Short answer: Perplexity SEO, the work of getting cited by Perplexity, comes down to four things. Allow PerplexityBot, serve your content as HTML that reads without JavaScript, write passages that answer one question each with a real date on anything time-bound, and keep the third-party pages that get cited in your category accurate about you. Then count citations the strict way: only sources marked in the answer text, by number of distinct questions rather than raw volume, and kept apart from whether you were recommended. Perplexity cites sources on nearly every answer, so both access mistakes and measurement mistakes surface here faster than on any other engine.
That last point is the reason to start with Perplexity even if it is not your biggest engine. On an assistant that rarely shows sources, a blocked crawler or an unquotable page shows up only as a vague absence. On Perplexity it shows up as a missing citation on a question you can name.
Two agents, and why blocking PerplexityBot is rarely the decision you mean
Perplexity's bot documentation names two agents, and they do different jobs:
| PerplexityBot | Perplexity-User | |
|---|---|---|
| What it does | Indexes pages so they can be surfaced and linked in Perplexity answers | Visits a page because a user's question requires it, and links it in the response |
| Used for model training? | No, per the documentation | Not described as a training crawler |
| Follows robots.txt? | Yes | Generally not, because the request came from a person |
| What blocking it costs | Removal from Perplexity answers, visible in citation data almost immediately | A visit a real user asked for fails |
| How to stop it, if you must | A robots.txt group | Only an edge rule at your CDN or firewall |
The first row changes the usual reasoning. Many sites block AI agents to keep their content out of model training. By Perplexity's own account, PerplexityBot is not a training crawler, so blocking it achieves nothing on that front and removes you from Perplexity answers. If training is your concern, the agents to address are the training crawlers such as GPTBot and ClaudeBot, which is covered in how to block GPTBot without disappearing.
# Perplexity: indexing for answers. Not used for foundation-model crawling,
# per Perplexity's bot documentation, so there is no training reason to block it.
User-agent: PerplexityBot
Allow: /
# Perplexity-User fetches on a person's request and generally does not
# follow robots.txt. Listed here for the record, not as access control.
User-agent: Perplexity-User
Allow: /
Because Perplexity-User generally ignores robots.txt, the only way to stop it is an edge rule at your CDN or firewall, and that rule fails a request a real person made. Most sites should not write one. Note also that the two groups only matter if nothing broader overrides them: a crawler obeys the most specific group that names it, so a site with a catch-all User-agent: * block and no PerplexityBot group is blocking Perplexity whether it meant to or not.
Perplexity SEO starts with access: check that it can read the page
A permissive robots.txt in front of a CDN rule or bot-management product that returns 403 to these agents is a block in practice, and the file will keep saying otherwise. Test with the real agent names and a browser-like control in the same run, so a site that is simply down is not mistaken for one that blocks AI agents:
for ua in "Mozilla" PerplexityBot Perplexity-User; do
printf "%-16s" "$ua"
curl -s -A "$ua" -w " %{http_code}\n" https://example.com/guide/ | wc -c
done
Read the result against the control row, never on its own:
| Control (browser) | Perplexity agents | What it means |
|---|---|---|
| 200, full page | 200, full page | Reachable. Move on to the content. |
| 200, full page | 403 or a challenge page | Something in front of the site overrides your intent: a CDN, WAF or bot-management rule. |
| 200, few bytes | 200, few bytes | The page depends on JavaScript. Most AI crawlers do not execute it, so they receive an empty shell. |
| Timeout or 5xx | Timeout or 5xx | The site is down or slow for everyone. This says nothing about AI access; test again later. |
A 200 with an empty body is the result people most often misread as a pass. Our guide to JavaScript rendering and AI crawlers covers the fixes, and the PerplexityBot glossary entry keeps the agent facts in one place.
Perplexity SEO at the page level: what makes a passage citable
Once Perplexity can read a page, the question is whether a passage on it is worth quoting next to a claim in an answer. The patterns below are not Perplexity-specific, but its frequent citations make the difference easy to see.
