Perplexity

AI search engine, which answers questions by citing sources, is called the "Google killer"

AI search

https://www.perplexity.ai/

Visit Website

In-depth Report

Perplexity AI is an AI answer engine company founded in San Francisco, USA in 2022, founded by former Google research scientist Aravind Srinivas and others. It subverts the traditional search presentation method: users ask questions in natural language, and the system searches multiple online sources at the same time, synthesizes the results into a directly usable answer, and attaches numbered, clickable inline quotes, allowing readers to return to the original source for verification with one click. This "citation-first" design has led to it being dubbed the "Google Killer" by the media since its inception, but it is closer to the truth as a "research accelerator" - it is good at quickly narrowing topics, comparing sources, and giving briefings with evidence, rather than completely replacing search engines. As of mid-2026, Perplexity has approximately 45 million monthly active users and handles approximately 1 billion queries per month; the E-6 round of financing completed in January 2026 pushed the valuation to approximately US$22.6 billion, with cumulative financing of approximately US$1.72 billion, and annualized revenue (ARR) of approximately US$450-500 million. What's more critical is the strategic shift: it will completely abandon advertising in 2026 and switch to a pure subscription model. Through the Comet browser and Computer intelligent agent, it will expand its business from a "tool that answers questions" to a "digital employee who performs tasks for users" and is fully embedded in enterprise workflows such as Microsoft 365.
Perplexity was co-founded in August 2022 by Aravind Srinivas (CEO, former Google Brain research scientist), Denis Yarats, Johnny Ho and Andy Konwinski. It is headquartered in San Francisco. The public version of the answer engine will be launched on December 7, 2022. The company was immediately favored by top institutions, with investors spanning strategic chip factories and growth funds: NVIDIA, Jeff Bezos, SoftBank Vision Fund No. 2, Accel, IVP, NEA, Databricks, Bessemer and Thrive Capital, etc. The pace of financing is extremely fast - as of mid-2026, approximately 11 rounds have been completed, with a cumulative total of approximately US$1.72 billion; the E-6 round closed on January 9, 2026, set the valuation at approximately US$22.6 billion, more than double the approximately US$9 billion at the end of 2024, and an increase of approximately 175 times from US$121 million in April 2023. On the revenue side, ARR exceeded US$450 million in the first quarter of 2026 and remained at US$450-500 million in the second quarter. Based on the current valuation, the trading multiple is about 45-50 times ARR, which is still a high premium. Strategically, starting in the spring of 2026, the company's main line will shift from "traffic engine" to "enterprise intelligence": In June, Perplexity Computer was officially embedded in Microsoft 365's Word, Excel, PowerPoint, Outlook and Teams, directly calling emails and file contexts to draft documents and analyze data, marking its attempt to become the infrastructure layer for knowledge work.
Perplexity's most distinctive engineering choices are "multi-model" and "reference-first". By default, the answer is provided by the self-developed Sonar model family - Sonar is based on Meta Llama 3.3 70B fine-tuning and is specially optimized for fast, factual, and referenced search. Officially, it has an F value of approximately 0.773 on the SimpleQA factual benchmark, and is comparable to cutting-edge models in terms of search satisfaction at a lower cost. Paying users can freely switch between the underlying models in a single session: GPT-5.2, Claude Opus 4.6, Gemini 3 Pro, Grok 4 and Sonar series; there are also two automatic modes: "Best" automatic model selection and "Research" mixed model grouping. In terms of product functions, Quick Search gives free users unlimited basic questions and answers; Pro Search takes more steps to conduct in-depth multi-source searches; Deep Research takes several minutes to crawl dozens of sources and directly produce multi-page research reports. Six Focus modes (Web, Academic, Reddit, YouTube, News, Wolfram Alpha) allow users to limit searches to vertical sources such as academic, community or video; Spaces provides a resident research space with preset instructions and uploaded files; Labs can generate small applications and dashboards from one sentence. The technical base is real-time web indexing and model synthesis. The answer "can be verified" is the fundamental difference between it and general chatbots. It is also the logical basis for its courage to completely abandon advertising in 2026 and rely on subscriptions.
The actual feeling of using Perplexity is more like a "research presentation with citations" than a "chat". You ask a question, and it searches multiple sources horizontally at the same time, gives a direct answer after synthesis, and marks the source, eliminating the process of jumping back and forth between more than a dozen tabs, comparing, and transcribing. For researchers, analysts, journalists, students, and knowledge workers, it is most commonly used as an alternative to traditional search: to quickly narrow a topic, check facts, compare sources, and find recent context. The Academic mode can provide peer-reviewed documents with citations in seconds, and Deep Research compresses the "opening of tags, reading, and synthesis" that used to take half an hour to two minutes. By 2026, the main line of experience will further move from "answering" to "doing things": the assistant in the Comet browser can read pages, fill out forms, and compare prices across tabs, and the Computer agent can continuously operate Gmail, Slack, Notion, and spreadsheets in an isolation sandbox, and complete cross-step long tasks at once. What needs to be said bluntly is that it is not a qualified substitute for Google - it is still not as smooth as traditional search in purely "finding URLs and doing daily tasks" scenarios. A more accurate value is "research accelerator". The free version already includes the complete answer engine and Comet browser, which is why it is widely used and its monthly active users have reached 45 million.
