Perplexity

By Bill Hartzer · Updated 19 August 2026

What Perplexity is, and where its results actually come from

Perplexity is an answer engine. You ask a question in ordinary language and it returns a written answer with numbered inline citations pointing at the pages it drew from, plus suggested follow-up questions. There is no list of ten blue links; the links exist as supporting evidence for prose. That format was Perplexity's core contribution to the field, and everyone else has since copied it.

The belief most readers arrive with is that Perplexity is a wrapper around Bing. That was substantially true once and it is out of date now, which makes it the most common and most stale misconception about the product.

In its early period, roughly 2022 to 2024, Perplexity was widely and credibly reported to be serving retrieval through the Bing Search API, with the differentiation sitting entirely in the language-model summarisation layer rather than in finding the documents. Today it documents its own crawler and publishes figures for its own index. Its engineering write-up of September 2025 describes "an exabyte-scale index and crawling apparatus" tracking "over 200 billion unique URLs" across tens of thousands of CPUs and hundreds of terabytes of RAM.

One caveat belongs in the same breath. Perplexity has never published a dated statement saying it stopped using third-party search feeds, and it has never affirmatively claimed it uses none. For structured verticals — maps, weather, finance, shopping — partner APIs are undisclosed and almost certainly involved. What can be said with confidence is that general web retrieval now runs on infrastructure Perplexity built and operates.

The crawlers, the index, and the transition nobody dated

Perplexity documents two automated agents, and the difference between them matters more than any other technical detail on this page.

PerplexityBot is the indexing crawler. It publishes a full user-agent string and a live JSON endpoint listing its IP ranges. Perplexity states that what it fetches is used for surfacing and linking sites in results, not for training models.
Perplexity-User is a live fetcher: it retrieves a specific page in response to a specific question a person just asked. Perplexity's own documentation states that this agent generally ignores robots.txt, on the reasoning that the request is user-initiated rather than a crawl. Publishers dispute that reasoning, and it is the root of most of the conflict described further down.

On the index itself, Perplexity's public claims are specific: hundreds of billions of webpages covered, over 200 billion unique URLs tracked with capacity for many hundreds of billions more, and tens of thousands of indexing operations per second to keep results fresh. The ranking pipeline it describes is multi-stage — lexical and embedding scorers generate candidates, cross-encoder rerankers refine them, and retrieval happens at both document and passage level. That last detail is telling: passage-level retrieval is an architecture built to feed a language model, not to render a results page.

A structural event corroborates the independence claim. Microsoft retired the Bing Search APIs entirely on 11 August 2025, decommissioning existing instances and pushing licensees toward a considerably more expensive Azure product. Every product that had been reselling Bing had to move or break. Perplexity's own index was already in place, and the product did not skip. A wrapper could not have survived that date intact.

How it makes money — including the reversal

Perplexity has four revenue lines, and the mix has shifted materially:

Consumer subscription — a free tier with capped access to the strongest models and features, and paid tiers unlocking model choice, higher Deep Research volume, document search and API credit. This is now the primary line.
Advertising, introduced and then abandoned. Perplexity launched sponsored and AI-integrated ad formats in late 2024. In February 2026 it discontinued that advertising strategy and moved to a subscription-first model. This is a significant and underreported reversal: the flagship AI answer engine tried advertising and backed out roughly fifteen months later.
Metered APIs — the Sonar API for grounded generation and, from 25 September 2025, a Search API selling raw access to the index. Selling index access is only possible because it owns an index.
Commerce and agentic shopping — merchant-side monetisation around its browser's buying agent, which is legally embattled.

There is also a publisher revenue-sharing programme that shares revenue with cited publishers. It is best understood as a response to copyright complaints rather than a core business line.

Prices are not stated on this page and should not be. Tiers have changed repeatedly, and a page that prints a number is wrong within the year. The structural point is more durable: Perplexity is now funded by the people using it rather than by advertisers, and its own valuation marks — from roughly $1 billion in April 2024 to around $21.2 billion in early 2026 — sit alongside a query volume that is a rounding error against Google's. Both facts are true at once, and the gap between them is the interesting part.

