What Is Perplexity AI? Answer Engine Explained

Perplexity AI is an "answer engine": a search product that reads live web results and has a large language model write a short answer, with numbered citations pointing to the sources. Where Google returns a page of links to sort through, Perplexity returns a written response, suggested follow-up questions, and a source list. That design makes it fast for research and gives you a way to check its work, but it also moves new risks into the open: wrong attributions, publisher disputes, and a business model still being worked out.

Where Perplexity came from

Perplexity AI was founded on August 22, 2022, by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski. Srinivas is the chief executive. According to Wikipedia's company entry, the search engine launched publicly on December 7, 2022, and by February 2023 the company reported two million unique visitors.

The funding record shows how quickly investors rated the idea. The same entry lists $165 million raised in April 2024 at a valuation above $1 billion, a $500 million round in June 2025 at $14 billion, and a $20 billion valuation in September 2025. A 2026 report in The Decoder put the figure at $18 billion and said the company reports more than 100 million users, while Wikipedia lists $21.21 billion after a later round. Valuations and user counts for private companies are self-reported or leaked, and they differ by source, so treat them as rough indicators of momentum rather than audited numbers.

How a Perplexity answer is built

Perplexity's own Help Center describes four steps: it interprets your question with language models, searches the web in real time, summarizes the most relevant findings, and returns an answer with numbered citations linking to the original sources. The company does not publish the details of its retrieval system, so the description below combines that official outline with the general technique it belongs to.

That technique is retrieval-augmented generation, usually shortened to RAG. A plain language model answers from patterns it absorbed during training, which stops at a cutoff date. A RAG system first fetches relevant documents, then hands them to the model along with your question and asks it to answer from that material.

The easiest analogy is the difference between a closed-book and an open-book exam. A chatbot without search is a student who studied hard and must answer from memory. An answer engine is the same student allowed to walk to the library mid-exam, pull three books, and write the answer with page references. The student's reasoning is unchanged, but the facts are fresher and checkable.

Step by step

1. Query understanding. Your question may be rewritten into one or more search queries. A vague question like "is it worth switching" becomes concrete once the system uses the conversation so far.
2. Retrieval. The system searches an index of web pages and pulls candidate documents or passages. Quality here caps everything downstream, because a model cannot cite a page it never retrieved.
3. Synthesis. A language model reads the retrieved text and writes the answer. Each sentence is tied to one or more sources so that a footnote can be attached.
4. Presentation. The interface shows the answer with numbered links, a source list, and suggested follow-ups that carry your context forward.

Why citations are a design feature, not a guarantee

A citation shows where the model claims it got a statement. It does not prove the statement is correct, because two failures are possible. The source can be wrong, since the web contains plenty of errors. Or the source can be right but not actually support the sentence it is attached to. The second failure is the one that surprises people, and the legal section below shows it is serious enough to appear in a federal lawsuit.

The related explainer on why AI search tools cite sources differently looks at how products differ in what they cite and how they link claims to text.

The models behind the answers

Perplexity is not tied to a single model. It uses its own Sonar models and lets paying users pick from third-party models.

The company announced Sonar on February 11, 2025. Its blog post says Sonar is built on Llama 3.3 70B, further trained to improve factuality and readability, and that it ran at 1,200 tokens per second on Cerebras inference hardware, which the company called nearly 10 times faster than comparable models such as Gemini 2.0 Flash. A token is a chunk of text, roughly a short word or word fragment. The same post says the API version did not yet run on Cerebras at that time, so the speed claim applied to the consumer product. These are company-reported figures from a launch post, not independent measurements.

Wikipedia also lists R1 1776, an in-house model based on DeepSeek R1, among Perplexity's models. For paid plans, the Help Center says users can choose advanced models from other labs, naming OpenAI's GPT series, Anthropic's Claude, and Google's Gemini. The exact versions change frequently, so check the model selector instead of relying on a list in an article.

That flexibility matters for a practical reason. Perplexity's value is the retrieval and citation layer wrapped around a model, so the product can swap in whichever model performs best. It also means a Perplexity answer is only as good as the combination of retrieval and model, which can differ from the same model used in a plain chat.

Modes and products

Perplexity has grown beyond a single search box. The list below separates the main surfaces.

