Who does Perplexity AI?

Why Search Broke — and Why It Matters Who's Fixing It

Type a question into Google today and you wade through ten pages of affiliate listicles before finding a human being's actual opinion. That frustration isn't imaginary — it's measured, and it's the reason a new generation of answer engines exists at all.

A 2024 German study of roughly 7,400 product review queries reached a blunt conclusion: search results are increasingly "SEO spam," optimized to rank rather than to inform. The same research found that more than 50% of consumers reported sometimes feeling misled by search results, and a quarter often ended up somewhere unexpected after clicking.

So people migrated to AI chatbots. But that migration created its own problem: a confident, fluent answer with no sources is arguably worse than a spammy link. You can't verify what you can't trace, and AI tools that answer without citations simply shift the trust problem rather than solving it.

This is the gap Perplexity AI was built to fill. The company's founding rationale rests directly on this decline in search quality, and its CEO Aravind Srinivas has framed the stakes plainly: accuracy and transparency are "prerequisites to making AI-powered search ubiquitous." Every answer the platform produces includes citations linking back to original sources, per its own documentation, so users can verify claims rather than take them on faith.

Why does the organization behind a tool matter? Because when the underlying search infrastructure is degrading, whoever rebuilds it sets the rules for what "trustworthy information" means:

• Whether answers are sourced and verifiable, or generated blind
• Whether the business model rewards accuracy or rewards engagement
• Whether the company is independent or captured by the incentives that broke search in the first place

Notably, Sacra's research profile reports that Perplexity dropped advertising as a monetization channel, citing concerns it would hurt user trust in AI — a direct response to the incentive structure that produced the SEO-spam era.

That same verification-first logic applies beyond search. At SkillzBench, we're building a free directory where every AI agent skill is security-scanned, license-checked, and benchmarked with published results, because an unverified AI component carries the same risk as an unsourced answer. Trust in AI has to be earned with evidence, not asserted with confidence. Understanding who built Perplexity — and whether their incentives align with yours — is the first step toward knowing which AI tools deserve a place in your workflow.

The Four Founders and the Mission to 'Serve the World's Curiosity'

Long before Perplexity AI became a household name in AI search, it was four people with a simple conviction: search should give you answers, not a list of links to dig through.

Perplexity AI launched in 2022 in San Francisco, founded by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski, according to Contrary Research's company profile. Each founder brought distinct credentials to the table, and the mix proved unusually complementary for a company aiming at one of tech's most entrenched incumbents.

Srinivas, the CEO, holds a Master's in Engineering from IIT (2017) and pursued a PhD at UC Berkeley in reinforcement learning before interning at both DeepMind and OpenAI. Yarats, now CTO, connected with Srinivas over email after the two published nearly identical research papers at Berkeley and NYU. Ho joined as an ex-Quora engineer, and Konwinski brought deep infrastructure credibility as a cofounder of Databricks. The four formally started the company in July 2022.

The founding rationale was rooted in a real problem: trust in traditional search was eroding. A 2024 German study of 7,400 product review queries concluded that search results were increasingly "SEO spam," and a 2020 survey found more than 50% of consumers sometimes felt misled by what search engines served them.

Srinivas's ambition was never modest. He explicitly launched Perplexity to disrupt Google Search, inspired by Larry Page's original vision of providing direct answers rather than links. The company's stated mission — "serve the world's curiosity" — comes with a core vision of becoming "the best platform for answers and information."

That mission has a non-negotiable prerequisite. As Srinivas wrote in a January 2024 blog post: "we're built on the idea that accuracy and transparency are prerequisites to making AI-powered search ubiquitous." The mechanism is citations — every answer links back to its original sources so users can verify what they're told. In Srinivas's words, "citations are a great way to marry search and LLMs."

The founders' key bets paid off quickly:

• Perplexity Ask reached 2 million monthly active users within four months of its January 2023 launch
• The company grew from a GPT-3.5/Bing wrapper into a platform with in-house models and its own browser
• Perplexity now claims 100M+ monthly active users, per Sacra estimates

That obsession with verifiable answers is part of a broader shift in how people evaluate AI tools — one SkillzBench is building on too, with its own approach of pinning, scanning, and benchmarking agent skills so users can see exactly what a tool does before trusting it. Transparency, it turns out, is becoming the price of admission.

From GPT-3.5 Wrapper to $20B Answer Engine

Perplexity's origin story is a lesson in how fast an AI startup can outgrow its own first idea. In December 2022, the company launched Bird SQL, a tool that used OpenAI's Codex to translate natural-language questions into Twitter searches. When Twitter ended free API access in February 2023, the team pivoted to general search — and the results were dramatic.

Perplexity Ask, launched in January 2023, reached 2 million monthly active users within four months, according to Contrary Research. The early product was modest: it pulled top Bing results and passed them into GPT-3.5 to generate cited summaries. That wrapper approach allowed a fast launch without building its own crawler or index.

The company then moved deliberately up the stack. It fine-tuned open-source models like Mistral 7B and LLaMA-2, and in February 2025 released Sonar, its first productized in-house model, built on LLaMA 3.3 70B and delivering over 1.2K tokens per second through a Cerebras partnership. In July 2025 it launched Comet, a Chromium-based browser, following the acquisition of browser startup Sidekick. Enterprise offerings, Perplexity Labs, and a Computer agent that orchestrates up to 19 models in parallel followed.

