What does Perplexity actually cite?

Most advice about AI search is written from the outside. This is written from a stored record: eight questions a buyer of our own category actually asks, put to Perplexity through its API on 19 September 2026, run twice about a minute apart, with every answer and every citation kept. The records are 2026-09-19T04-22-26-292Z.json (run A) and 2026-09-19T04-23-37-261Z.json (run B) in web/data/self-audit/, and the numbers below can be recomputed from them.

The shape of the answer set

• Questions asked8
• Answers returned and analysed8
• Total citations90
• Distinct URLs (run A)86
• Distinct URLs (run B)88
• Distinct domains (run A)74
• Distinct domains (run B)75
• Domains cited exactly once (run A)65
• Domains cited exactly once (run B)66
• Citations to a site homepage, either run0

Seven of the eight answers carried exactly ten citations. The eighth — “Best AEO tools for tracking brand visibility in ChatGPT and Perplexity” — carried twenty. A broader, list-shaped question pulled twice the evidence; the rest were capped at the same round number, which is worth knowing before you read anything into a citation count.

Finding 1: it never cited a homepage

Every one of the 90 cited URLs pointed at an interior page. Not a single root domain. The path depth broke down like this:

This is the single most actionable thing in the data. A homepage is a page about a company. An answer engine is not looking for a company; it is looking for the paragraph that answers the question in front of it. If everything you know is on your homepage and your product pages, there is nothing here for it to cite.

Finding 2: the URL nearly restates the question

We checked every cited path against the words of the question it was cited for. Here is the rule in full, because the number is worthless without it, and “roughly how we did it” is not good enough to be checked:

On that rule, 74 of 90 matched in run A, and 75 of 90 in run B. The stem clause is worth spelling out that precisely because it moves the answer: with no stemming at all the figure is 68 of 90 in run A, and stripping only a plural -s gives 73 of 90 in run A. If you implement “a simple stem” differently from us you will get a different number and reasonably conclude we are wrong, so the exact rule above is what we mean, and scripts/check-citation-rule.py in our repository is the executable version of it. That script also runs in our checks and fails the build if this page ever states a figure it cannot re-derive. For one question the match was total:

The three-letter floor is doing real work, and it is the most interesting part of the rule. A four-letter floor would exclude aeo — and aeo alone carries 11 of run A’s 74 matches, all of them on the two questions about budgets and subscriptions, where the cited paths are things like /reviews/hubspot-aeo/, /best-aeo-tools/ and /resources/affordable-aeo-agencies. Move the floor to four letters and the figure drops to 63 of 90 in run A, and 64 of 90 in run B. The acronym is how this category names itself in a URL, so a rule that cannot see three-letter words cannot see the category at all.

The strictness elsewhere is worth dwelling on, because this is where analyses like this usually go soft. Take “How much does answer engine optimization cost per month?”. Six of its ten cited paths contain the literal word cost. A seventh, /blog/answer-engine-optimization-pricing-2026/, matches anyway — not on pricing, which is not a word in the question, but on answer, engine and optimization, which are. The three that score zero are /blog/aeo-and-geo-pricing-guide/, /aeo-pricing and /blog/aeo-agency-pricing — obviously the same topic, and all three turn on pricing, which the rule does not accept.

Relax it by exactly two synonym pairs — cost matching pricing and price, and cite matching citation and citations — and the total goes to 80 of 90 in run A and 81 of 90 in run B. That is the whole of the relaxation, and you can check it. We publish the stricter 74.

The 16 misses in run A are not concentrated anywhere in particular: two on the category question, five on the broad tools question, two on the monthly-subscription question, three on the cost question and four on the citation-tracking question. The two largest account for nine of the sixteen. Most are pages plainly on topic that name it with a different word.

We are not claiming the URL slug is a ranking factor — nobody outside Perplexity can claim that, and correlation with the page’s title and heading is obvious. What the data shows is the shape of the winning page: one question, one page, and the question visible in the address bar.

