Perplexity SEO: what the engine documents and what it does not

In Ahrefs' US database the query perplexity seo carries 300 searches a month, a keyword difficulty of 4 and a cost per click near 7 USD, measured on 5 September 2026. Difficulty 4 beside a seven dollar click is an odd pair, and the results page explains it: vendors selling the service, each with its own list of ranking factors, none of them quoting the two pages where Perplexity says what its crawler does.

How do you optimize for Perplexity?

Perplexity SEO is allowing PerplexityBot through robots.txt and the firewall, being indexed, and writing passages that carry a claim without their page around them. A new site rarely gets attached to an answer, a mid-authority site wins narrow questions, and a consolidated one competes for the handful of slots an answer actually attaches.

Quick answer

• Two crawlers, two jobs. PerplexityBot exists so pages can be surfaced and linked in answers. Perplexity-User fetches a page because a person asked for it, and the documentation says that fetcher generally ignores robots.txt rules.
• Blocking is not an off switch. Perplexity's help centre states that a blocked page may still contribute its domain, its headline and a brief factual summary.
• robots.txt is one door of two. The same documentation asks for the published IP ranges to be permitted, and gives firewall rules for doing it, which is the step none of the four guides I read mentions.
• Nothing is published about selection. Not one Perplexity page I opened says how a source gets chosen, ranked, or attached to a particular sentence.

What kind of problem this is

The belief that sends this work wrong is not the usual one.

With ChatGPT, the model a buyer arrives with is that it is a search engine wearing a different interface. Here it is more specific and harder to shake: that somewhere there is a published list of Perplexity ranking factors, because every page that ranks for the query publishes one. Nine signals. Six pipeline stages. Three reranking layers. The lists are confident, they are numbered, and they agree with each other closely enough to look like reporting.

They are not reporting. They are inference from watching answers, which is a legitimate activity and a different thing, and almost none of them says so. The sequence that does have a documented basis is short and unglamorous: crawler files first, passages second, and very little in between. It is worked through for the other engine, where the crawler names are published in the same way and the selection rule is missing in the same way, in how to get cited by ChatGPT.

The discipline those prerequisites belong to is the one with an actual academic paper behind it, and the paper is more specific than the sales language built on top of it about which techniques moved anything and which made things worse. That reading, with the numbers, is in what generative engine optimization is.

Drop the list and the work changes shape. Two things replace it and both can be verified in an afternoon: opening the doors the company documents, and opening the engine to see whether your name is inside it.

The two doors that are documented, and the one that is not

Perplexity publishes a crawlers page, and it is short enough to read in three minutes.

The first door is robots.txt. The page names PerplexityBot, says it is "designed to surface and link websites in search results on Perplexity", adds that "it is not used to crawl content for AI foundation models", and recommends allowing it. It publishes the full user agent string, Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot), which makes the first check a one line request against your own site. Mine returned 200 with that user agent, through Cloudflare, on 5 September 2026.

The second door is the firewall, and none of the four guides I read mentions it. The same page does not stop at robots.txt: it asks for requests from the published IP ranges to be permitted, lists the endpoint that carries them, and then walks through Cloudflare and AWS WAF rules combining a user agent condition with an IP condition. A site can allow the crawler in a text file and refuse it at the edge, and the text file is the half everyone checks. A plain request from your own machine only clears the user agent rule, since the IP it arrives from is not one a WAF condition is looking for.

The third door cannot be opened, because it has no name. The help centre article on robots.txt, last updated on 16 July 2026, says that Perplexity "partners with third-party crawlers to help build our search index" and that the agreements with those providers were updated so they also respect robots.txt. Which providers is not stated anywhere I could find. A crawler you cannot name is a crawler you cannot allowlist by user agent, cannot allowlist by IP, and cannot see in a log without guessing.

Underneath all three sits the ordinary gate. Answers are composed from material a crawler already stored, so a page missing from search indexes is not a candidate for anything, and the order for diagnosing that is in why a page is not indexed.

