Perplexity SEO: how pages get cited, sentence by sentence

Perplexity searches its own index, one passage at a time

Perplexity began on third-party search APIs and replaced them, its September 2025 architecture write-up explains, because they were costly, stale or slow, and returned whole documents when a model needs the one paragraph that answers. Its own index now tracks “over 200 billion unique URLs” and serves Perplexity’s apps, Comet and its developer APIs alike. Each stage below comes from that document, with what you control at each one.

1. 1

   A model decides when each URL is recrawled

   An ML model predicts whether a URL needs indexing and when, weighing how important it is and how often it’s likely to change. Documents from authoritative domains and under-covered topics get priority for fast storage.

   Your leverUpdate pages on a real cadence and cover the questions competitors skip: importance and change frequency are what the scheduler weighs.

2. 2

   Pages are parsed into self-contained spans

   A content-understanding module extracts the meaningful text with rules adapted per site; list- and table-heavy sites get more formulaic parsing. Each document is then split into spans that are retrieved and ranked on their own at query time.

   Your leverMake every section make sense in isolation, and put specs, prices and comparisons in real HTML tables.

3. 3

   Keyword and semantic retrieval run together

   The index is queried lexically and by embeddings in parallel, and the two result sets merge into one broad candidate pool.

   Your leverUse the exact terms buyers type, such as product names, model numbers and units, inside natural explanation.

4. 4

   Stale and off-topic candidates are dropped

   Prefilters remove clearly non-responsive or stale content before any expensive scoring. Every result Perplexity’s API returns carries a date and a last_updated field, and developers can filter on both.

   Your leverShow a visible published and updated date, mirror them in datePublished and dateModified, and change them only when the content changes.

5. 5

   Cross-encoders make the final cut

   Fast lexical and embedding scorers narrow the pool, then cross-encoder rerankers score what’s left at both document and sub-document level. The rankers keep training on signal from the millions of requests Perplexity serves every hour.

   Your leverPut the answer in the first sentence of the section that owns the question.

6. 6

   The answer pins each claim to its sources

   The model writes from the passages that survive and attaches sources claim by claim. A standard answer cites a handful; Research mode “performs dozens of searches” and reads hundreds of sources before it writes.

How an answer sentence maps to its sources

Perplexity’s help center says every answer includes numbered citations linking to the original sources. In the web app, as of October 2026, that numbering is folded into a compact chip at the end of a claim. Four pieces make up the citation UI:

• The claim chip. After a sentence or clause: the first source’s site name, plus +N when more sources back the same claim.
• The source card. Hover or tap the chip and a card pages through that claim’s sources (1/4, 2/4), each with its title, the snippet Perplexity read and, for rated domains, a shield label.
• The sources panel. The full list for the answer, opened from the favicon row under it (“10 sources”) or the panel beside it.
• The Links tab. The same sources as a plain list of results, next to Answer and Images.

Underneath, the numbering is literal. In the Agent API, which took over when support for Sonar Chat Completions ended on September 27, 2026, the answer text carries markers such as [1][3], and each number is the id of a search result with its own title, url, snippet, date and last_updated. Abridged, with illustrative values:

"output": [
  {
    "type": "search_results",
    "results": [
      { "id": 1, "title": "Pricing: free for up to five users",
        "url": "https://www.example.com/pricing",
        "date": "2026-03-02", "last_updated": "2026-10-01",
        "snippet": "The free plan covers up to five users…", "source": "web" }
      // …
    ]
  },
  {
    "type": "message",
    "content": [
      { "type": "output_text", "text": "Its free plan covers teams of up to five[1][3]." }
    ]
  }
]

The shield: Government, Academic and Trusted labels

Some citations now carry a small shield icon. Perplexity’s help center explains that a source review process rates whole websites, not individual pages, and gives the ones that clearly pass one of three labels. Hover the citation and the card shows the label with one line on what the site is known for.

The review asks objective questions: whether a site corrects its mistakes, whether it says who wrote each piece, and whether it keeps news apart from advertising and opinion. Perplexity says partnerships and payments don’t affect labels, that most domains have none, and that having none is not a negative judgment.