• Search and Pro Search. The default experience is a fast answer. Pro Search, according to the Help Center, researches across multiple sources and produces more detailed answers, and lets you choose the model.
• Research (Deep Research). A multistep mode covered in the next section.
• Create files and apps. A Pro feature that, per the Help Center, builds reports, spreadsheets, dashboards, and simple web apps from natural-language prompts.
• Spaces. Saved collections of threads and files, useful for ongoing projects.
• Comet. Perplexity's AI browser. Wikipedia dates its launch to July 2025, initially limited to the top subscription tier, and says it became free to download in October 2025.
• Perplexity Computer. A multi-model agent system listed by Wikipedia as launching in February 2026.
• Search and Sonar APIs. Developer access, priced in the table below.

Deep Research, step by step

Perplexity launched Deep Research on February 14, 2025. According to the launch post, it works in stages: it searches, reads documents, reasons about what to look for next, and refines its plan as it goes before synthesizing a report. The post says most tasks finish in under three minutes, and the Help Center describes reports as taking about two to four minutes. Reports can be exported as a PDF or document, or converted into a Perplexity Page.

The company reported that the mode scored 21.1 percent on Humanity's Last Exam, a benchmark of more than 3,000 expert-level questions across more than 100 subjects, and 93.9 percent on SimpleQA, a factuality benchmark. Those are the company's numbers from launch day. Benchmarks are useful for comparison, and they measure performance on a fixed question set, not accuracy on whatever you ask. The explainer on deep research tools compares the approach across products.

Why does a multistep approach help? A single retrieval pass is like asking one question at a library desk. A research loop is like a researcher who reads the first source, notices a gap, and goes back with a better question. The cost is time, and more steps also create more chances for an early mistake to carry through.

What Perplexity costs

Plans and prices change often, and Perplexity's official pricing page is the authoritative source. The figures below come from a Finout pricing analysis dated April 27, 2026, which says its details were correct as of April 2026.

Plan	Price	Notable inclusions
Free	$0	Unlimited basic searches, about 5 Pro Searches a day
Pro	$20 per month or $200 per year	Unlimited Pro Search, advanced models, image generation, $5 monthly API credit
Max	$200 per month or $2,000 per year	Unlimited Labs and Research, Perplexity Computer, Comet, highest-tier models
Enterprise Pro	$40 per seat per month	Team features, SSO and SCIM, SOC 2 Type II, no training on customer data
Enterprise Max	$325 per seat per month	Everything in Enterprise Pro plus higher limits and Comet Max

The Help Center confirms the structure, listing Pro features and describing Max as a higher tier with access to all advanced models and early access to new features, though the Help Center page itself does not state prices.

Developer pricing

The same analysis lists per-million-token rates for Perplexity's Sonar API models. Sonar is listed at $1 for input and $1 for output, Sonar Pro at $3 and $15, and Sonar Reasoning Pro at $2 and $8, with additional per-request fees that depend on how much search context is used. Perplexity's pricing documentation separately lists a Search API at $5.00 per 1,000 requests, or $1.00 per 1,000 with Fast Search. Developers should confirm current rates before budgeting.

Token fees and request fees are separate meters, which is why a single "per request" figure can mislead. A call that pulls in more search context costs more in both.

How an answer engine differs from other tools

Search products now blur together, so it helps to separate what each design is for.

Design	What you get	Main strength	Main risk
Traditional search	Ranked links	Discovery, navigation, local results	You do the reading and synthesis
Answer engine (Perplexity)	Written answer with numbered citations	Fast synthesis you can verify	Claims attached to sources that do not say them
Chatbot without search	Answer from training data	Reasoning, drafting, code	Out-of-date or invented facts
Chatbot with browsing	Answer that may search when triggered	Flexible, general-purpose	Retrieval is optional, so freshness varies

Perplexity's differentiator is that retrieval is not optional. The product was built around search from the start, whereas general chatbots added search later. In practice that heritage shows up in how dense and prominent the citations are.

For a direct comparison of search products, the guide to the best AI search engines ranks the field, and the explainer on AI search versus traditional search covers what changes for users and publishers. Publishers also care about how these tools choose what to cite, which is the subject of generative engine optimization.

