Ask Perplexity a question and you get one answer with a short row of numbered sources beside it. Not ten blue links. If your page is not one of those sources you are invisible on that query, whatever your Google position says. That is the entire problem Perplexity AEO exists to solve.
How Perplexity builds an answer, and why it is not a ranked list
Perplexity answers a question by running its own searches, fetching a handful of live pages, and writing a summary that cites them by number. The output is one composed paragraph with attributions, not a ranked list of ten results. Winning means being retrieved and quoted inside that paragraph, not sitting at position three.
A search engine ranks documents. An answer engine assembles one. A ranking system rewards the page that best matches a query overall, while an answer engine has to find a passage that resolves the specific question in front of it. Those are rarely the same paragraph. A three-thousand-word pillar page can rank beautifully and never get cited, because nothing inside it states a claim compactly enough to be lifted out. Meanwhile a plain page that answers one question in its first four sentences gets pulled in again and again.
The second thing to understand is that retrieval happens at question time. Perplexity is not answering purely from what a model absorbed during training; it goes and fetches pages while the user waits. A page you publish this month can be cited this month, with no retraining cycle to wait on and no visibility into one either. In my experience, of all the answer engines, this is the one where ordinary publishing work still moves fastest.
My take: it is the most winnable answer engine precisely because it shows its work. You can read the sources it chose, compare them to your page, and see exactly what they did that you did not.
What makes a page citable rather than merely crawlable
A citable page states its answer in plain language near the top, attaches that claim to something verifiable, and makes the origin of the claim obvious. Sentences that carry a clear, attributable claim are the ones I see quoted. Bury the answer under three paragraphs of preamble and a retrieval system will take the competitor who said it in line one.
- Answer first. The forty to seventy words under each heading should stand alone. Paste them into a message with no context: if they still make sense, they are quotable.
- Claims carrying a date. An undated assertion is a weaker thing to attribute than a dated one. Put a reviewed-on date on the page and dates inside the sentences that need them.
- A named human. An author with a reason to be believed, a real bio, and a consistent identity across the site.
- Specifics instead of adjectives. Named tools, named constraints, real trade-offs. Adjectives are unquotable.
- One idea per heading. A heading that covers three things produces a paragraph that answers none of them cleanly.
The paragraph test I run before anything goes live
I copy the first paragraph under each heading into a blank message and read it back cold. If I have to scroll up to work out what "this approach" refers to, the paragraph fails and gets rewritten with the subject named. Pronouns pointing back at the previous section are the most common reason a genuinely good page never gets quoted anywhere.
The technical layer Perplexity AEO actually depends on
Perplexity AEO depends on unglamorous infrastructure: HTML that contains the answer in the source, headings that describe the content honestly, a robots policy that permits the AI crawlers you want, and machine-readable files describing the site. If the answer only appears after JavaScript executes, assume a retrieval fetch may never see it.
- Server-rendered answers. Client-only rendering is the failure I find most often. The page looks perfect in a browser and arrives at a fetcher as an empty shell.
- An intentional robots.txt. The exclusion protocol is a published standard (RFC 9309), and AI crawlers announce themselves under their own user-agent names. Plenty of sites opt out of citation by accident, through a wildcard rule a plugin wrote for them.
- Headings phrased as questions. Retrieval works on passages. Headings are the boundaries of a passage.
- Structured data. Article, FAQPage and Organization types from the Schema.org vocabulary give a machine an unambiguous read on authorship and dates.
- An llms.txt file. A short map of your important pages at the root. Our free llms.txt generator builds one in a minute. It is not a ranking factor and nobody should sell it to you as one.
None of this is exotic. It is the same hygiene good SEO has always wanted, with a lower tolerance for pages that hide their meaning behind a framework.
Classic SEO and AEO optimise the same page for different jobs
Classic SEO and AEO ask different questions about one page. SEO asks whether the page deserves to outrank nine others for a query. AEO asks whether a single passage inside it can be lifted, attributed and trusted inside a composed answer. Most of the underlying work overlaps. The parts that do not are where citations get lost.
| Element | Classic SEO optimises for | AEO optimises for |
|---|---|---|
| Unit of success | The page ranking in a list | A passage quoted in an answer |
| Title | Click-through from a results page | Topical clarity for a retriever |
| Answer position | Anywhere, if the page satisfies intent | Immediately under the heading |
| Length pressure | Longer, to cover the topic fully | Compact, self-contained blocks |
| Freshness | Helps rankings decay slower | Decides whether a claim is safe to quote |
| Reporting | Impressions, positions, clicks | Referrals, crawler hits, manual prompt checks |
If you want the wider argument about how the two disciplines are converging, we made that case in AEO vs SEO in 2026. This article assumes you have already accepted the premise and want the Perplexity-specific mechanics.
What can you honestly measure?
You can honestly track four things: referral sessions from perplexity.ai in your analytics, AI crawler hits in your server logs, manual prompt checks against a fixed list of buying questions, and branded search volume over time. You cannot track impressions or citation share, because no answer engine publishes them. Anyone quoting you a precise citation rate is estimating.
Referral sessions
Answer engines send a small but unusually qualified stream of traffic. Segment it in analytics on day one so you have a baseline before you change anything.
Crawler hits in logs
Server logs name which AI user agents fetched which URLs, and when. It is the earliest signal that a new page has entered the retrieval pool at all, and it arrives long before any human clicks through from an answer, which makes it the metric I check first after a publishing push.
A fixed prompt panel
Asking a fixed set of prompts on a schedule is the only direct read on whether you are being cited. Building that panel is a separate exercise from the page work here.
Branded demand
Being named inside answers lifts searches for your company name first. Watch that curve.
Why I refuse to report a visibility score
Every AI visibility number on the market is a sample of prompts a tool chose, run from a location a tool chose, on a day it chose. The number moves when the sample moves. I would rather hand a client twenty recorded answers with the sources visible than one index that nobody can audit or reproduce.
When I would tell you to skip Perplexity work
Perplexity AEO is not worth paying for when nobody researches your category in text. A restaurant chasing walk-ins, a barbershop, a boutique whose customers arrive from Instagram, a contractor living entirely on referrals: these businesses will see almost nothing from it. Answer engines reward considered, researched, question-heavy purchases with a long decision window.
The second case is a site that has not done the basics. If your pages load as empty shells, your service pages say nothing specific, and you publish twice a year, buying an AEO retainer is paying for a multiplier on zero. Fix the content and the rendering first. The AEO layer costs very little once the underlying page is worth quoting.
So how do you tell which side of the line you sit on? Ask your last ten customers what they typed before they contacted you. If the answers are questions with several clauses in them, comparisons, requirements, regional constraints, then answer engines are already in your buying journey. If the answers are your brand name or nothing at all, spend the money somewhere else this quarter and revisit it next year.
Where to spend your next four weeks
Start with the ten questions your customers ask before they buy, and rewrite the page that answers each one so the answer sits in its first paragraph. That single pass does more for citation than any file you can drop into your root directory. Everything technical is a multiplier on content that already answers something clearly.
Do the rewrite yourself if you have a writer who understands the product; the mechanics are in our AI search optimization guide. Bring in help when the bottleneck is the technical layer or the volume of pages, which is the work our AI Search Optimization service exists to carry. Either way, pick the ten questions this week. The retrieval pool refreshes constantly, and it cannot pick a page you have not written.