One file that gets recommended for this engine specifically has no documented role in it. Perplexity serves an llms.txt for its own developer documentation, which says it publishes one, not that it reads yours, and the evidence for anything else is thin enough to be worth reading before spending an afternoon on it: what llms.txt is.

What ten sources and three citations look like

I opened the engine twice on 5 September 2026, signed out, and wrote down what came back.

For how to get cited by perplexity, the answer was four bullets. Each bullet carried an inline marker labelled with the source's domain name rather than a number: tryhikoo, upgrowth, authoritytech twice. Three distinct domains attached to four claims. Under the answer, a panel said ten sources. So seven of the ten retrieved documents were fetched, listed, and never attached to anything. Not one of the ten came from perplexity.ai.

For perplexity seo, the answer was two sentences with a single marker covering three sources, and the source panel again said ten. Two of the ten cards carried a small label reading Trusted domain. One was Perplexity's own hub page about using the engine for SEO research, which answers the other half of an ambiguous query; the other was a vendor. Eight of the ten were agency or service pages selling optimization for this engine.

Three things in that are worth keeping, and one is worth discarding.

Retrieval is not citation, and the gap is most of the set. Ten documents in, three domains attached. Being in the source panel and being the evidence behind a sentence are different outcomes with different value, and any tracker that counts the panel is counting the easier one.

The label is visible and the criterion is not. Two cards said Trusted domain. No Perplexity page I opened defines what makes a domain trusted, how a domain enters that set, or what the label changes. It is the clearest example on the whole surface: the outcome is displayed, the rule behind it is not.

The interface described in the guides is not the one on screen. Several describe numbered references. What I saw were chips carrying the source's domain. Small, and it is the tell of a copied description rather than an opened tab.

And a single check proves nothing, which is why the observation above is written as two answers on one date and not as a rate. Ask the same question tomorrow and the panel can hold a different set. What makes the method worth anything is repetition: a fixed list of customer questions, put to the engine monthly, recording which domains get attached rather than which appear. Signed out, the free allowance ran out on my third query of the day, so the battery needs an account to be repeatable. Nothing from that exercise reaches Search Console, which holds the rest of the evidence and has no column for this.

The guides that rank for this query, and what they cite

Four of the first results opened cleanly on 5 September 2026, and they are the reason the keyword difficulty is 4.

Three of the four publish a list of what Perplexity rewards, with nothing attached to any item: relevance, authority, clarity, recency, structured data, in one case a set of pre-determined authoritative domain lists organised by topic. The fourth is the exception and attributes most of its claims to named third-party analyses, which is a real improvement and still leaves the reader one step from a measurement. None of the four quotes the crawler page. None of them quotes the help centre sentence about a blocked page keeping its domain, its headline and a summary. None of them says when this work is not worth doing.

The figures travel worse than the factor lists. One source card in Perplexity's own panel carried a description stacking two averages for the same metric in a single sentence, 8.2 sources per answer and 6.71, each credited to a different tracker. Both cannot describe the same thing, and the page presents them as agreement. Naming a publisher is not the same as showing a sample, which is the failure mode across this whole subject rather than a fault of one vendor, and the version of it that runs through answer engine optimization is the same shape.

The two claims worth separating before spending anything are the ones the phrase itself confuses: optimizing for the engine, and using the engine as a research tool. They share a keyword and share nothing else, and that split is the one drawn in SEO for AI.

When Perplexity is not worth the work

When the business runs on proximity. Somebody looking for an emergency plumber is not asking an answer engine, and the same hours put into the profile and the reviews return more, which is the argument set out in local SEO for a small business.

When the site is not indexed anywhere. The candidate pool is drawn from search indexes, so indexing is the prior problem and no amount of passage editing jumps over it.

When there is nothing on the site anyone would quote. If every claim restates a measurement somebody else made, the work is producing one of your own, and that is a longer project than formatting.

When the goal stated out loud is traffic. The answer is read inside the engine. A citation buys recognition and occasionally a visit, and a plan that converts the first into the second is inventing the rate.