Finding 3: there is no trusted club

Of the 74 domains cited in run A, 65 were cited exactly once, and 69 appeared for only one of the eight questions. Just five domains turned up for more than one question:

The most-cited domain in the whole run accounted for five of 90 citations — under 6%. HubSpot does not win here because it is HubSpot. It wins because blog.hubspot.com/marketing/answer-engine-optimization-tools is a page about exactly that, and it was cited for three separate questions that all touched tooling.

That tail is the reason this is a winnable game for a company nobody has heard of. The engine is not consulting an approved list. It is assembling an answer out of whichever pages most directly address the question, and most of those pages belong to organisations with no more authority than yours.

Finding 4: almost nothing came from the places SEO advice sends you

Across all 90 citations, in both runs, there were zero from Reddit, YouTube, Wikipedia, Quora, G2, Capterra, Trustpilot or Product Hunt. Each run had one LinkedIn post and one Chrome Web Store listing. Mainstream publishers accounted for five of 90 in run A — techradar.com twice, and finance.yahoo.com, coursera.org and sitepoint.com once each. In run B the same set of publishers accounted for three of 90: techradar.com, finance.yahoo.com and sitepoint.com once each, and coursera.org was not cited at all.

That leaves 83 citations in run A, and 85 in run B, from company and agency websites writing about their own field. The domain list is full of names that did not exist two years ago, on the TLDs that go with that: in run A, 54 of the 74 domains were .com, eight were .ai, seven .io, two .org, and one each .co, .life and .site; in run B, 56 of 75 were .com and only one was .org, the rest unchanged.

Be careful with how far you take this. It is one category, and ours is an unusually new one where Reddit threads and G2 categories barely exist yet. In a mature category the mix would almost certainly be different. What it does establish is that a citation does not require a review-platform presence, because in this run nobody’s citation did.

Finding 5: a vendor’s own marketing pages get cited

The tools question cited peec.ai four times — on /chat-gpt-visibility-tracker, /ai-mode-visibility-tracker, /for-agencies and /ai-instructions. It cited ahrefs.com/brand-radar, ahrefs.com/faq and an Ahrefs help article. The services question cited conductor.com, seo.com, elsner.com and searchbloom.com on their own service pages. In run B it cited tryprofound.com itself.

Those are not editorial sources. They are product pages, and the engine used them as evidence about what the product does. Every one of them is a narrow page with the feature name in the URL. A generic “Platform” or “Solutions” page is not in that list anywhere.

Finding 6: the run is stable at the top and noisy underneath

We ran the identical eight questions twice, roughly a minute apart. The headline metrics were identical — 8 of 8 answers, 90 citations, the same brands named, the same visibility score of 0. Underneath, things moved:

• 71 domains appeared in both runs. Run A cited 74 distinct domains, and run B cited 75. Run A had three that run B did not (coursera.org, thatmarketingbuddy.com, vismore.ai); run B had four run A did not (discoveredlabs.com, marceldigital.com, semai.ai, tryprofound.com).
• Per-domain counts moved. ahrefs.com went from 3 to 4, blog.hubspot.com from 5 to 4, peec.ai from 4 to 2.

So treat a single run’s per-domain count as an estimate with roughly a citation of noise on it, and treat the trend across months as the signal. Any report that gives you one run’s per-domain figures to two decimal places is selling precision it does not have. Our full handling of this is on how we measure AI visibility.

What we did with the finding

We scored zero on this run: not one of the 90 sources was bungad.com. The analysis above is the reason we then built a library of pages, each titled as one question, rather than buying links or rewriting the homepage. Whether that works is an open question, and we will publish the answer either way — the same eight questions, the same engine, every month. If bungad.com is still absent by the third month, the thesis on this page is wrong and we will say so.

How to reproduce this

1. Get a Perplexity API key and list eight to twenty-five questions a buyer of your product would actually type. Not keywords — questions.
2. Ask each one and store the whole response, including the citation array. The analysis above is entirely from that array.
3. Run it twice in the same session. If your two runs disagree wildly, your question set is ambiguous, not the engine.
4. Count four things: citations per answer, distinct domains, how many domains appear exactly once, and how many cited URLs contain a word from the question.
5. Then look for yourself in the list. That number — yours over the total — is your citation share.

Ours is on our score page, in full, including the eight questions and every domain cited instead of us.

Common questions