How Perplexity Search Works: From Questions to Citations

Understanding Perplexity Search by Separating "Search" and "Generation"

Perplexity is an AI search engine that searches the web for information, generates answers based on multiple sources, and cites the pages that serve as evidence. The official Perplexity documentation also describes it as a service that searches the web and presents integrated answers and citations based on multiple sources in response to questions.

When understanding the mechanism, it is crucial not to treat the "process of finding information from the web" and the "process of the LLM creating an answer" as the same thing.

Broadly speaking, after receiving a question, Perplexity searches for the necessary information, retrieves candidate information, and narrows down the information to be used for the answer. The LLM reads this information, integrates it into an answer to the question, and presents the sources as citations.

Even in Perplexity's API, the Search API, which returns only ranked web search results without using LLM, and the function that generates answers using search results are provided separately. Viewing search and text generation as separate processes is practically important for understanding Perplexity's behavior.

How search, retrieval, and generation are connected across AI search as a whole is explained in detail in the related article "Illustrated Guide to AI Search Mechanism."

On the other hand, the entire internal algorithm of how Perplexity evaluates search candidates and ultimately adopts which pages for answers or citations is not disclosed.

Therefore, in this article, we will organize the flow from search to citation by distinguishing between facts that can be confirmed with official information and parts that can be understood from the general structure of RAG and information retrieval.

Viewing the Process from Question to Citation in Six Steps

Perplexity search can be more easily understood when divided into the following six steps.

The official Perplexity does not disclose all internal processes for general user products under these six names.

However, in the official documentation, the processes of web search, retrieval of multiple sources, integration of information, answer generation, and citation can be confirmed. Pro Search also explains the flow of performing multiple web searches, analyzing and integrating the retrieved information, and returning answers and citations.

The key point here is that the "pages searched," "pages used for answer generation," and "pages ultimately cited" are not necessarily the same set.

Being able to make this distinction allows you to consider separately whether your site is not displayed on Perplexity because it is not crawled, not a search candidate, not used for generation, or just not cited.

Questions May Be Converted and Decomposed into Search Queries

The question text input by the user into Perplexity and the search queries executed to find information from the web are not always the same.

For example,

"Which companies should a BtoB SaaS company compare when starting AI search optimization?"

This question includes multiple conditions such as "BtoB SaaS," "AI search optimization," "companies," and "comparison."

To answer such a question, it is easier to gather information by expanding the search targets or changing the search expressions for each necessary point rather than searching with just one string as it is.

The official Perplexity explains that Pro Search performs multiple searches on the web in response to questions and retrieves information from articles, academic papers, forums, videos, etc. The published example of the Agent API also confirms a structure where multiple search queries and search results are generated from a single input.

However, it cannot be definitively stated that normal Perplexity search always performs the same method and number of query decompositions for all questions.

The important thing is to consider the "question input by the user" and the "query used by the search system for information exploration" as separate entities.

If you judge based only on user input that "it is searched with just this keyword," you may narrowly define the conditions for your page to become a candidate on AI search.

Specific steps and real data are summarized in the related article "Query Fan-Out Investigation." Understanding how AI expands searches from a single question makes it easier to see why traditional SEO search keywords and AI search information retrieval units do not necessarily match.

Web Search Retrieves Candidate Information and Narrows Down Information for Answers

After query processing, the next step is information retrieval from the web.

Perplexity is officially described as a search service that retrieves current information from the web in response to questions and uses it for answers.

The structure of the search layer can be better understood by looking at the Perplexity Search API.

The Search API returns web search results that include titles, URLs, snippets, publication dates, and last update dates for searches. This API does not perform answer generation by LLM and is a function for obtaining the ranked search results themselves.

In other words, it can be viewed in two stages: first, "web pages that are candidates for search" exist, and then the information is used by the answer generation side.

However, not all pages retrieved as search results are used for answer generation as they are.

Including information with low relevance to the question or information that does not contain the necessary evidence in the answer can lower the accuracy of the answer.

Therefore, it is important to narrow down the information needed for the answer from the retrieved candidates and form the context to be passed to the LLM.

Being able to distinguish between "searched" and "used for the answer" is important in AI search analysis.

Understanding the Role Division of Search and LLM through RAG

A useful concept for understanding the difference between search and generation is RAG (Retrieval-Augmented Generation).

RAG is a mechanism widely known from a study published by Patrick Lewis et al. in 2020, which generates text using information obtained from external sources as context, rather than relying solely on the knowledge within the LLM.

In simple terms,

the search side is responsible for "what to bring as answer material," and the LLM side is responsible for "what kind of answer to create from the provided material."

The internal implementation of Perplexity is not exactly the same as the paper by Lewis et al.

However, even in Perplexity's official API documentation, web search and text generation are treated as different roles. The Search API returns search results, and the generation function generates answers using the obtained search information.

Understanding this role division helps avoid judgments like,

"Because it's ranked high on Google, it will definitely be cited on Perplexity,"

"Because PerplexityBot crawled it, it will be used for the answer,"

"Because it was used for answer generation, it will definitely be displayed as a citation."