The publisher conflict, and why it is not just complaining

Perplexity's user-facing privacy posture is unremarkable to adequate: it says crawled content is not used for model training, offers account controls to exclude chats from training, and sells enterprise tiers with retention commitments. Its publisher-facing conduct is another matter, and it is the most contested of any major AI search product.

On 4 August 2025, Cloudflare published research alleging that Perplexity used undeclared "stealth" crawlers with spoofed generic browser user-agents and rotating networks to fetch content from sites that had explicitly blocked PerplexityBot in robots.txt and by firewall rule. Cloudflare removed Perplexity from its verified-bot programme and shipped managed rules to block the traffic. Perplexity disputed the characterisation, arguing the fetches were user-initiated rather than crawling. The distinction to hold onto is that this came with technical evidence from a credible infrastructure provider, not simply publisher grievance.

Separately, and not as an allegation at all: Perplexity's own documentation concedes that Perplexity-User ignores robots.txt. That is a stated policy.

The legal record is long. Forbes accused Perplexity of plagiarising its reporting in June 2024, with Wired publishing parallel findings on scraping and fabricated summaries. News Corp, through Dow Jones and the New York Post, sued for copyright infringement in 2024, including allegations that Perplexity hallucinated quotes and attributed them to those publications. The BBC demanded it stop scraping in June 2025. Japanese publishers Yomiuri Shimbun, Asahi Shimbun and Nikkei sued in August 2025. Reddit sued in October 2025.

Comet, agents, and the Amazon case

Perplexity ships Comet, a Chromium-based AI browser launched in July 2025 and made free in October 2025. Its agentic assistant can act inside pages on the user's behalf — including, on shopping sites, completing purchases.

That capability produced the most consequential litigation in AI search to date. In November 2025 Amazon sued Perplexity in the Northern District of California, pleading the Computer Fraud and Abuse Act, over Comet agents accessing customer accounts in a way Amazon characterised as covert. On 10 March 2026 a federal judge granted Amazon a preliminary injunction blocking the shopping agent. On 4 August 2026 the Ninth Circuit overturned that injunction, finding the evidence did not support a CFAA violation and that the injunction "would not serve the public interest."

The underlying suit continues and Amazon may seek rehearing or Supreme Court review, so anyone citing this should check its current state. As of 19 August 2026, this is the leading US appellate authority on whether an AI agent may act on a website on a user's behalf, and it is two weeks old. Reporting of the injunction as a settled defeat for Perplexity is now out of date. Coverage of the Ninth Circuit ruling sets out the reasoning.

What it is good at, and who should avoid it

Three things Perplexity does genuinely well. Its citation-grounded answer format remains the best-executed in the category for a research question where you want a synthesis and the ability to check it. Its freshness is a real advantage on news-adjacent queries, a direct consequence of an indexing pipeline tuned for recency. And it is model-neutral: retrieval is always Perplexity's, but paid users choose which vendor's model generates the answer, from OpenAI, Anthropic and Google as well as Perplexity's own Sonar family. Not being married to one lab's model is a structural advantage over ChatGPT Search and Microsoft Copilot.

It is a poor choice for several kinds of user:

Anyone who wants to browse a ranked list. Perplexity is an answer engine, not a search engine, and there is no non-AI mode. If you want to scan sources, compare them yourself and do exploratory reading, it actively gets in the way.
Anyone doing work where a plausible-but-wrong synthesis is expensive and who will not click through to verify. Citations create an appearance of verifiability that the generation does not always earn.
Publishers and site owners who object on principle to documented robots.txt non-compliance.
Anyone in a regulated setting who needs a documented, auditable provenance chain for how a source was retrieved.

One more risk belongs on the list. A company that launched advertising and killed it fifteen months later, while carrying copyright suits on three continents and a live CFAA case, is still working out how it pays for itself. That is not a reason to avoid the product, but it is a reason not to build a workflow that assumes it will look the same in two years.

Common misconceptions

"Perplexity is a Bing wrapper." It was, in substance, early on. It is not now. It runs a documented crawler and an index it puts at over 200 billion URLs, and it kept working after Microsoft decommissioned the Bing Search APIs on 11 August 2025, which a reseller could not have done.

