When you ask ChatGPT a question, the text that comes back is a single piece: the model merges what sits in its own weights with, if search is on, whatever it distilled from a handful of pages, and the sources, when they exist, sit under the text as a secondary layer. Perplexity inverts that order. The product's default behaviour is to build every answer on top of a live search and to place numbered citations next to the sentences; when the user taps the [1] or [3] badge, they land on the page that claim was taken from.
That structural difference has a practical consequence. Being visible in Perplexity does not mean "the model remembers you", it means "one of your pages was selected into the source list for that answer". The second of those is an event that can be observed, repeated and influenced at the page level. Getting your brand name lodged in a model's general knowledge is a long haul; we covered that side in how your brand gets recommended in ChatGPT and Gemini. Because Perplexity behaves much more like a search engine, a single well-structured page can surface in the source list in a relatively short time.
How Perplexity assembles an answer
Several separate steps sit between the sentence the user types and the paragraph that appears on screen. We describe those steps from the product's externally observed behaviour; the internal details are closed, but the order that shows on the outside is consistent.
The query is split into sub-questions. A compound question such as "how long does corporate legal advisory take in Ankara and which documents are required" does not turn into one search query. The system breaks it into several narrow queries and runs a separate search for each. So a page enters the source list by matching not the whole sentence the user typed, but one of the sub-questions derived from that sentence.
A live search runs. What the answer rests on is not the model's training data but the web results pulled at that moment. Classic search visibility therefore remains a precondition: a page that does not appear at reasonable positions in any search index will struggle to reach the candidate pool at all.
Candidate pages are filtered down. The system does not take dozens of results into the summarisation step, it takes a handful of sources out of them. The filter is not ranking alone; whether the page answers the sub-question directly, how extractable the text is, and the identity of the source all play a part.
The summary is written and citations are attached. The model produces sentences from the selected pages and marks every claim with a number. That last step says something precise: to be cited, it is not enough for your page to be "good", it has to be possible to pull a sentence out of it that stands on its own.
So there are two separate contests. The first is getting into the search results, the second is being picked out of those results for summarisation. Most sites work on the first one and never think about the second.
What cited pages have in common
Pages that turn up in source lists over and over share a few traits. These are not a magic formula, only ordinary editorial decisions that raise extractability.
The answer is finished at the top of the page. On cited pages the information being looked for is usually given on the first screen, without any context-building first. A long introduction, a brand story, or a warm-up paragraph of the "this subject has been growing in importance in recent years" kind pushes the answer down and lowers extractability.
The date is visible. The publication and update dates are present both in readable form on the page and in the structured data. On subjects that change over time, such as pricing, regulation or product features, an undated page is at a disadvantage next to a dated competitor.
The author and the organisation are identifiable. Sites where it is unclear who wrote the text, where there is no contact information and no about page, produce no identity signal. An author name, a title, the organisation behind the work, and structured markup supporting all of it make the source easier to tell apart.
The page stays on a single topic. Vast guides that answer ten different questions at once match more weakly than narrow pages that give a clear response to one sub-question. A query that gets broken into sub-questions rewards narrow pages.
Together these four form a threshold. None of them on its own guarantees a citation, but if all four are missing, the page is open to elimination at the summarisation step even when it ranks well in search results. Going through your existing content with these four items as a checklist produces results faster than writing new pages.
PerplexityBot and access: the technical preconditions
However good your content is, if the crawler cannot read the page the chance of a citation is zero. This section comes before content.
robots.txt. Perplexity defines two separate agents: PerplexityBot, which crawls content and takes it into an index, and Perplexity-User, which goes out to fetch the page at the moment a user asks a question. Closing one and leaving the other open produces different outcomes. Taking a blanket "block the AI bots" decision and then waiting to be cited is a contradiction. We collected the details of managing which door is open to which bot in what llms.txt is and how to manage AI crawlers.
Server-side content. If the page's text only comes into being after JavaScript has run in the browser, there is no guarantee that any bot will see that text. The safe assumption is this: a sentence you cannot see in the HTML source with view-source may not be cited. Static generation or server-side rendering removes that risk entirely.
The CDN and the security layer. In practice this is the most insidious obstacle. Your robots.txt file may be open to AI bots, while a "block AI bots" switch left on in the CDN dashboard in front of it rejects the requests before they ever reach your server. On this site robots.txt explicitly gives Allow: / to GPTBot, ClaudeBot, PerplexityBot and the other agents, and for exactly that reason the automatic AI blocking feature in the CDN panel is deliberately kept off. Because either layer can silently cancel the other out, the two have to be verified together. The fastest way to verify is to send a request to the page with the relevant bot user-agent and see that it really returns 200 and full HTML.
Page speed and stability. Crawlers do not have infinite patience. Pages that respond slowly, return intermittent errors, or carry heavy redirect chains drop out of the candidate pool quietly.
Writing content that can be quoted
Once the technical door is open, the job comes down to the shape of the text. The aim is for a language model to be able to extract from the page an answer sentence that stands up on its own.
