How Does Perplexity AI Work? (2026)

How Perplexity answers a question, in five steps:

1. It reads your question. Perplexity's help center calls this step "Understanding your question."
2. It searches the web in real time. It draws on its own index, built by its crawler PerplexityBot together with unnamed third-party crawlers, and can visit a page live while it answers. Pro Search runs several searches per question.
3. In our tracking, it kept to a small pool of pages. Perplexity's help center lists no pool size. In the answers we tracked, 96.8% drew on exactly 10 sources.
4. A model writes the answer. By default, an automatic option called Best picks the model. Paid plans can choose one by hand.
5. It numbers the citations. In Perplexity's words, "Each answer includes numbered citations linking to the original sources."

These numbers come from LoudFace's tracking in Peec of 146 B2B SaaS agency-buying questions, 9 to 22 October 2026 (method below). ChatGPT figures for sources and citations cover the 1,624 answers where it searched. Google AI Overviews' sources list is mostly the links it shows, so nearly all of it counts as cited. The day-to-day overlap rests on 1,873 Perplexity day pairs, 455 for ChatGPT and 1,746 for AI Overviews.

Where does Perplexity get its sources?

From three places: its search index, live visits to pages, and licensed data.

The index. Perplexity's crawler, PerplexityBot, "is designed to surface and link websites in search results on Perplexity," according to its developer docs on bots. It does not crawl alone. The help center says "Perplexity partners with third-party crawlers to help build our search index," and names no partner. You will read on some sites that Perplexity runs on Google and Bing. No Perplexity page we read says so.

Live visits. A second agent, Perplexity-User, fetches pages while someone is asking. The docs describe it plainly: "When users ask Perplexity a question, it might visit a web page to help provide an accurate answer and include a link to the page in its response." In a 4 August 2025 blog post replying to Cloudflare, Perplexity wrote that its user-driven agents "do not store the information or train with it."

Licensed data. Perplexity's premium data sources add "licensed content from publishers and data providers that is not available on the open web." Wiley, PitchBook Essentials, CB Insights, Statista and Midpage are built in for all users. FactSet, Crunchbase and other licensed connectors work for Pro, Max and Enterprise users who hold an eligible license with the data partner.

Perplexity also says your pages will not be used for pre-training: "Perplexity does not build foundation models, so your content will not be used for AI model pre-training." That is a real difference from ChatGPT, which we covered in where ChatGPT gets its information.

Does Perplexity follow robots.txt?

PerplexityBot does. Perplexity-User generally does not. Both statements come from Perplexity, on two different pages, and a site owner who reads only one of them gets the wrong picture.

The help center's robots.txt article says: "Our crawler, PerplexityBot, will not index the full or partial text content of any site that disallows it via robots.txt." It adds that the feature that let users summarize a blocked page "has been disabled to prevent misuse." Blocking still leaves a trace: "we may still index the domain, headline, and a brief factual summary."

The developer docs describe the other agent. "Since a user requested the fetch, this fetcher generally ignores robots.txt rules." So robots.txt controls PerplexityBot. If you want to stop Perplexity-User, you need a rule at your firewall that matches its user agent and Perplexity's published IP ranges.

Most B2B SaaS companies want the opposite. Perplexity's published advice on appearing in its answers is about access. It recommends allowing PerplexityBot and requests from its published IP ranges, and where a firewall sits in front of your site, "you may need to explicitly whitelist Perplexity" bots. Neither change is instant. The docs say robots.txt settings can take "up to 24 hours for our systems to reflect changes," and that "Changes to WAF configurations may take some time to propagate." Before you spend a cent on content, check your security settings. A blocked crawler makes the rest of this work pointless.

How does Perplexity choose which sources to cite?

Perplexity doesn't publish how it ranks or picks sources. Its help pages say it searches the web and identifies trusted sources, and stop there. Anyone selling you a list of Perplexity ranking factors is selling a guess.

