How Perplexity Answers With Sources

Perplexity is an answer engine. You ask a question, it searches the web, picks the most useful passages, and writes an answer with numbered citations you can click. You met the general recipe, retrieval-augmented generation, in Research With AI: Fast Answers You Can Actually Cite. This article looks at one real system that uses it.

Perplexity has published part of its design. In September 2025 it described the search system behind its Search API, which it says is "the same global-scale infrastructure that powers Perplexity's public answer engine" (Introducing the Perplexity Search API). Other parts, such as exactly how it prompts the model to cite, it has not published. Below, we keep the two apart: what Perplexity has said, what is common practice in systems like it, and what nobody outside the company knows.

Think of a well-run newsroom. Reporters collect and file clippings from everywhere (the index). When a question comes in, an editor pulls the few clippings that actually answer it, first quickly, then carefully (retrieval and ranking). A writer then drafts the story with a footnote on each claim (generation with citations). Perplexity's published design covers the first two jobs in some detail and the third hardly at all.

Step 1: Perplexity runs its own index of the web

Perplexity does not just borrow another search engine's results. Its engineering post, Architecting and evaluating an AI-first search API, says its index tracks over 200 billion unique URLs, with room for many hundreds of billions more. Keeping that fresh is a balancing act, because every recrawl of an old page competes with indexing a new one, so Perplexity trains a model to predict which pages need to be indexed or refreshed.

Its crawler is called PerplexityBot. Perplexity's crawler documentation says PerplexityBot exists to surface and link websites in Perplexity's search results and "is not used to crawl content for AI foundation models", and the engineering post says it follows the limits sites set in their robots.txt files. A second agent, Perplexity-User, visits pages when a user's question needs them.

Step 2: pages are cut into self-contained spans

The web is messy, so Perplexity built what it calls a content understanding module: parsing rules that adapt to each kind of site, which a language model reviews and helps improve over time. The goal is to break each document into "self-contained spans, each of which can be individually retrieved and ranked at query time."

Why bother? Because a model's context window is limited, and irrelevant text makes answers worse. Handing the model the one paragraph that answers your question beats handing it a whole page. This is the same chunking decision covered in Chunking: The Decision Everything Downstream Inherits.

Step 3: retrieval with keywords and meaning together

Perplexity says it does not choose between lexical (keyword) and semantic (meaning-based) retrieval: it queries its index both ways and merges the results into one hybrid candidate set. At this stage the goal is to miss nothing; precision comes later.

Keyword search catches exact names, codes and rare terms. Meaning-based search compares embeddings, lists of numbers that capture what a passage is about, so it finds relevant passages that use different words. Hybrid Search and Re-Ranking: The Cheapest Quality Win explains why the combination usually beats either alone. In February 2026 Perplexity also released its own embedding models, pplx-embed, which it describes as built for real-world, web-scale retrieval and makes available through its API.

Step 4: ranking in stages

Next come prefilters: simple rules that remove content that is clearly off-topic or stale. Then several ranking stages, each slower and smarter than the last. Perplexity says the early stages use fast keyword and embedding-based scorers, and the final stages use cross-encoder rerankers, models that read your question and a passage together and judge how well the passage answers it. The ranking stack also learns from the signals in the millions of requests Perplexity serves.

That is as far as Perplexity's own description goes. You may come across outside analyses that claim much more specific details, such as particular models or score thresholds. Perplexity has not confirmed them, so this article leaves them out.

The scale behind these steps, from the same engineering post:

Step 5: writing the answer with citations

Here Perplexity has published little. What you can see is the result: an answer with numbered citations pointing to the sources, and in its developer API, an Agent API that returns "web-grounded answers with built-in citations" (Perplexity API docs).

The common way systems like this produce citations, described in Grounding and Citations: Answers You Can Check, is to number the retrieved passages in the prompt and tell the model to answer only from them, marking each claim with the matching number. Because the numbers come from real retrieved pages, this stops the model from inventing a source that does not exist. It does not stop misattribution: the model can still attach a real source to a claim that source does not support. That is why you still need to click through, as Research With AI: Fast Answers You Can Actually Cite recommends.