What to expect from where you start

The middle row is where the return sits, and the reason is structural rather than encouraging. A site with no index presence is not in the retrieval set at all, and a site the engine already reaches for wins nothing by being asked twice. In between, the difference between being retrieved and being attached is a writing problem, and writing problems are the only ones on this list that money and patience reliably fix.

Mistakes that repeat

• Treating a factor list as documentation. Three of the four ranking guides on page one publish one, and none of them cites a source for it.
• Allowing the crawler in robots.txt and blocking it at the edge. The documentation asks for both, and only the first tends to get checked.
• Blocking the bot and calling it removed. The help centre says the domain, the headline and a short factual summary can survive the block.
• Counting the source panel as citations. Ten documents were retrieved in both answers I read, and three domains were attached to a sentence.
• Copying a description of the interface instead of opening it. Several pages ranking for this query describe numbered references; the two answers I read carried domain name chips.
• Reading one answer as a measurement. The same question asked twice can return a different panel, which is why the method is a fixed list repeated monthly.

Data and transparency

The crawler names, their stated purposes, the full PerplexityBot user agent string, the published IP endpoints and the Cloudflare and AWS WAF configuration guidance come from Perplexity's crawlers documentation, opened on 5 September 2026, which also carries the statement that Perplexity-User "generally ignores robots.txt rules" because a user requested the fetch. That PerplexityBot will not index the full or partial text of a site that disallows it, that a blocked page may still be indexed as domain, headline and a brief factual summary, that Perplexity does not build foundation models, and that the company "partners with third-party crawlers to help build our search index" without naming them, come from the help centre article on robots.txt, written by Perplexity Support and last updated on 16 July 2026, opened the same day. That citations have been included in every answer since launch, and the first partner list of the Publishers' Program, come from Perplexity's announcement of 30 July 2024, opened the same day. No page among those three says how a source is selected, ranked or attached to a sentence, which is why every claim about selection in this article is written as observation.

The two answers described, their source counts, the domain name markers, the Trusted domain labels and the signed-out rate limit are my own reading of the engine on 5 September 2026, signed out, in Chrome, on the queries named in the text. They are two answers on one day and are written as such: a rate would need a fixed question set repeated over weeks. That my own site returns 200 to a request carrying the documented PerplexityBot user agent was tested the same day from an ordinary address, which exercises the user agent rule and not the IP allowlist.

That three of the first four results publish unattributed ranking factors and that the fourth attributes most of its claims to named third-party analyses is my own reading of that results page on 5 September 2026; the four pages read were mrs.digital, seo.com, flow-agency.com and digitalsuccess.us. The two averages of 8.2 and 6.71 sources per answer are reproduced from a source card description as it appeared in Perplexity's own panel, as an example of two figures for one metric presented as agreement, not as measurements this article adopts. No figure appears here for how often Perplexity cites any site, for its market share or for the traffic it sends: the numbers in circulation for all three vary widely between publishers and none I read arrives with a declared method. The search demand for this query, 300 monthly searches in the United States with a keyword difficulty of 4 and a cost per click near 7 USD, was measured in Ahrefs on 5 September 2026 for this blog's keyword research. The operational judgment throughout comes from work across a portfolio recording more than 300 million impressions a year in Search Console. Verified as of September 2026.

Primary sources, opened on 5 September 2026: Perplexity's crawlers documentation; the help centre article on robots.txt; the Publishers' Program announcement.

What this changes

The reason to start with this engine is not its size. It is the feedback loop.

Perplexity shows its working. The sources are on screen, the markers say which claim came from where, and a question you care about takes forty seconds to test. Everywhere else in this territory the loop runs in months and mostly in the dark. So the useful way to hold it is not as a channel with its own budget line but as the instrument you measure the others with: the place where you find out, cheaply and today, whether anything you publish is quotable at all.

That reframing has a cost attached, and it is the honest close. If the engine is your instrument, then a month where nothing of yours gets attached is not a Perplexity problem to solve with Perplexity tactics. It is the instrument telling you that the pages are not carrying a claim anybody could stand behind, which is the same finding for every engine, and the one no crawler setting will fix.

Frequently asked questions