What Perplexity does publish is a set of source labels. "Perplexity marks some cited domains with a Government, Academic, or Trusted label." The labels judge whole sites ("Perplexity rates the whole website a source comes from, not each individual page"), and the source labels page says "partnerships, payments, and other business arrangements do not affect" them. It also says "Most domains on the web do not have a label." Nothing on that page ties a label to which sources get picked. Don't make "get a Trusted label" a goal.

Perplexity has changed how it cites before. In July 2024, launching its Publishers Program with TIME, Der Spiegel, Fortune, Entrepreneur, The Texas Tribune and WordPress.com, it wrote: "Recently, we updated how our systems index and cite sources."

What we can measure is the funnel. Across the 2,044 Perplexity answers in our tracking, the engine read 20,959 pages and gave 6,294 of them a numbered citation: 30.0%. In 93.4% of those answers it cited between 1 and 5 pages. Only 21 answers (1.0%) cited none. On these questions the pool was the first filter, and it was tight. ChatGPT, when it searched on the same questions, listed about twice as many pages (a median of 21 against 10) and cited a far smaller share of them (11.2% against 30.0%).

Nine in ten of Perplexity's 6,294 citations on these questions went to company and agency websites:

• Company and agency websites: 5,684 (90.3%)
• Publishers and editorial sites: 176 (2.8%)
• LinkedIn: 155 (2.5%)
• Unclassified sites: 120 (1.9%)
• Other reference sites: 107 (1.7%)
• Other user-generated sites: 44 (0.7%)
• Review sites: 3
• Institutional sites, as Peec classes them: 3
• Reddit and forums: 2
• YouTube and Wikipedia: 0

Reddit was retrieved 39 times and cited twice. On the agency-buying questions we tracked, Perplexity read and cited company websites, including agencies' own lists and service pages. A caveat on order: Peec stores cited sources at the top of each list, so the list order tells you nothing about Perplexity's own ranking. Being cited is not the same as being ranked first.

How is Perplexity different from ChatGPT and Google AI Overviews?

In our tracking, Perplexity searched every time, read a small pool and cited about a third of it. ChatGPT searched in 79.5% of answers and, when it did, read a median of 21 pages and cited 2. Google AI Overviews read about 8 and cited nearly all of them.

The control points differ too. ChatGPT search inclusion runs through OAI-SearchBot, with GPTBot as a separate training crawler. Perplexity documents no training crawler at all. For AI Overviews, Google's guidance says a page "must be indexed and eligible to be shown in Google Search with a snippet." There is no separate AI crawler to allow. Our guide to Google AI Overviews optimization covers that side.

The sources barely overlap. On 974 occasions when Perplexity and ChatGPT both cited something for the same question on the same day, only 15.4% of Perplexity's cited domains were also cited by ChatGPT, and 60.7% of those pairs shared no domain at all. The shared figure for Perplexity and AI Overviews (1,934 pairs) was 15.5%. On these questions, Google also leaned on sources Perplexity barely used: YouTube made up 6.9% of AI Overviews citations (0 on Perplexity), Reddit and forums 3.3% (2 citations on Perplexity), and LinkedIn 6.6% (2.5% on Perplexity). We saw the same split in the agencies each engine names in our study of whether AI engines recommend the same agencies.

On the 146 questions we tracked from 9 to 22 October 2026, Perplexity was also the steadiest engine. When the same question was asked on consecutive days, its cited domains scored an average overlap of 0.527, on a scale where 0 means no shared domain and 1 means the same set, against 0.228 for ChatGPT and 0.262 for AI Overviews. That matches our earlier study of whether ChatGPT gives the same answer to everyone, where Perplexity was the most consistent engine on all four day-to-day measures.

Peec almost never records the search queries Perplexity runs (1 of its 2,044 answers carried any), so we can't tell you how Perplexity rewrites a question. For what query rewriting is and why it matters, see our explainer on fan-out queries.

How does a B2B SaaS get named in Perplexity?

On the B2B SaaS agency-buying questions in our sample, Perplexity read about 10 pages, cited about 3, named about 2 companies and changed its cited domains less from one day to the next than ChatGPT or AI Overviews did.