Search, retrieval, generation, and citation are continuous processes, but they are not a single process that passes through the same evaluation as it is.

Not All "Retrieved Pages" Are Cited

One of the major features of Perplexity is the citation that allows you to verify the sources along with the answer.

However, when understanding citations,

it is necessary to separate the retrieved pages from the pages ultimately displayed as citations.

The Agent API of Perplexity reveals a structure where search results are first retrieved, and then the model generates text and associates citation numbers in the generated text with the search results.

From this published specification, it can be understood that, at least on the API, "retrieval of search results" and "attaching citations to generated text" are not the same.

From the perspective of your site, it is easier to judge by considering the state as follows:

Crawlable or fetchable
↓
Retrieved as a search candidate
↓
Used as information for answer generation
↓
Displayed as a citation in the answer

Just because the previous stage is passed does not mean it will necessarily proceed to the next stage.

Moreover, the complete formula of which ranking signals Perplexity evaluates for each page and under what conditions the final citation is determined is not disclosed.

Therefore, in AI search optimization, it is more appropriate to observe which stage your site has reached and separate the causes rather than definitively creating "ranking factors for being cited."

PerplexityBot and Perplexity-User Have Different Roles

When considering how your site is retrieved by Perplexity, it is necessary to distinguish between PerplexityBot and Perplexity-User.

The official Perplexity documentation defines these two as separate User Agents.

PerplexityBot is a search crawler that collects and indexes information to display websites in Perplexity's search results.

On the other hand, Perplexity-User is a user-initiated fetcher that accesses the necessary web pages in response to user questions.

Furthermore, the handling of robots.txt is not the same.

For PerplexityBot, site operators can manage access with robots.txt. Meanwhile, Perplexity-User, being a user-requested retrieval, is generally explained in the official documentation to ignore robots.txt instructions.

For sites using WAF, it is advised not to permit access based solely on User Agent but to verify in combination with the IP range published by Perplexity.

From this, it is clear that the technical issue of "whether it is retrieved by Perplexity" and the content/search issue of "whether it is cited after retrieval" should be checked separately.

The Difference Between Google Search and ChatGPT Search Lies More in "Experience Design" Than "Entry"

Previously,

"Google returns links to search results,"

"Perplexity returns AI-generated answers,"

was enough to explain the differences between the services to some extent.

Currently, this distinction alone is insufficient.

Google Search provides a search experience that combines AI-generated answers using web information and links to web pages through AI Overviews and AI Mode.

ChatGPT Search also performs web searches and provides answers using the latest information and links to sources. The official OpenAI documentation explains that user input may be rewritten into one or more search queries and sent to the search service.

Thus, as of 2026, it has become difficult to distinguish Perplexity, Google, and ChatGPT solely based on whether they search the web.

When looking at differences, it is necessary to consider the design aspects such as,

"How search results are integrated into answers,"

"How sources are presented,"

"How far exploration can continue with additional questions,"

"What kind of experience is provided for search and generation."

Perplexity can be understood as a service that places web search, answer generation, and citation confirmation as a continuous information exploration experience at the forefront.

Pro Search and Research Have Different Depths of Search

When investigating Perplexity's search behavior, it is also important not to treat normal search, Pro Search, and Research under the same conditions.

Pro Search is an advanced search function that performs multiple searches on the web for complex questions, reads diverse sources, integrates information, and creates answers.

Research is a function that further deepens exploration by repeating search, document reading, and reasoning about what to investigate next, ultimately integrating the research content into a report.

The important point here is not the sheer number of searches.

The depth of investigation into the question and the process of information exploration change depending on the search mode.

Therefore, when observing the citation status of your site, it is safer not to compare the results confirmed with normal search and those confirmed with Research as the same population.

Only by fixing the mode, question, and timing of the measurement can it become easier to compare citation changes.

Confirmation Points to Avoid Misjudgment in Perplexity's Answers

Having citations makes it easier to verify the answers.

However, the existence of citations itself does not guarantee that all the content of the answer is correct.

Perplexity itself recommends checking the original sources when accuracy is important.

When confirming, at least separate the following three points:

Does the Cited Page Actually Have the Evidence?

Check if the content stated in the answer text is really written on the cited page.

Even if the correct page is cited, the summary or integration by the LLM may result in stronger expressions than the original text or missing conditions.

Is the Information Date Consistent with the Question?

For information that changes over time, such as fees, systems, product specifications, news, and officials, check the publication or update date.

Even if the source itself is reliable, if the timing of the question and the information do not match, it may not be an appropriate answer.

Can You Confirm a Source Closer to the Primary Information?

For corporate information, the official site; for systems, administrative agencies; for research, the original paper, etc., the closer you get to the original information, the easier it is to enhance confirmation accuracy.

For how to prevent AI mis-summarization or mis-citation on your content side, refer to the related article "Risk Management and Practical Avoidance Techniques to Prevent Misinformation in AI Measures."