Perplexity doesn’t say labels change ranking, so treat them as a trust cue for readers rather than a ranking factor. The checks are cheap to pass anyway: a public corrections policy with dated correction notes, a byline and author page on every article, and sponsored content clearly marked and kept apart.

What robots.txt controls on Perplexity, and what it can't

Perplexity documents two user agents, each with its own published IP list. Each setting works independently, and changes can take up to 24 hours to reach Perplexity’s systems.

• Builds the index

  PerplexityBot

  Allow

  Surfaces and links websites in Perplexity’s search results. Perplexity says it isn’t used to crawl content for AI foundation models, follows robots.txt rate limits and backs off when a site is struggling. IPs: perplexitybot.json.

  Obeys robots.txt

• Fetches for a user

  Perplexity-User

  Allow

  Visits a page when a user’s question needs it, for example to check a detail live. Not used for crawling or for training. IPs: perplexity-user.json.

  Generally ignores robots.txt

  Perplexity: "this fetcher generally ignores robots.txt rules."

# Perplexity: search index and live fetches
User-agent: PerplexityBot
Allow: /
Disallow: /cart/
Disallow: /account/

User-agent: Perplexity-User
Allow: /

Sitemap: https://www.example.com/sitemap.xml

Outside crawlers feed the index too. The help center says Perplexity partners with third-party crawlers and has updated those agreements so they also respect robots.txt, particularly on news publishers’ sites.

Make every passage fetchable

Perplexity has no webmaster console, takes no URL submissions and isn’t an IndexNow participant. Whether a passage gets to compete comes down to the crawler receiving clean HTML, quickly, with the text already in it.

• Text in the server HTML

  Required

  Vercel’s December 2024 crawler study found that none of the major AI crawlers, PerplexityBot included, render JavaScript. Glenn Gabe’s August 2025 test found Perplexity failed to find the content on every client-rendered URL he tried.

• Both IP lists allowed

  Required

  As of October 2026, perplexitybot.json lists 8 addresses and perplexity-user.json lists 4. Let them through bot-fight modes, managed AI rules and rate limits, and refresh from the JSON instead of hard-coding.

• Fast, clean responses

  Helps

  Perplexity-User fetches while someone waits. A timeout, 403, 429 or challenge page means the answer gets written from another site’s passage.

• Honest dates

  Helps

  Visible published and updated dates, plus datePublished and dateModified. The API exposes both dates per result, and stale content is filtered before ranking.

• HTML tables and lists

  Helps

  List- and table-heavy content gets more formulaic parsing, so the facts survive extraction intact.

• Canonicals on syndicated copies

  Helps

  In Columbia’s Tow Center test (March 2025), Perplexity Pro cited syndicated copies of Texas Tribune articles for three of ten queries. Make partners canonicalize to your original.

• XML sitemap

  Unconfirmed

  Perplexity doesn’t document sitemap use. Keep one with honest lastmod values, referenced from robots.txt.

• llms.txt

  Unconfirmed

  Perplexity hasn’t said PerplexityBot reads other sites’ llms.txt files or ranks with them.

• IndexNow

  No effect

  Perplexity isn’t on IndexNow’s list of participating search engines.

# 1. Is the text in the raw HTML PerplexityBot receives?
curl -s -A "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot)" \
  https://www.example.com/pricing | grep -c "free for up to five users"
# 0 means client-rendered text, or a WAF challenge

# 2. The current IP allowlists (the files redirect, hence -L)
curl -sL https://www.perplexity.com/perplexitybot.json | jq -r '.prefixes[].ipv4Prefix'
curl -sL https://www.perplexity.com/perplexity-user.json | jq -r '.prefixes[].ipv4Prefix'

Write passages Perplexity can lift

Perplexity’s citations overlap with Google’s results more than any other assistant’s, so classic SEO carries over. What’s specific to Perplexity is the unit: it scores spans, filters for freshness and pins claims to sources. Write for that.

The same facts, before and after

Comet, Computer and agents in the user's browser

Perplexity runs agents in four places, its September 2026 security post says: Computer in the cloud, Comet in the browser, Portable Computer on the user’s own device, and coding agents on its engineers’ laptops. The first two can land on your site, alongside the answer engine’s own live fetches.