Where answer engines go wrong

Answer engines fail in patterns, and knowing the patterns makes you a better user.

• Citation drift. The cited page is on the topic but does not contain the specific claim. The model paraphrased several sources and attached one footnote.
• Weak source selection. The retrieved pages include forum posts, content farms, or outdated pages, and the model gives them equal weight.
• Stale or partial pages. Retrieval may see a cached or truncated version of a page, so the answer misses a correction or a caveat.
• Overconfident synthesis. When sources disagree, a smooth paragraph can hide the disagreement.
• Hallucinated specifics. Numbers, quotes, and names can still be invented, particularly when retrieval returns little.

A worked example: checking one claim

Suppose you ask an answer engine how long a particular regulation has been in force and it replies with a date and a footnote. The checking steps are mechanical. Open the footnote and search the page for the date. If the date appears, note whether it is in the regulator's own text or in a commentator's summary. If it does not appear, the citation has drifted, and the date may have come from another retrieved page or from the model's own training data.

Next, look at the other sources in the list. If three independent pages agree, confidence rises. If they conflict, the smooth sentence in the answer has hidden a dispute, and the right move is to read the primary document. The whole check takes about a minute, and it is the habit that turns the product's main feature, visible sources, into actual protection.

A short verification routine

None of this means the answers are unusable. It means the footnotes are there to be used. A quick routine for any claim you plan to rely on is to click the citation, find the exact sentence that supports the claim, check whether the source is primary (a study, a filing, an official statistic) or a secondary summary, and compare against a second source when the stakes are high. For health, legal, or financial questions, an answer engine is a starting point for finding sources, not a substitute for professional advice.

Publisher disputes and lawsuits

Perplexity's approach depends on reading publishers' pages and summarizing them, which has produced conflict. The timeline below is drawn from Wikipedia's company entry unless noted.

Date	Event
June 2024	Forbes publicly accused Perplexity of copying its content (criticism, not a lawsuit)
October 2024	The New York Times sent a cease-and-desist letter; Dow Jones and the New York Post filed suit
August 2025	Cloudflare reported that Perplexity used undeclared crawlers that ignored no-crawl directives (a research finding, not a lawsuit)
August 8, 2025	Japan's Yomiuri Shimbun sued; Asahi and Nikkei followed later that month
September 2025	Encyclopaedia Britannica and Merriam-Webster sued
October 2025	Reddit sued in federal court in New York

The most detailed ruling so far is in the Dow Jones case. According to a Loeb & Loeb summary, the publishers of the Wall Street Journal and New York Post allege copyright infringement on two grounds: copying their works into the retrieval index, and producing outputs that reproduce portions of articles verbatim. A third claim under the Lanham Act alleges that Perplexity's outputs contain invented text attributed to the publishers' brands.

In a ruling summarized in August 2025, the Southern District of New York denied Perplexity's motions to dismiss for lack of personal jurisdiction and improper venue, and its alternative request to transfer the case to California. That ruling concerned where the case may proceed. The summary does not report any decision on the merits, and the case had not reached one at the time of the sources reviewed. Nothing in these filings is a finding that Perplexity infringed anyone's copyright.

The Lanham Act claim is the one most relevant to readers. It treats a false attribution as potential brand harm to the publisher, which is the legal version of the citation-drift problem described above.

A bigger swing: the Chrome bid

On August 12, 2025, Perplexity offered $34.5 billion for Google Chrome, according to Wikipedia. The offer is a reminder that the company's ambitions extend past search, toward owning the interface through which people browse and ask questions, which is also the logic behind Comet.

Advertising and the business model

Perplexity was one of the first generative AI companies to test advertising, beginning in 2024. In February 2026, the Financial Times reported, as summarized by The Decoder, that the company was retreating from ads because they made users question every answer. The Decoder quotes a Perplexity executive describing the company as being in the "accuracy business," and says the company is betting on subscription plans priced from $20 to $200 per month. Different outlets describe the wind-down as gradual rather than instant, so the accurate summary is that Perplexity announced it was moving away from advertising during 2026, not that ads vanished overnight.