Turn headings into questions. A ## heading that reads "how long does the mediation process take" instead of "our services" makes sub-question matching directly easier. The more the heading resembles the sentence the user typed, the higher the chance that section gets selected.
Finish the answer in the first paragraph. The paragraph immediately under each heading should give the response to the question even if nothing after it is read. Detail, exceptions and nuance are left to the paragraphs that follow. This is the same inverted pyramid journalism uses, and it is the single highest-return edit for being quotable.
Write a definition sentence. Full sentences in the "X is a Y that ..." pattern are easy for models to lift out. Do not use pronouns when you are explaining a concept: instead of "this method delivers the following", write "generative engine optimization delivers the following". Pronouns lose their meaning once the quote is torn out of its context.
Do not leave numbers bare. If you are going to give a figure, state its source and its date in the same sentence. Percentages without a source both lower the trust signal and carry the risk of being quoted wrongly.
Back lists and tables up with prose. Bulleted lists are useful for scanning, but pulled out of context they are usually meaningless. Put a sentence above or below every list that explains on its own what the list says. The same holds for tables: a line that sums up the table's conclusion in one sentence is quoted more often than the table itself.
We went through the common patterns that do the opposite of these items, one by one, in seven mistakes people make writing content for AI.
How to track your Perplexity visibility
There is no console in the classic sense in this area; the measurement is largely set up by hand. Done with discipline, it is still reliable enough.
Prepare a fixed query set. Write 15-30 questions your customers would genuinely type into a table. They should not contain your brand name; they should express a need, in the shape of "how do I choose employment law advisory in İzmir". The set has to stay fixed, otherwise periods cannot be compared with each other.
Record three separate things. First, whether your domain appears in the answer's source list. Second, if it did appear, which URL was selected, because it is often not the page you expected and that fact on its own is directive. Third, whether your brand name is mentioned inside the text without a citation; those two are different signals.
Measure under the same conditions. Being logged in or logged out, and the language and location settings, change the result. Run the measurement the same way every time and note the date. A single observation is noise; what carries meaning is the trend of the same set over weeks.
Look at the server logs. Filtering PerplexityBot and Perplexity-User requests out of your access records shows which of your pages are genuinely being fetched. This is a data layer independent of the source list observations and more objective than they are.
If you want to place this measurement inside a wider frame, the metric set in how to measure AI visibility: five core metrics will do the job.
Three mistakes people make often
Blocking the bots and then expecting to be cited. Shutting AI crawlers out over content-theft concerns is a defensible preference, but the consequence has to be accepted: a page behind a closed door cannot get into the source list. The decision should be taken deliberately, not arrive by itself because of a switch someone forgot in a dashboard.
Piling everything into one enormous page. The "ultimate guide" format is good at collecting links and weak at sub-question matching. Splitting a subject into a few narrow pages that each answer clearly usually produces more source appearances than keeping the same content in a single file.
Opening with the brand narrative. An introductory paragraph that explains how many years you have been in business pushes the answer to the question below the fold. Company history can stay on the page, but it has to come after the answer.
These three corrections can usually be applied without producing any new content, by changing the structure of the pages you already have. If you want to see which of your pages meet these criteria, you can find the checklist we follow on our solutions page.
Frequently Asked Questions
What does being cited as a source in Perplexity mean?
Perplexity grounds every answer in a live web search and puts numbered citations next to the sentences it produces. Being cited as a source means one page from your site appears in that numbered list and the user lands on your site when they tap the badge. It is different from the model memorising your brand: it happens at the page level, it can be observed one instance at a time, and it can be influenced directly through content and technical changes.
What happens if I block PerplexityBot in robots.txt?
Your page cannot enter Perplexity's crawling and indexing flow, so it is not expected to appear in source lists. Perplexity also uses the Perplexity-User agent, which fetches the page at the moment a user asks a question; the two are defined separately and can be managed separately. A decision to block can be legitimate, but the loss of visibility has to be accepted along with it. The real danger is a block left switched on in a CDN dashboard without anyone noticing.
Should I write separate content for Perplexity?
No. Rather than writing a separate text, it is more productive to restructure the pages you already have: put headings into the shape of the question a user would ask, finish the answer in the first paragraph under each heading, build full definition sentences instead of pronouns, and make the date and author information visible. These edits do not lower classic search performance either, and generally improve it.
Can I be cited if my site is rendered with JavaScript?
That would be a risky assumption. There is no guarantee that an AI crawler will run the JavaScript on your page, and even if it does, the extra delay raises the chance of being filtered out. The safe criterion is this: do not expect a sentence that does not appear as plain text in the HTML source to be quoted. Static site generation or server-side rendering removes that uncertainty completely and improves page speed as well.
Is there an official tool for measuring Perplexity visibility?
There is no official reporting panel comparable to the search consoles. In practice two sources are combined: asking a fixed question set by hand at regular intervals and recording whether your domain comes up in the source list, and filtering your server access logs for which of your pages the Perplexity agents fetched. A single observation is noisy; the meaningful signal is the trend of the same set over weeks.