How Perplexity Chooses Its Sources and How to Become One
Amit Malvi
Founder, SearchAxe. Former Senior SEO Manager, Invideo AI
Published Updated 9 min read
Quick answer
How Perplexity chooses its sources is documented in unusual detail. It searches its own index of more than 200 billion URLs, crawled by PerplexityBot, using keyword and meaning-based retrieval together. It filters out stale or off-topic results, ranks individual passages with several layers of models, and hands the best spans to a language model that cites them with numbered links. To become a source, be crawlable, current and quotable passage by passage.
On this page
- How Perplexity chooses its sources, step by step
- What kind of content does Perplexity's index favour?
- Does ranking on Google or Bing help you get cited by Perplexity?
- How does Perplexity treat brand sites versus third-party pages?
- Does blocking PerplexityBot remove you from Perplexity?
- How do you become a Perplexity source?
- Is Perplexity worth optimizing for on its own?
Key takeaways
- Perplexity runs its own search index of more than 200 billion URLs, so PerplexityBot access matters more than your Google or Bing listing.
- It ranks and cites passages, not whole pages, so a self-contained paragraph can win a citation even on a long page.
- Stale content is filtered out early in its ranking pipeline, which makes visible, accurate update dates and real updates worth the effort.
- Reddit made up 6.6% of all Perplexity citations in Profound's analysis, so community threads compete with your own pages.
- Perplexity sent 7.23% of AI referral traffic in 2026 in SE Ranking's data and was the only major AI engine whose traffic declined, so weigh it against the engines your buyers use most.
Perplexity chooses sources by searching its own web index, ranking the passages that answer each part of the question, and citing the pages those passages came from. If you run growth at an AI startup and want Perplexity to quote your pages when buyers research your category, this is how its pipeline works and what to change. This article covers Perplexity's pipeline. How AI search engines decide which sources to cite compares it with ChatGPT, Claude, Gemini, Google AI Overviews, Google AI Mode and Copilot.
How Perplexity chooses its sources, step by step
Perplexity's September 2025 research article on its search infrastructure describes a pipeline that processes 200 million queries a day, and its Search API announcement says the API runs on the same infrastructure that powers the public answer engine. Read together, they give an unusually detailed picture of how a page becomes a Perplexity source.
From your page to a numbered citation
- Crawl and indexPerplexityBot crawls the web into an index tracking over 200 billion unique URLs. A machine learning model decides which URLs to index and when to refresh them, based on importance and how often they change.
- Parse into spansA content understanding module extracts the meaningful content and splits each page into self-contained spans that can be retrieved on their own.
- Retrieve by keywords and meaningEach query runs as both a lexical and a semantic search, and the results are merged into one candidate set that favours coverage over precision.
- FilterPrefiltering removes clearly non-responsive or stale content before any expensive ranking happens.
- Rank in stagesFast lexical and embedding scorers narrow the set, then cross-encoder rerankers make the final cut, scoring both whole documents and individual spans.
- Answer and citeA language model writes the answer from the top spans, and each answer includes numbered citations linking to the original sources.
The last step uses whichever model the user picks. Perplexity's help article on how it works lists models such as GPT-5 and Claude 4.6 Sonnet in Pro Search, and says its Research mode performs dozens of searches and reads hundreds of sources for one report. The sources are chosen by Perplexity's own search stack before any of those models write a word.
What kind of content does Perplexity's index favour?
Perplexity's pipeline favours content that is fresh, from domains it treats as authoritative, and easy to split into clean, self-contained passages. Each preference is stated in its research article, so these are design choices, not guesses from the outside.
- Freshness: the indexing model is trained to match refresh timing to how important a URL is and how often it changes, and stale content is removed at the prefiltering stage.
- Authority: when choosing which documents to keep in fast storage, Perplexity prioritizes documents from authoritative domains and documents on undercovered topics.
- Clean structure: its parser uses different rules for well-structured list and table content than for freeform user-generated content, and good parsing is what lets a page be cut into retrievable spans.
- User signal: ranking models are trained on signal from the millions of queries Perplexity serves each hour, combining AI-driven evaluations with live human feedback.
Two practical consequences follow. A page whose key facts sit inside images, accordions or a single long paragraph is hard to split into spans, so it loses to a page with one answer per section. And a comparison or pricing page that has not changed in two years competes against pages refreshed last month, in a pipeline built to drop the stale one. The formats that survive this are covered in how to structure a web page so AI engines can quote it.
Find out which AI engines recommend you
A free consultation on where you appear today across AI engines and search.
Schedule a Free ConsultationDoes ranking on Google or Bing help you get cited by Perplexity?
A Google ranking helps more with Perplexity than with any other assistant Ahrefs tested, even though Perplexity runs its own index. Pages that rank well tend to be the pages any good retrieval system finds relevant, and Perplexity's citations line up with Google's results noticeably more often than its rivals' do.