The scale today is striking, though the numbers come from private-market estimates rather than audited filings. Per Sacra's research:

• Annualized revenue of roughly $750M as of August 2026, up from $232M at the end of 2025 — 268.3% year-over-year growth in 2025
• A $20B valuation as of September 2025, up from $9B in December 2024
• Approximately $1.5B in total funding from investors including Accel, SoftBank Vision Fund 2, Nvidia, and Jeff Bezos
• 100M+ monthly active users and tens of thousands of enterprise clients

The most interesting part of the story is strategic, not financial. Sacra's analysis puts it plainly: "Their bet is that owning the search data and user relationships will prove more defensible than the language models themselves." In other words, Perplexity believes the engagement data from millions of daily questions — not the underlying model — is the real moat. That's why it keeps rebuilding the search stack AI-natively, from crawler to index to ranking to personalization.

There's a lesson here for anyone evaluating AI tools: the wrapper-to-platform arc shows how quickly today's thin layer over someone else's model can become tomorrow's full-stack product. It's also why verification matters at every layer — whether that's a search engine citing its sources or, as we're building at SkillzBench, agent skills benchmarked in a sandbox so you can see each one's measured Lift before you trust it. Transparency about where answers come from is becoming the differentiator across the whole AI ecosystem.

Citations as the Business Model — and What It Means for Trust

When Aravind Srinivas explained why he built Perplexity, he didn't talk about models or compute. He talked about citations: "citations are a great way to marry search and LLMs," he said, framing transparency as the product itself rather than a feature bolted on afterwards.

That framing runs through everything the company says about itself. Its own help documentation states that every answer includes citations linking to original sources, so users can verify information or explore further. Srinivas put it more bluntly in a January 2024 blog post: accuracy and transparency are "prerequisites to making AI-powered search ubiquitous."

The monetization story tests how deep that commitment really goes. According to Sacra's private-market analysis, Perplexity has dropped advertising as a monetization channel, citing concerns that ads would hurt user trust in AI. Yet another analysis from Cognativ describes limited, relevance-focused ads supporting the free user base as part of the revenue mix. The two accounts contradict each other — and that contradiction is itself instructive. Sacra's account is more recent and specific, but the honest takeaway is that a company's transparency claims deserve the same scrutiny as its features.

For anyone adopting AI tools at work, that's the broader lesson: verify what's under the hood. Perplexity shows its sources; most AI tooling doesn't. Research on agent skills found roughly 26% carry at least one vulnerability, yet most users install them with no idea what's inside.

SkillzBench applies the citation principle to that problem. Its Verify pillar pins every skill version to a commit SHA, runs license checks and security scans, and publishes benchmark Lift scores — the percentage-point improvement a skill delivers in a sandboxed test. Money never touches scores, and Lift is always shown separately from ratings.

• Commit-SHA pinning, so the exact version you install is the version that was tested
• Security scans and license checks on every skill
• Published Lift scores, generated with the same model, with and without the skill
• A trust score built from security, quality, license, maintenance and community signals

The directory and its free MCP connector are still being built, but the waitlist is live. If Perplexity's bet — that owning user trust matters more than owning the model — proves right, the same logic will apply to agent skills. Browse the free skills directory on SkillzBench and see what verification looks like before you install anything.

How to Evaluate AI Answer Engines Like an Insider

Evaluating an AI answer engine isn't about trusting the marketing page — it's about running a few insider checks before you commit your workflow (or your budget) to it.

Start with freshness. In Zapier's hands-on testing, Perplexity retrieved articles published just 30 minutes earlier, while ChatGPT cited information from months prior. That single test tells you more about real-world usefulness than any feature list. If a tool claims real-time web access, verify it with a question about something that happened this morning.

Next, demand citations. Perplexity's own help center states that every answer includes citations linking to original sources so you can verify information — and its CEO has said "citations are a great way to marry search and LLMs." An answer engine that won't show its sources is a search engine you can't audit. Make source visibility a non-negotiable filter.

Then understand the pricing ladder before you upgrade:

• Free tier — enough for occasional questions, but limited model access.
• Pro at $20/month — the standard individual plan for regular research work.
• Max at $200/month — includes access to the Comet browser, a genuine strategic bet by the company.
• Enterprise Pro at $40/seat/month — for teams, with compliance claims worth scrutinizing.

That last point matters more than most buyers realize. Perplexity's enterprise page reports figures like 125+ hours saved per week per user and a 90% adoption rate — but these are self-reported marketing claims, not independently verified results. Treat vendor case studies as hypotheses to test in your own environment, not as proof.

The same skepticism applies to company-level numbers. Sacra's estimates put Perplexity's valuation at $20B and annualized revenue at roughly $750M, but these are private-market estimates, not audited figures. And the business model itself is contested: Sacra reports Perplexity dropped advertising to protect user trust, while another analysis still describes limited ads in its revenue mix. When sources disagree, that's your signal to probe deeper.

Finally, apply the same verification mindset to the broader AI tooling you adopt. Just as you'd check an answer engine's citations, you should check what powers your AI agents — published research has found roughly 26% of agent skills carry at least one vulnerability. That's why SkillzBench is building a free directory that verifies agent skills with security scans and published benchmark results, plus a free MCP connector to attach only the skills you choose to any agent.

When you're ready to go beyond getting answers and actually put AI agents to work, join the SkillzBench waitlist — verified agent skills and the free MCP connector are the next step.