- One question per heading, answered in the first sentence. An assistant lifts sentences that answer a constraint: a dimension, a material, a return window, a compatibility. The sentence has to name its subject so it still makes sense when cut out and pasted into a reply.
- Facts, not adjectives. Copy written for brand tone gives an engine nothing to cite, however well it reads. Put the specific fact in a sentence, not only in a widget or a spec table.
- Sources, quotations and statistics on the page. The original GEO paper (Aggarwal et al., accepted to KDD 2024) reports that its methods raised visibility in generative engine responses by up to 40%. Citing sources, adding quotations and adding statistics performed best, while keyword stuffing offered little to no improvement. That was measured on a research engine, not on Perplexity itself, but it points the same way as everything else here: give the engine something checkable to quote.
- A real date on anything time-bound. Questions that imply a current answer favour dated, recently updated pages. An undated page, or one whose modified date is bumped by every deploy, loses to a page that states a real date. Content freshness and dates covers how to stop the build pipeline from faking it.
- Method stated on comparison pages. Comparison and round-up pages match the shape of a buying question, which is why they dominate citation lists in commercial categories. A round-up that states its method and criteria is far more quotable than one that simply lists products.
A useful test is to copy the first two sentences under each heading into an empty document. If they still answer the heading's question without the paragraph above them, the passage is liftable. If they begin with "This" or "As mentioned", they are not.
When Perplexity cites someone else
Perplexity's own citation habits explain where much of the competition comes from. Profound's analysis of 680 million citations found that Reddit was the single most-cited domain in Perplexity, at 6.6% of all citations, with YouTube second at 2.0%. ChatGPT's leader was a reference work instead, Wikipedia at 7.8%.
Our own benchmark points the same way from a different angle. In the product categories we measured, a large share of citations went to marketplaces, review sites and publisher round-ups rather than to brand sites. Your product page can be perfect and still lose the citation to a listing page or a discussion thread that carries the same facts in a third party's words.
That shifts part of the work off your own site. Find the pages Perplexity cites for the questions you care about, then check whether they carry your current facts, prices and availability. If a frequently cited marketplace page still shows an old price, that old price is what an answer citing that page will draw on. For discussion sites, the honest move is to make sure accurate information exists where people ask, not to seed threads; a planted answer that is later exposed costs more than the citation was worth.
Keep in mind what a citation buys. Most answers are read without a click, so for a publisher, being the cited source is a brand outcome rather than a traffic one. Measure it on the answer side, not in your analytics.
How to count Perplexity citations without inflating them
Perplexity's habit of showing sources makes citation data plentiful, and plentiful data is easy to inflate. These are the rules we apply in our own product, most of them added after a mistake:
- Only sources marked inside the answer text count. An assistant typically retrieves a dozen or more pages and cites a few. Counting the retrieval pool instead of the in-answer markers overstated our own citation figures by two to four times before we fixed it.
- A marker that does not resolve to a real URL is discarded. Recording one citation fewer is better than recording one that is not real.
- Breadth is reported alongside volume. One answer citing twelve of your pages is not twelve times the reach. The number of distinct questions that cited your domain is much harder to inflate.
- Cited and recommended are kept apart. Your page can be cited as a source while the answer recommends a competitor. Merging the two flatters the report and hides the real gap.
Run the same fixed set of buyer questions each time, with your brand name kept out of them, and report Perplexity on its own before any cross-engine average. In our own benchmark one model named 3.5 providers per answer and another 2.1, so a blended number describes neither engine. Treat a change as real only when it holds across two consecutive runs, because answers vary between identical runs. Citation analysis walks through reading the result, and the Perplexity engine page shows how often it cited sources across the categories in our benchmark.
Sources
- Perplexity, crawler documentation (PerplexityBot and Perplexity-User).
- Profound, citation patterns across AI platforms.
- Aggarwal et al., GEO: Generative Engine Optimization, arXiv, accepted to KDD 2024.
- OpenAI, crawler documentation (training versus search agents).
- Our own product rules and incident records, July to September 2026, and our Q3 2026 benchmark (first-party).