Which model writes the answer? Perplexity has trained its own: Sonar, launched in February 2025, was "built on top of Llama 3.3 70B" and further trained for factual, readable answers (Meet new Sonar). Paid subscribers can also pick models from other companies, including OpenAI, Anthropic and Google. Perplexity's list of included models calls itself a snapshot as of 17 October 2026 and says the model selector in the app is always the source of truth.

How good are the sources it cites?

An independent benchmark gives one answer. SourceBench (Jin et al. 2026, SourceBench: Can AI Answers Reference Quality Web Sources?) rated 3,996 web sources returned or cited by 12 systems for 100 real-world queries, using eight measures: content relevance, factual accuracy, objectivity, freshness, ownership accountability, author accountability, domain authority and layout clarity. GPT-5 scored highest overall. Perplexity's Sonar Pro model was among the three lowest weighted overall scores, alongside Gemini 2.5 Flash and the Tavily search tool.

That result covers one model at one point in early 2026, not every mode of the Perplexity app. The lesson holds for every tool: a citation tells you where to look, not that the claim is true.

Quick Reference

Frequently Asked Questions

Q: Does Perplexity ever generate an answer without citing sources? It aims to cite its claims, but not every sentence gets a citation, and citations can be wrong. A real source can end up attached to a claim it does not support. Treat each citation as a pointer to check, especially for numbers and quotes.

Q: How does Perplexity handle conflicting information from different sources? Perplexity has not published how it weighs sources that disagree. Its answers often present more than one view with separate citations, but that is no guarantee. When sources conflict on something that matters, open them and judge for yourself.

Q: Can I use Perplexity's retrieval and citation pipeline with my own documents? Partly. Perplexity's API offers an Embeddings API you can use to build search over your own text, a Search API that returns ranked web results, and an Agent API for web-grounded answers with citations (Perplexity API docs). But the index, retrieval and ranking described above apply to web search, not to your own documents.

Q: Why use keyword search alongside meaning-based search instead of vectors alone? Keyword search is good at exact matches, such as a product code, a person's name or a rare term, which meaning-based search can blur. Meaning-based search finds relevant passages that use different words. Perplexity's engineering post makes the same argument: systems over-optimized for one kind of search fall short for AI answers.

Q: Does Perplexity use my questions to train its AI? On Free, Pro and Max accounts, a setting called AI Data Retention is on by default, which lets Perplexity collect your data for AI training. You can turn it off in Account settings, on the Preferences page; the opt-out covers only data collected afterwards. Perplexity says Enterprise data is never used for AI training (Data Collection at Perplexity).

Q: How does Perplexity keep its answers current for breaking news? Perplexity says freshness is a design goal: its index balances refreshing known pages against adding new ones, and a model predicts which pages need attention. Even so, very recent events can be thin or wrong in any search engine, so check the dates on the sources it cites.

Test yourself

You ask Perplexity for the market share of a software product. The answer gives a precise figure, 34%, with citation [3]. You open source [3]: it is a relevant industry article, but it never mentions 34%. Using what this article describes, explain what probably happened and what you should do next.

Answer: Retrieval worked: source [3] is on topic, which is what Perplexity's hybrid retrieval and staged ranking are built to find. The failure happened at the writing step. Numbering real passages stops a model from inventing sources, but it does not stop misattribution, where the model attaches a real source to a claim it does not support. The 34% may come from another retrieved passage, from the model's own memory, or from nowhere. Next, check the other citations for the figure. If none contains it, search for the original data, such as the analyst firm or the company's own report, and use that source instead. If you cannot find it, leave the number out. The citation told you where to look; it did not prove the claim.

If you want this kind of breakdown every week, how real systems actually work under the hood, subscribe to Internals Decoded at internalsdecoded.com.

Sources

• Introducing the Perplexity Search API (Perplexity, September 2025)
• Architecting and evaluating an AI-first search API (Perplexity, September 2025)
• Perplexity crawlers (Perplexity docs)
• pplx-embed: embedding models for web-scale retrieval (Perplexity, February 2026)
• Meet new Sonar (Perplexity, February 2025)
• What advanced AI models are included in my subscription? (Perplexity)
• Perplexity API documentation
• Data Collection at Perplexity (Perplexity)
• SourceBench: Can AI Answers Reference Quality Web Sources? (Jin et al. 2026)