The flip side is that most of Perplexity's citations in that study, everything outside that 28.6%, did not rank in Google's top 10. Organic rankings on Google and Bing are a head start, not the decider. A page ranking twelfth with a crisp answer paragraph can be cited over the page ranking first with a slogan.
How does Perplexity treat brand sites versus third-party pages?
Perplexity documents no rule for or against vendor sites, but its citations lean heavily on community and video platforms. In Profound's analysis of 680 million citations from August 2024 to June 2025, Reddit made up 6.6% of all Perplexity citations and 46.7% of citations among its top 10 domains, with YouTube at 2.0%.
In practice that splits by question type. A factual sub-question such as "Does [tool] have a free plan" is best answered by your pricing page, and a well-parsed pricing page can win that citation. A judgement question such as "Which AI transcription tool do podcasters prefer" pulls in Reddit threads, reviews and roundups, because those are the spans that contain opinions. You need your own pages for the first kind and other people's pages for the second, which is why earning Reddit mentions and getting into best-of lists belong in the same plan.
Does blocking PerplexityBot remove you from Perplexity?
Blocking PerplexityBot keeps your pages out of Perplexity's index, which removes you from almost every answer. Perplexity's crawler documentation says PerplexityBot is designed to surface and link websites in search results, is not used to train foundation models, and that robots.txt changes can take up to 24 hours to take effect.
A second agent, Perplexity-User, visits pages when a user's request needs them and generally ignores robots.txt, because a person initiated the fetch. So a blocked site can still be read when someone pastes its URL, but it will not be retrieved for ordinary questions. Since PerplexityBot does no training, there is little reason for an AI startup to block it; the wider trade-offs are in should you allow GPTBot, ClaudeBot and PerplexityBot to crawl your site.
How do you become a Perplexity source?
Make your pages reachable by PerplexityBot, split them into passages that each answer one question, keep them current, and get discussed on the third-party pages Perplexity already cites. The checklist below covers the work in the order we do it.
- PerplexityBot is allowed in robots.txt and not blocked by your CDN or firewall, checked against the IP list Perplexity publishes.
- Each important page answers its main question in the first two sentences under a clear heading.
- Each section stands on its own, with the product name and the fact in the same paragraph rather than "it" and "this".
- Specifications, pricing and limits are in HTML text, lists or tables, not in images or collapsed tabs.
- Pricing, comparison and integration pages are reviewed quarterly and show an accurate last-updated date.
- Your product is named in the Reddit threads, reviews and roundups that Perplexity cites for your category prompts.
- You re-run your category prompts monthly and log which URLs Perplexity cites, yours and everyone else's.
To see where you stand today, run the questions your buyers ask and open every numbered source. The no-cost method is in how to check whether ChatGPT and Perplexity recommend your product without paid tools, and the clicks Perplexity sends can be isolated in analytics with AI referral tracking in GA4.
Is Perplexity worth optimizing for on its own?
Usually not on its own, and the traffic data says why. SE Ranking's June 2026 study of 101,574 websites found Perplexity sent 7.23% of AI referral traffic in 2026, third among the platforms tracked, and that its share of US AI traffic fell from 19.73% in the first four months of 2025 to 6.85% in 2026. Perplexity was the only major AI search engine in that study whose traffic declined slightly.
The good news is that almost nothing on the checklist is Perplexity-only. Passage-level answers, fresh pages and third-party mentions help every engine. The Perplexity-specific work is crawler access, freshness on the pages it cites and the Reddit presence it weights so heavily. If your buyers are researchers, analysts or developers who use Perplexity daily, that is where a Perplexity SEO agency earns its keep; if they are not, fold Perplexity into your wider AI visibility work.
Frequently asked questions
Does Perplexity use Google or Bing to find sources?
Perplexity runs its own index. Its September 2025 research article says the index tracks over 200 billion unique URLs, crawled by PerplexityBot, and that early versions of its answer engine used third-party search APIs before it built its own. Its citations still overlap with Google more than other assistants do: Ahrefs found 28.6% of Perplexity citations ranked in Google's top 10 for the same prompt.
Does PerplexityBot use my content to train AI models?
No. Perplexity's crawler documentation says PerplexityBot is designed to surface and link websites in Perplexity's search results and is not used to crawl content for training AI foundation models. A separate agent, Perplexity-User, fetches pages when a user's request needs them and generally ignores robots.txt. Allowing PerplexityBot is therefore a visibility decision, not a training decision.
How often does Perplexity re-crawl pages?
Perplexity does not publish a fixed schedule. Its research article says a machine learning model decides when to index or refresh each URL based on its importance and how often it changes, and that the system processes tens of thousands of index updates per second. Pages that change regularly and matter to many queries get refreshed sooner, which is one more reason to keep key pages current.
Why does Perplexity cite Reddit instead of my website?
Usually because the question asks for opinions or experiences, and Reddit threads contain those spans while vendor pages do not. Profound found Reddit made up 6.6% of all Perplexity citations, the largest share of any domain. Your pages can still win factual sub-questions about pricing, features and integrations, so answer those clearly and earn mentions in the threads Perplexity cites for judgement questions.