Points to Consider for Your Site to Be Recognized and Cited by Perplexity

When adapting your site to Perplexity search, starting with "how to get cited" alone makes it difficult to separate causes.

The first thing to check is whether your page is in a state that can be retrieved by Perplexity.

Check if PerplexityBot is unintentionally blocked by robots.txt, if WAF is blocking legitimate access, if the HTTP status is normal, and if the main information can be read on HTML. The official Perplexity also guides sites intending to be listed in search results to allow PerplexityBot and check the published IP range.

Next, check if, when the page becomes a search target, it is clear "what the page is about" and "what it can answer regarding the question."

Prepare specific information that AI can use as the basis for answers, such as titles, headings, body text, product specifications, fees, comparison conditions, survey data, primary information, and update dates.

Then, input actual questions into Perplexity and observe states such as,

"Your page does not appear,"

"It appears as a search candidate but is not cited,"

"There are questions where it is cited and questions where it is not."

By separating crawling, becoming a search candidate, being used for answers, and being cited in order, it becomes easier to determine whether it is a technical retrieval issue or an information design issue.

If you want to understand why Queue Inc. designs AI search measures based on RAG and query fan-out, the related article "What Does Queue Inc. Do? An Overview of the Company Providing AI Search Optimization SaaS "umoren.ai"" introduces the background of research and implementation.

If you want to check which questions your company is exposed to and which sources competitors are cited from, you can check the status on AI search with the current analysis of umoren.ai.

Frequently Asked Questions

Does Perplexity Search the Web Every Time a Question Is Asked?

Normal Perplexity is officially described as an AI search engine that retrieves current information from the web at the time of questioning.

However, the data handled changes depending on the mode used and the connected information sources, so it is more appropriate not to generalize to "always searches only the public web under all usage conditions."

Does Perplexity Have Its Own Search Index?

Perplexity publishes PerplexityBot as a crawler to display websites in search results.

However, the entire structure of the search index for general user search and the role division of each search infrastructure are not completely disclosed.

Therefore, it is best to avoid concluding "completely independent search engine" or "uses only specific external search engines."

Are Search Rankings and Citation Rankings the Same in Perplexity?

They cannot be judged as the same.

The Perplexity Search API can only obtain ranked web search results. On the other hand, the generation-based API returns answers and citations using search results.

At least in the published specifications, the process of obtaining search results and the process of presenting citations in the answer are separate.

Therefore, it is safer not to directly equate being ranked high in search with citation ranking.

If There Is a Citation, Is the Answer Correct?

A citation is a clue to verify the content of the answer, but it does not guarantee accuracy itself.

It is necessary to open the cited page and check whether the answer text is truly supported, whether the information is not outdated, and whether conditions or exceptions are not omitted. The official Perplexity also recommends checking the original sources for important information.

Does the Search Result Change with Follow-Up Questions?

It may change.

Perplexity allows you to delve deeper into questions while continuing the conversation, and Pro Search and Research perform multiple searches and additional information exploration.

If the content of the question or the premises change, the necessary information also changes, so the same sources are not necessarily fixed and cited every time.

Conclusion

When understanding the mechanism of Perplexity search, it is important to think of search and generation, retrieval and citation as separate processes.

Perplexity receives a question, searches for web information, retrieves candidate information, and forms the necessary context for the answer. Then, the LLM generates the answer text and presents the sources as citations.

Understanding this structure helps avoid judgments like,

"Because it was searched, it will be cited,"

"Because it was crawled, it will be used for the answer,"

"Because it is ranked high on Google, it will be cited on Perplexity."

When analyzing your site, first confirm "whether it can be retrieved," then observe in order whether it "becomes a search candidate," "is used for answer generation," and "is cited," making it easier to identify areas for improvement.

For practical AI search measures, the related article "Practical Strategy Guide to Become "Chosen Information" in Search Results with AI Overview Tools" summarizes practical confirmation points.

AI Search Practices to Proceed After Understanding Perplexity's Mechanism

Even if you understand the processing flow of Perplexity search, deciding where to apply it in practice is a separate decision. From the next articles, concretize content design, measurement, and prioritization.

• How Should Owned Media Change in the AI Search Era? Design Strategy and Practical Guide Not Dependent on PV
  AI search-oriented site design method when wanting to break away from PV dependency

• Practical Guide to Integrate LLMO and Content Marketing: SEO × AI Optimization for a Cited Site Design
  Integration design method when unable to balance SEO and LLMO

• What Should Companies with Few Branded Searches Do First with LLMO? Practical Steps for Cited Primary Information and Structuring in AI Search
  Practical steps for LLMO for companies with few branded searches

• How Will Inquiry Routes Change with the Spread of AI Search? Attraction Strategy to Achieve Results Even with Reduced Inflow
  AI search indicator sorting method when confused by KPI judgment due to reduced inflow

• Explaining the Period Estimate Until LLMO Measures Show Results and Measures to Accelerate Effectiveness and Timeline
  Priority of measures when the timing of LLMO results is unreadable