Get one of your pages into the 10. We treat that pool as the gate: if none of the roughly 10 pages Perplexity reads for a question mentions you, we don't expect it to name you. One audit we ran of 120 AI answers to agency questions, 40 per engine, points to the same gate: every answer that named LoudFace had a loudface.co page in its sources. With 90.3% of Perplexity's citations going to company and agency websites on these questions, your own pages and the third-party lists that name you are where that happens. YouTube and Reddit barely register on Perplexity for this kind of question. Google AI Overviews cites them far more often, as the comparison above shows.

Then be one of the 3 it cites. On these questions, Perplexity's median answer ran to 54 words, which left little room for your nuance. Our view at LoudFace is that the page that gets quoted opens with the answer itself. Put your short answer, with its number and its source, in the first lines of the page.

Treat each engine as its own channel. When Perplexity and ChatGPT both cited sources for the same question on the same day, only 15.4% of Perplexity's cited domains were also cited by ChatGPT. Winning ChatGPT will not carry you into Perplexity on its own. Of the three engines we tracked, Perplexity's cited domains also changed the least from one day to the next. Its day-to-day overlap score (domains cited on both days, divided by all domains cited on either day) averaged 0.527, against 0.228 for ChatGPT and 0.262 for AI Overviews. When an engine's answers to the same question cited sources on two consecutive days, a domain cited on the first day, yours or a rival's, was cited again on the second in 62.8% of cases on Perplexity, against 33.3% on ChatGPT and 35.1% on AI Overviews. On a question where no one has locked in the citations yet, though, a well-structured page from a site with some existing authority can be cited by Perplexity within 24 hours of publishing. On a question a competitor already owns, expect a slower climb.

Leave the myths alone. You don't need a Trusted label or a list of secret ranking factors. Allow PerplexityBot, check that your firewall lets it through, and put a sourced answer at the top of the pages that compare you with competitors.

Which AI models does Perplexity use in 2026?

Several, and the line-up moves fast. Perplexity's model page calls its list "a snapshot as of October 17, 2026." For Search on the Pro and Max plans it lists GPT-6.1 Sol, Gemini 3.8 Flash, Claude Sonnet 5.5, Kimi K3, GLM 5.3, Grok 4.7 and Nemotron 3.5 Super VL. Max subscribers also get GPT-6 Astra, Gemini 4 Argon, Claude Opus 5.5 and Claude Fable 5.1.

The page itself warns: "The model selector in Perplexity is always the source of truth for your account." Take that seriously. Perplexity's own "How does Perplexity work?" article, the page Google ranked first for this question when we checked in October 2026, still names GPT-5 and Claude 4.6 Sonnet. The help pages lag the product.

• Best is the default. "It is the automatic model option inside Search." On the Free plan "you cannot choose an advanced Search model manually." And the same model behaves differently inside Perplexity, because "Perplexity integrates these models with its own search, citations, tools, prompts, safety systems, and usage limits."
• Sonar is Perplexity's own model. At its relaunch in February 2025, Perplexity said "Built on top of Llama 3.3 70B, Sonar has been further trained to enhance answer factuality and readability." Sonar does not appear in the October 2026 Search table, and the current default-model page names no model, saying only that the default "is optimized for quick searches and web browsing."

What can Perplexity do beyond a quick answer?

The plain search is one of several modes now. Per Perplexity's help center:

• Search gives "Fast answers to everyday questions." Pro Search, inside it, "conducts multiple searches across the web, drawing from articles, academic papers, forums, videos, and more."
• Research (Deep Research) "performs dozens of searches, reads hundreds of sources, and reasons through the material to autonomously deliver a comprehensive report." Perplexity gives the timing three ways. The Research page says it "completes most research tasks in under 3 minutes" and, in the same sentence, that it "can take around 4 to 5 minutes to generate a response." The how-it-works article says "in just 2-4 minutes."
• Model Council lets you "Consult several models at once for research, complex decisions, or verification," on Max and Enterprise Max.
• Comet is Perplexity's browser, which runs tasks on web pages.
• Perplexity Computer, launched on 25 February 2026, handles long-running work. Perplexity's split: "Perplexity answers your questions. Computer does your work." It runs on credits, and "Today, 100 credits equals $1."