The strategic logic is straightforward. In traditional search, ads sit beside links the user chooses between. In an answer engine, the answer itself is the product, so an ad inside or next to it can look like an endorsement. The trade-off is that subscriptions must cover the cost of running search plus large-model inference for every query, which is one reason a free tier is limited.

Strengths and weaknesses in practice

Where it tends to work well

• Current events and fast-moving topics, because retrieval is built in.
• First-pass research where you want to see sources quickly and decide which to read.
• Product and feature comparisons that depend on current pages.
• Summarizing a topic across many pages, followed by manual verification.

Where it tends to struggle

• Tasks that need long, careful reasoning without any web input, where a dedicated chat model may do better.
• Niche topics with thin or low-quality sources, where retrieval has little to work with.
• Anything where the exact wording of a source matters, such as legal text or medical guidance.
• Paywalled material, which raises both access and ethical issues.

A useful rule is to treat Perplexity like a fast research assistant who always shows the sources. You still read the sources that matter.

Common misconceptions

"Perplexity has its own giant model." It has its own Sonar models, built on Llama 3.3 70B in the first generation, but it also routes to third-party models. The retrieval and citation layer is the core of the product.

"Citations mean the claim is verified." Citations show provenance, not truth. A claim can be attached to a page that does not support it.

"Answer engines will end the web." Answer engines depend on pages published by others, and publishers' reaction, from licensing deals to lawsuits, shapes what they can read. The relationship between the two is unsettled.

"It is just a chatbot with search." The architecture differs: retrieval comes first on every query, and the interface is organized around sources. Whether that distinction matters depends on how you use it.

The takeaway

Perplexity's contribution to search is a shift in what a result looks like: from a list of places to look to a sourced answer that tries to save the looking. That works best when you treat the footnotes as the product and the paragraph as a draft. The company's own history shows the tensions in the model, from a retreat from advertising because it undermined trust to lawsuits from publishers whose pages feed the answers.

For readers, the durable lesson is about method rather than brand. Any answer engine, Perplexity or otherwise, is a retrieval system plus a language model, and its reliability is limited by both parts. Check the source before you rely on the sentence, and expect the products, prices, and legal landscape to keep moving.

Frequently Asked Questions

Q: Is Perplexity AI free? A: Yes, there is a free tier with unlimited basic searches and a limited number of Pro Searches per day, about five according to one 2026 pricing analysis. Paid plans start with Pro at $20 per month and go up to Max at $200 per month.

Q: What models does Perplexity use? A: It uses its own Sonar models, the first of which was built on Llama 3.3 70B, and lets paid users choose advanced models from other labs such as OpenAI, Anthropic, and Google. The exact lineup changes, so check the model selector in the app.

Q: How is Perplexity different from ChatGPT? A: Perplexity is built around retrieval: it searches the web and cites sources as the default behavior. A general chatbot answers from its training data and may or may not search depending on the product and mode. Both can be wrong, so checking sources matters with either.

Q: Does Perplexity hallucinate? A: Retrieval reduces some errors but does not eliminate them. A citation can point to a page that does not support the claim, and a Dow Jones lawsuit alleges that invented text was attributed to publishers. Click through to the cited page before relying on a specific fact.

Q: Who founded Perplexity AI? A: Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski founded the company on August 22, 2022. Srinivas is the chief executive.

Q: What is Perplexity's Sonar API? A: It is a developer API that returns web-grounded answers from Perplexity's Sonar models. Pricing combines per-token rates and per-request fees, so costs depend on the model and the amount of search context requested.

Q: Does Perplexity show ads? A: Perplexity tested ads from 2024, and in February 2026 reports said the company was dropping them over user trust concerns. Reports describe the removal as gradual, and the company now leans on subscriptions.

Sources

• Perplexity Help Center: How does Perplexity work?
• Perplexity: Introducing Perplexity Deep Research (February 14, 2025)
• Perplexity: Meet new Sonar (February 11, 2025)
• Perplexity API pricing documentation
• Finout: Perplexity pricing in 2026 (April 27, 2026)
• The Decoder: Perplexity pulls advertising from its search engine (February 18, 2026)
• Loeb & Loeb: Dow Jones & Company v. Perplexity AI (August 2025)
• Wikipedia: Perplexity AI