Comet is a key piece in Perplexity’s bet on the “next entry point”: a proxy browser based on Chromium with a built-in AI assistant instead of a plug-in. It will first be open to Max users with a monthly fee of US$200 in July 2025, and will be free globally from October 2. It will then be launched on Android (November 2025) and iPhone (March 18, 2026), completing four-terminal coverage. Comet Assistant resides in the sidebar and can summarize the current page, organize tab pages, and compare information across pages. Each search has clickable inline references. Cross-tab reasoning can directly compare quotes and data on different pages. Perplexity's logic in providing it for free is very clear: the browser is the starting point for the intelligent agent to "click for you and place orders for you". Whoever occupies this surface will block the upstream of the agency economy. Computer, launched in February 2026, takes its capabilities to another level - it can continuously operate hundreds of applications in a cloud isolation sandbox, and automatically select the most appropriate one among more than a dozen top models for execution; in June, it was further embedded in Microsoft 365 and became a permanent "digital employee." Risks also exist: at the end of 2025, security research disclosed the "CometJacking" vulnerability. Malicious instructions buried in web pages can silently steal email and calendar data, highlighting the problem of permission boundaries for agents operating in logged-in sessions.
Perplexity will form a "free + four levels of paid + usage billing API" structure in 2026. The free version includes unlimited basic searches, several Pro searches per day, Comet browser, file uploads and Spaces, making it a complete answer engine in itself. The Pro level is $20/month (annual payment is $16.67/month), which is the most popular choice for individuals. It provides hundreds of Pro searches per day, model switching, Deep Research, Labs, and packaged Comet Plus; however, in early 2026, Pro’s Deep Research quota was significantly reduced (from hundreds of times per day to about 20 times per month), and heavy loads were pushed to the Max level of $200/month. Max is additionally given to the Model Council (the same problem is concurrently combined with three cutting-edge models), the highest model permissions and priority support; on the enterprise side, there are Enterprise Pro ($40/seat/month, including SSO, SOC 2 Type II, data is not used for training) and Enterprise Max ($325/seat/month). Comet Plus is a $5/month publisher profit-sharing package that distributes about 80% of revenue to partners such as CNN, Condé Nast, and the Washington Post. It is said that approximately $42.5 million has been allocated. The developer-side Sonar API is billed based on usage and has zero data retention. In February 2026, the company completely abandoned advertising and switched to pure subscription. The core bet was whether subscription and per-volume revenue could outperform the inference cost of per query-this is also the sustainability bet on the valuation multiple of about 45–50 times ARR.
The overall reputation of real users is in favor of approval, but the complaints from heavy users are very specific. The positive comments focus on three points: first, the "citation of sources" allows answers to be verified, and the research is highly trusted; second, the research speed is fast, and Deep Research compresses the half-hour label jump to two minutes; third, the multi-model access and ad-free experience are cost-effective. Some users call the $20/month Pro "the most worth buying AI research tool." Academic mode and Focus mode are frequently praised by students and researchers. The negative feedback is also clear: the reduction of the Pro quota in early 2026 made heavy users dissatisfied, thinking that "the price/performance ratio has dropped significantly" and they were forced to jump to Max; the agent-style function is unstable in actual operation, and some reviewers encountered tasks that were slower than manual operations, AI hallucination errors during booking, and memory usage exceeding 4GB; occasional factual errors or quotation deviations remind users that "citation does not mean exemption." Generally speaking, it is most pleasing to the workflow of "starting with research", rather than ordinary search users who pursue omnipotence; it is almost the default entry for researchers, analysts and students who need to bring evidence presentations.
The media and analysts' judgment of Perplexity has returned from the early "Google disruptor" to the "high-growth but awkward middle class." It directly competes with Google search, ChatGPT search, Microsoft Copilot, Gemini and You.com, but the size gap is huge: in February 2026, Perplexity.ai had about 6 million daily visits, while Google's global mobile terminal had 2.2 billion daily users. In AI referral traffic, Perplexity's global share is about 7.2% (in January 2026, down about 13% year-on-year), which has been surpassed by Gemini (a surge of 231% year-on-year, ranking second); its share in the US market is about 19.7%, significantly higher than the global average, indicating that the basic market is more biased towards professional research groups. Even more alarming is the decline in overall traffic share: global web traffic share fell from 2.0% in March 2026 to about 1.3% in May. The industry uses "AI Darwinism" to describe this flow - user loyalty is weak, models are transferred as soon as they are improved, and first-mover advantages are not enough to maintain status. The core argument of the bears has always been that with the capital and distribution advantages of Google and OpenAI, whether Perplexity can enable subscription and volume-based revenue to continue to outperform the cost of inference is still an pending verification item.
Perplexity is caught between two lines of litigation: the “upstream content party” and the “downstream platform party”. On the copyright side, as of mid-2026, it has been sued by the New York Times (at the end of 2025, another Lanham Act was proposed to claim that hallucinatory content impersonates its brand), CNN (in May 2026, it said that it had crawled more than 17,000 pieces of content and generated similar original text), Dow Jones/News Group, Tribune, Encyclopedia Britannica and Webster (September 2025), and the Yomiuri Shimbun (claiming about 21.7 Billion yen), Asahi and Nikkei (about 2.2 billion yen each), Italian Mediaset and a group of writers (involved in pirated book training) sued. The core dispute "whether AI crawling is infringement" has not yet been decided, and the company defended it with "acts without will" and "fair use". On the platform side, Amazon sued Comet agents for automatically logging in to place orders, which was unauthorized access (the ban was once in effect, but the appeals court suspended it, and the court session will be held in June 2026); Reddit was also sued. There are also security risks such as CometJacking on the technical side. Applicable recommendations: It is most suitable for researchers, students, analysts and professionals who need "research with citations"; it is not suitable for those who use it as an all-in-one search substitute or who are extremely concerned about privacy. Looking forward, agencyization (Comet/Computer plus M365) is its biggest bet. There is no clear IPO timetable in the near future, and the jurisprudence of "whether AI consumption content should pay content tolls" will determine the red line of the entire answer engine category.

