How Perplexity decides what to cite

Perplexity is the cleanest target for AI visibility experiments because the Sonar API exposes the citation list in structured form. You don't have to scrape an answer, you don't have to guess what got cited. The model just hands you back an array of URLs with snippets.

From watching which pages Perplexity actually cites, a few patterns stand out about which features of the cited pages correlate with citation likelihood. Here's what we've noticed.

#The signals that mattered

Three came out clearly above the noise.

#1. Topical specificity of the URL

Pages whose URL slug matches the query tend to be cited noticeably more often than pages where the slug was generic. A query for "GPTBot robots txt" was much more likely to cite gpt-bot-robots-txt-guide than ai-crawler-overview, even when the latter contained more relevant content.

This is in line with Perplexity's stated reliance on hybrid retrieval (dense plus sparse), and it suggests slug optimization is more valuable for AI search than for Google in 2026.

#2. Recency, weighted by dateModified

Two pages with similar content. One updated last month, one updated two years ago. The recent one usually wins. The catch: it's dateModified (from schema or Last-Modified header), not datePublished. A page from 2022 that you updated last week beats a page from 2023 that you haven't touched.

For SaaS content this is real leverage. Most product comparison pages are written once and never refreshed. A quarterly refresh pass on your top 20 pages is one of the highest-ROI moves available.

#3. Concrete numbers in the first 200 words

Pages with at least one concrete statistic in the first paragraph (a percentage, a dollar figure, a benchmark, a count) tend to be cited more often than pages with abstract or descriptive intros. Perplexity's training appears to favor sources that "lead with the data."

This is also the easiest fix on the list. Rewrite the lead of your top pages so the second sentence contains a number.

#The signals that didn't matter (as much as expected)

Some of the "AI SEO" advice we tested either had no measurable effect or a smaller effect than expected.

• Page length. Once content quality is accounted for, length past a point doesn't seem to help. Very long pages don't appear to earn more citations just for length.
• Inline FAQ schema. Helped slightly, but nowhere near the lift from Article with proper author and dateModified.
• External link density. A handful of outbound links is fine; piling on many more doesn't help and may slightly hurt.
• Read time. No effect.

#What we tested but didn't have enough signal on

A few hypotheses we couldn't confirm or deny with this sample size:

• Does HTTPS / HTTP/2 matter? Couldn't separate from "site quality" generally.
• Does Cloudflare interstitial harm citations? Likely yes, but we couldn't get a clean test.
• Does llms.txt change Sonar's behavior? Saw a slight lift on sites that had a good one, but it's confounded with "site is generally well-maintained."

#Three things to do this week

If you have an hour:

1. Refresh dateModified on your top 10 pages, AND actually edit them (Perplexity catches "drive-by" date updates pretty well).
2. Rewrite the first paragraph of each so it leads with a number or a concrete claim.
3. Audit your URL slugs against the queries you care about. If your pricing page lives at /p/sku-3001, fix it.

If you're a Crawlmind customer, the audit checks slug quality, dateModified currency, and statistic density on every page audit by default. None of it is mysterious. It's just rarely measured.