"Perplexity built its own index from day one." Also false. The company has never published a clean date for the transition, and both extreme versions of the story are wrong.

"Citations mean it is accurate." A citation means a document was retrieved. It does not mean the sentence it is attached to reflects that document. Forbes and Wired documented cases in June 2024 where it did not. Open the links.

"It respects robots.txt." PerplexityBot does. Perplexity-User does not, by Perplexity's own documentation, and Cloudflare alleged additional undeclared crawlers in August 2025.

"It trains its models on the pages it crawls." Perplexity states that PerplexityBot content is not used for training. That is a company claim rather than an audited fact, but no evidence specifically contradicts it.

"Amazon won and the shopping agent is blocked." True between 10 March and 4 August 2026. The Ninth Circuit vacated the injunction and the case is still live.

Frequently asked questions

Does Perplexity have its own search index?

Yes, as of 2025. Perplexity operates PerplexityBot, a documented crawler with a published user-agent and IP list, feeding an index the company describes as exabyte-scale and covering over 200 billion unique URLs, refreshed at tens of thousands of indexing operations per second. It has never stated that it uses no third-party feed at all, and structured verticals such as weather or finance are almost certainly served by partner APIs, but general web retrieval runs on its own infrastructure.

Does Perplexity use Bing?

Not any more. In its early years, roughly 2022 to 2024, Perplexity was widely reported to be serving retrieval through the Bing Search API, with its language model providing the differentiation. Microsoft retired the Bing Search APIs completely on 11 August 2025, decommissioning existing instances. Perplexity's own index was in place by then and the product continued without disruption, which is fairly strong evidence that it was no longer dependent on Bing.

Is Perplexity a search engine?

Not in the conventional sense. It is an answer engine: it returns written prose with numbered citations rather than a ranked list of results, and there is no mode that turns the AI off and gives you ten blue links. That is a deliberate product decision rather than a limitation. If your task involves scanning many sources, comparing them yourself, or exploratory browsing, a conventional results page will serve you better.

Does Perplexity respect robots.txt?

Partly, and the split is documented by Perplexity itself. PerplexityBot, the indexing crawler, does follow robots.txt. Perplexity-User, which fetches a page live in response to a specific question, generally does not, on the argument that a user-initiated request is not a crawl. Publishers reject that reasoning. Separately, Cloudflare published research on 4 August 2025 alleging undeclared stealth crawlers that bypassed blocks, and de-listed Perplexity as a verified bot.

Who owns Perplexity?

Perplexity AI, Inc., an independent, privately held American company. It has never been acquired. It was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho and Andy Konwinski, with Srinivas as chief executive. Its investors include Nvidia, Jeff Bezos, Databricks, Nat Friedman, Tobias Lutke and 1789 Capital. Private funding rounds have marked its valuation from roughly $1 billion in April 2024 to about $21.2 billion in early 2026.

How much does Perplexity cost?

There is a free tier with capped access to the strongest models and features, and paid consumer tiers that unlock model selection, higher Deep Research volume, document search and API credit. Developer APIs are billed by metered usage. Specific prices are not listed here on purpose: Perplexity has restructured its tiers repeatedly, and since February 2026 it has been subscription-first, having discontinued the advertising strategy it launched in late 2024.

Is Perplexity being sued?

Yes, on several fronts. News Corp, through Dow Jones and the New York Post, sued for copyright infringement in 2024. Japanese publishers Yomiuri Shimbun, Asahi Shimbun and Nikkei sued in August 2025, and Reddit sued in October 2025 over scraping. Amazon sued in November 2025 under the Computer Fraud and Abuse Act over its shopping agent; an injunction granted in March 2026 was overturned by the Ninth Circuit on 4 August 2026, with the case continuing.

Is Perplexity accurate?

Its citation format makes verification possible, which is more than most AI answers offer, but it does not make the answer correct. Forbes and Wired both published findings in June 2024 alleging summaries of articles it had not properly accessed, and the News Corp complaint alleges fabricated quotes attributed to its publications. The practical rule is that citations show a document was retrieved, not that the sentence attached to it is faithful to that document.