So, when you type a question into Perplexity, two seconds later, you get a clean paragraph with little numbered citations hanging off the end.
It feels like talking to one very well-read AI.
But it isn't one AI. Behind that single answer box, Perplexity is routing your question across a stack of rival models, such as OpenAI's GPT, Anthropic's Claude, Google's Gemini, xAI's Grok, and stitching the live web into the reply.
The genius of Perplexity is that you never see the machinery.
And Perplexity's business is that it owns almost none of it.
Here is how the machine actually works, and why a company that builds neither the models nor the content underneath is worth $22.6 billion.
Let’s go!
The Value Prop: An Answer, Not Ten Blue Links
Aravind Srinivas and three other co-founders founded Perplexity in August 2022. Srinivas had a Berkeley PhD and internships at OpenAI, DeepMind, and Google.
The original idea was almost boring.
It was a tool that read several web pages and summarised the answer.
They named it after "perplexity," a machine-learning term for how well a model predicts the next word. The bet was that people don't want links.
They want the answer, with receipts. And it worked fast.
From a $520 million valuation in early 2024, it hit $9 billion by December, around $18 billion by mid-2025, and roughly $22.6 billion in 2026. Revenue also rocketed. About $100 million in annual recurring revenue in early 2025 to over $450 million by March 2026, up roughly 335% in a year.
More than 100 million people now use it every month.
When you ask Perplexity a question, five things happen in about a second.
It parses what you actually mean.
It then searches the live web, using both old-school keyword matching and modern semantic vector search.
It reranks the results through layers to surface the few passages that matter.
It assembles those passages into a prompt with citations already wired in.
Then it hands that prompt to an LLM for the final answer, constrained to the evidence it just retrieved.
This last step is the trick. Instead of insisting on its own model, Perplexity runs a "model-agnostic router," a system trained with reinforcement learning.
It chooses the model to answer based on speed, cost, and difficulty.
It keeps an in-house family (Sonar, built on Meta's open Llama) for cheap, fast, cited answers, and routes harder questions to frontier models from OpenAI, Anthropic, Google, xAI, and others. Pro users can even pick the model by hand.
In February 2026, it went further with "Model Council": ask one question, several models answer, and a synthesiser reconciles them into a single response.
Its agent product, Computer, launched in February 2026 at $200 per month, orchestrates as many as 19 models for long, multi-step tasks.
Why Owning Nothing is The Strategy
Here is the counterintuitive part. Perplexity sits on top of two things it largely does not own: the models and the content. That looks fragile. But it is actually the whole strategy.
Because Perplexity is model-agnostic, it wins the model price war without fighting in it.
Every time GPT, Claude, or Gemini gets cheaper, faster, or smarter, Perplexity's product gets cheaper, faster, or smarter for free. It is arbitraging a war between giants.
The hard, expensive layer of training frontier models is someone else's problem.
The layer Perplexity owns (retrieval, routing, citations, speed, the clean interface) is the layer that compounds with every model release.
That is the lesson: you do not have to own the hardest layer. You have to own the layer that gets more valuable as the rest of the stack commoditises.
The Google Question
If answer engines are the future, why hasn't Google, which has the models, the index, and billions of users, crushed Perplexity?
Srinivas's answer is sharp: "Google has had two years to kill Perplexity and hasn't."
His theory is the classic innovator's dilemma. Google makes its money selling ads against links. A perfect direct answer removes the links and the reason to click an ad.
As he put it, if AI tells you the basketball score, "how can you sell Ticketmaster ads?" The incumbent's revenue model is the startup's opening.
Also read: Will Perplexity Kill Google?
The Dark Side of the Machine
There is a catch in a machine built on other people's models and other people's content: the content owners are fighting back.
In 2025, Cloudflare published research claiming Perplexity dodged "do-not-scrape" blocks by disguising its crawler, changing its user-agent to look like an ordinary browser.
Perplexity called the report a "sales pitch."
Then the lawsuits came: the New York Times, News Corp (the Wall Street Journal and the New York Post), Nikkei, Encyclopaedia Britannica, and even Reddit have all aimed.
The answer engine needs the open web to work. The open web is suing it. A moat built on assets you don't own is a moat someone else can drain.
In a Nutshell
Perplexity looks like one smart AI. It is really a brilliant piece of plumbing.
A router and a retrieval pipeline that rents intelligence from its rivals and borrows facts from the web, then wraps both in the cleanest answer in the business.
That is enough to be worth $22 billion, and to get sued by half the internet. The lesson isn't "build a model." It's "own the layer that others make more valuable for you."
P.S. This Sunday (5th July) I'm building the content engine that creates content for all my channels. Five agents that find the best reels in your niche, rank them, write you 20 scripts, and get sharper the more you post. Last one held 95 people for 90 minutes on a Sunday. This one's bigger. Live only, no replay. Save your seat here.