User Reviews

AlfonsoRolón

—

After using Comet for a week, I almost never open Chrome again. What impressed me most was the cross-tab comparison. When researching two SaaS pricing plans, I asked directly next to the address bar. It read two tabs at the same time and produced a comparison table in 30 seconds, which took at least 20 minutes of manual work. The page summary is not the perfunctory "This article introduces XXX", it can capture the key arguments and include original text quotes. However, I only dare to let the Agent automatically fill in the form for price comparison and collect information, and I will not let it touch any step involving payment.

LawrenceWilson_20235

—

This function of Model Council is really useful for making high-stakes decisions. A question is thrown to GPT-5.2, Claude Opus and Gemini at the same time, and the areas of agreement and disagreement between the three are marked. When I was struggling with a technology selection before, looking at the reasons for each of the three models was much more practical than looking at just one answer. It is only available with Max subscription, and the threshold is not low.

CWilliams0

—

Perplexity's Finance model has become more and more convenient to use recently. It can provide analyst ratings, 52-week target prices, and jump directly to the corresponding paragraphs of SEC documents. Preparing before regular investment research meetings saves a lot of time in flipping through documents. However, if you really want to make a decision, you still have to check the original data yourself, and you cannot fully trust its synthesis.

Ryan_Walker_7

—

I can’t go back, and now my first reaction when looking up something is to open Perplexity.

trueNaomiRoger_pro

—

Deep Research can now directly ship deliverables. This update is so useful. Last week, I was asked to make a PPT for comparing competing products. After running it, it directly generated a slide outline with cited sources and a comparison table for me. Although I checked the two data myself and adjusted the order, the first draft was produced in 15 minutes after half a day. It’s much smoother than searching first and then pasting it into other tools.

LoganChavez_X7

—

As a reminder, Perplexity references are not omnipotent. I came across it quoting an actual FTC press release, but the threshold mentioned was actually changed by a later ruling and it was not mentioned. Therefore, for important conclusions, you still have to click into the source file and take a look for yourself, don't just read the summary it gives.