For a B2B buyer, the modes change how many sources an answer can draw on. Pro Search runs multiple searches, and Perplexity says its paid Pro plan gives "10x as many citations per answer." Our tracking cannot tell which plan or mode answered, so treat the 10-page pool we measured as a typical capture rather than a ceiling.

Why is Perplexity controversial?

Mostly over how it treats publishers' content. Two fights stand out: a crawling dispute and a run of lawsuits.

The crawling dispute. On 4 August 2025, Cloudflare wrote that it was "observing stealth crawling behavior from Perplexity" and said that when Perplexity's declared crawlers were blocked, it saw "a generic browser intended to impersonate Google Chrome on macOS" across "tens of thousands of domains and millions of requests per day," and that "we have de-listed them as a verified bot." Perplexity replied the same day: "It appears Cloudflare confused Perplexity with 3-6M daily requests of unrelated traffic from BrowserBase," and argued that agents acting for a user are not crawlers. The two accounts can't both be right, and we found no independent audit that settles it.

The lawsuits. Several publishers have sued Perplexity, and so has Amazon. These are three of the suits, and none had reached a verdict on the records we could read:

• Dow Jones & Company and NYP Holdings v. Perplexity AI, filed in October 2024 in federal court in New York. On 21 August 2025 the court denied Perplexity's motion to dismiss "in full." The scheduling order says "This case is to be tried to a jury." The docket shows no verdict.
• The New York Times v. Perplexity: the Times says "On December 5, 2025, we filed a lawsuit against Perplexity AI, Inc." alleging copyright infringement, trademark dilution and trademark infringement. Perplexity filed a partial motion to dismiss, then a renewed one in April 2026 after an amended complaint. In its filing for the quarter ended 30 June 2026, the Times reports no ruling on that motion and says it intends to keep pursuing the case.
• Amazon.com Services v. Perplexity AI, filed on 4 November 2025 in federal court. Trade press reported that the suit centers on Perplexity's Comet browser and its AI shopping agent. On 9 March 2026 the court granted Amazon a preliminary injunction, and Perplexity appealed on 10 March 2026. On 4 August 2026 the Ninth Circuit vacated the injunction and sent the case back to the district court, where Amazon filed an amended complaint on 21 September 2026. The docket shows no verdict.

Perplexity's answer to publishers has been money. The Publishers Program, launched in July 2024, promised revenue sharing, and Comet Plus, announced on 25 August 2025 as "a $5 standalone subscription," splits revenue "based on three types of internet traffic: human visits, search citations, and agent actions."

How did we measure this?

Every number marked as ours comes from LoudFace's own Peec project, first read on 23 October 2026 and pulled again on 24 October 2026, with identical results. Peec asks each tracked question once a day per engine from a US location. We dropped the 4 questions in the project that name LoudFace, leaving 146 B2B SaaS agency-buying questions. Over 14 days, 9 to 22 October 2026, that gave 2,044 answers per engine, with no failed pulls.

A retrieved (read) page is an entry in the answer's sources list. A cited page is one the answer attaches a numbered citation to. A named company is a tracked brand that appears in the answer text.

Peec does not show which Perplexity plan, mode or model answered, or whether the session was logged in, so the 10-page pool may reflect a default or free-tier view. The questions are agency-buying questions, so the source mix describes that question type, not Perplexity in general. Other topics may draw on very different sources.

In our tracking, Perplexity answered from a pool of about 10 pages, so the work that counts is getting your page into that pool.

LoudFace's free AI visibility audit checks where your brand stands across ChatGPT, Claude, Gemini, and Perplexity searches versus your top competitors. For the work itself, our generative engine optimization service is built to get your brand cited and recommended in AI answers.