So robots.txt governs Perplexity’s index and its live fetcher, not the person browsing with an AI assistant. Perplexity says the November 2025 Comet Assistant got new tools for perceiving and interacting with complex websites. What helps it finish a task is mostly what helps people: real <button> and <label> elements, prices and stock in the HTML, no CAPTCHA in front of public information, and nothing that only works on hover.

Product cards and checkout inside the answer

Shopping questions get product cards built around what matters for the question asked, with the relevant specs and reviews. Since November 25, 2025 that experience is free for all US users, and checkout runs through PayPal without leaving the conversation (Perplexity). The retailer stays merchant of record: it sees the customer, handles returns and keeps the relationship.

Perplexity said in 2024 that its product cards aren’t sponsored. Products reach them through:

• Your own product pages, crawled like any other page, with specs, price and availability in the HTML.
• Shopify, whose product data powered Perplexity’s shopping discovery from its November 2024 launch for stores that sell and ship to the US.
• The Merchant Program, free for large retailers that share product specs directly. Perplexity says fuller product data raises the chance of being a recommended product; members also get a dashboard of search and shopping trends (Perplexity).
• PayPal, which lets shoppers buy from participating retailers inside the answer.

Getting paid: the Publishers' Program and Comet Plus

Perplexity has paid publishers two ways. The Publishers’ Program, launched in July 2024 with TIME, Der Spiegel, Fortune and others and joined by Gannett in July 2025, shared revenue from sponsored follow-up questions. That ad business is going: in February 2026 Perplexity executives told the Financial Times it would wind advertising down by the end of 2026 over trust concerns (Campaign).

Comet Plus is the newer model: a $5-a-month subscription, included in Pro and Max, whose revenue goes to participating publishers minus a small share for Perplexity’s compute. It pays on three kinds of traffic, so a citation can earn money even when nobody clicks. Launch partners announced in October 2025 include Condé Nast, Fortune, Le Figaro, Le Monde, the Los Angeles Times and The Washington Post. Publishers apply by emailing publishers@perplexity.ai.

For everyone else the lesson is simpler. The ad product Perplexity built was sponsored follow-up questions, not paid spots in the answer, and it’s being wound down. The organic path in this guide is the way in.

Reading Perplexity in your analytics and logs

As of October 2026, citations in Perplexity’s web app link straight to the cited URL, with no redirect and typically no UTM parameters, and its pages send Referrer-Policy: strict-origin. Your analytics receives the bare origin, https://www.perplexity.ai/, never the thread or the question. Two consequences: you can’t see which prompt sent a visit, and a page still served over http:// gets no referrer at all, so the visit counts as Direct.

# Admin → Data display → Channel groups → copy the default group
# New channel "Perplexity", placed above Referral
# Condition: Session source matches regex
^(www\.)?perplexity\.ai$

# Or one channel for AI answers, Perplexity included
(^|\.)(perplexity\.ai|chatgpt\.com|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai)$

In logs, Perplexity-User requests are the best evidence that a page is being pulled into live answers, and PerplexityBot shows what’s being indexed. User agents are trivial to fake, so match the source IP against the published lists before trusting either.

# Hits per Perplexity agent
grep -oE "PerplexityBot|Perplexity-User" access.log | sort | uniq -c

# Pages fetched live for users, most requested first
grep "Perplexity-User" access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

What Perplexity doesn't offer

• Myth: Perplexity runs on Google's results.

  Reality: It runs its own crawler and an index of 200B+ URLs. Google rankings correlate with its citations (28.6% top-10 overlap) but aren’t what it searches.

• Myth: You can pay for a place in the answer.

  Reality: Its ad product was sponsored follow-up questions, not spots in the answer, and it’s being wound down. Source labels can’t be bought either.

• Myth: There's a console for submitting URLs.

  Reality: There’s no webmaster console, no URL submission and no IndexNow. Discovery runs on its crawler, partner crawlers and links.

• Myth: Blocking PerplexityBot removes you completely.

  Reality: Your domain, headline and a short summary may still appear, Perplexity-User generally ignores robots.txt, and syndicated copies can be cited instead.

• Myth: You need a shield label to be cited.

  Reality: Most domains have no label, Perplexity says no label isn’t a negative judgment, and it hasn’t said labels affect ranking.

16 moves, highest impact first