.
- Publish something only you can publish. Your own numbers, dated. Ours, for the record: across four sites launched in 2026, first-weeks impressions ran 2,200 to 16,500, click-through rates 0.1% to 0.6%, average positions 12 to 49. Modest figures - but specific, ours, and quotable in a way that "growth takes time" never is.
- Go and be mentioned elsewhere. Be useful in public where your category gets discussed. It is slow, unglamorous, and it outperforms every on-page tweak on this list.
- Re-run the prompt set on a schedule - monthly is plenty - and log the results with dates so you can tell movement from noise.
This site is a new domain running exactly this loop, with the prompt-set results published as they come in - including the rounds where nothing improves. Predictions are logged with dates before the outcome is known and graded when they come due, on the receipts page.
An honest note on what isn't known
Perplexity does not publish ranking factors. There is no documentation of how sources are weighted, ordered, or dropped, no quality guidelines to read, and no way to audit a decision after the fact. Everything above is inference from observed behaviour, and the system can change under you without announcement.
What makes it worth doing anyway is that the visible citation list gives you something rare in this field: a cheap, fast, honest verdict. You can be wrong about the mechanism and still measure the outcome. That is the entire case for using Perplexity as your test rig - and for treating anyone who claims to know its algorithm as a source of entertainment rather than information. The broader picture is on the GEO pillar, and the citation mechanics common to all assistants are covered in how to get cited by ChatGPT.
Want your prompt set run properly, on a schedule?
The GEO Audit runs your category's real buying prompts across Perplexity and the other major assistants, records who gets named instead of you, and returns the specific fixes - with a 90-day recheck so you find out whether they worked.