Pathicia404

—

I only dare to use half of Comet’s Agent automation now. Let's compare the price of the same keyboard across three e-commerce companies, and the success rate is about 60%. The remaining 40% either clicked on the wrong filter conditions, or got stuck on a loading page. Once I even clicked "Add to Cart" to checkout, which scared me and I quickly stopped. So my principle is to leave the information collection and comparison to it, and the actions involving money and submission are absolutely done by yourself. In addition, privacy should also be mentioned. It can read every tab of yours. I would open another browser to isolate sensitive company systems and bank logins.

安然527

—

The free one-year program for students ended in January, and the current student price is $10 a month for Education Pro, which can be used for free for one month after verification. It's quite friendly to the student party, but it's far worse than the previous wave of wool.

ryi4vt

—

It’s free and great, Comet is free on all platforms.

brownpeacock225

—

The free version only uses Pro Search 5 times a day, and I hit the ceiling before 10 a.m. when doing intensive research. After the collision, it fell directly back to normal search, and the experience gap was quite obvious. It's light enough for use, but if you really want to do research every day, you still have to drive Pro, otherwise it will be very uncomfortable to hit the wall.

万然

—

I really think there is something good about the Perplexity Computer agent. I built a skill for weekly competitive product monitoring, which automatically pulls financing news, product updates, and recruitment signals from five startups, and sends them directly to my email on Monday morning. Not having to set up a development environment is equivalent to having a research assistant who will never resign. Of course, it’s better not to touch anything involving account and password. I’ll do the sensitive operations myself.

掠影

—

I really didn’t expect to use Comet’s voice mode every day. One time I was in the kitchen with a spatula in one hand and asked if PostgreSQL 17 had native vector search. It responded instantly with a pgvector solution. When writing code, you don’t need to cut the window to ask questions directly, which is very freeing of attention.

Raymond.Johnson

—

yyds is the first choice for checking information.

吕珍

—

I quite agree with Perplexity’s insistence on not advertising. Now both ChatGPT and Google are inserting advertisements into AI searches, but it goes in the opposite direction, saying that advertisements will erode users' trust in answers, and it will completely withdraw from the advertising business by the end of the year. After using it, it is confirmed that the search results are not interfered by the position of anyone who bought it, which is quite rare in an era when sponsored content is becoming more and more normal. Of course, the confidence lies in the increase in subscription revenue.

JacobAllen

—

Pro Whether the twenty dollars is worth it depends on what you do every day. I do research every day, and Pro's model switching and unlimited Pro Search are really inseparable; but if you mainly use it to write things and code, then ChatGPT and Claude are obviously stronger, and you will lose money by using Pro.

GMyers_2021

—

After using it for more than half a year, I came to the conclusion that Perplexity is indeed the king in "factual search with source". It saves too many clicks to search for information compared to Google's direct answer. But its writing is really flat. I once asked it to draft a product narrative. It was logical but not spiritual. Claude’s version was more interesting. There is also the issue of hallucinations that has not been completely resolved. Occasionally, the quotation is correct, but the interpretation is a little off. Key conclusions must be checked by yourself with the source documents. It is positioned as a research accelerator, not an all-round assistant.

Alan_Young_X

—

When more than 30 tabs are opened, Comet consumes more memory than Chrome, so use it with caution on older machines.

Nicholas578

—

Spaces and file uploading are the most helpful tools for my research. Group several PDFs and web page threads into one Space. The files will not be automatically deleted and you can come back to them at any time. Compared with pasting information into the chat box every time, this method of building a knowledge base by project is much more convenient. It’s just that the team version of Spaces feels like it’s not done well yet, and the cost goes up quickly when there are more people.

greenbear805

—

Answer directly with the source, which is easier to use than ChatGPT.

DAkra

—

Google Killer lives up to its name and the search experience is great.

EllieKrauss

—

The professional version is $20 a month and is worth subscribing.

AHicks_Max12

—

An essential tool for academic research.

兰花153

—

Smarter than traditional search engines.

bigsnake851

—

Copilot is very powerful.

NodeKing2025

—

The free version is enough for daily use.

z8tslfs2

—

Love the Citations feature.

JoeJames3692

—

Strong ability to obtain real-time information.

EBarnesSr03

—

The enterprise version is a bit expensive.

Perplexity

In-depth Report

User Reviews

Similar Products