How to track Perplexity citations, and why one check means nothing.

This is about measuring where you stand. If you are earlier than that and want to be cited in the first place, start with how to get cited by Perplexity.

01: Four ways to see where Perplexity cites you

1. Ask it yourself, and read the sources

Ask Perplexity the questions your buyers actually ask, in a fresh thread each time so an earlier conversation does not shape the answer, and read the Sources list rather than just the answer text. Record three things alongside each result: the date, the country you asked from, and whether you were named in the text, cited as a source, or both. Those are different outcomes and it is worth keeping them apart. Free, slow, and fine for a first look.

2. Read perplexity.ai referrals in your analytics

Visits that arrive from a Perplexity answer show perplexity.ai as the referrer. This tells you which pages people actually clicked through to, which is the number closest to business value. It badly undercounts citations, though: most people read the answer and never click.

3. Read Perplexity’s crawler visits in your logs

Perplexity runs two crawlers, PerplexityBot and Perplexity-User. The second fetches a page when a user’s question needs it, so its visits are the closest thing in your own logs to “Perplexity read this page to answer someone”. Server logs, your CDN, and some analytics tools such as Microsoft Clarity report these by page. On this site, Clarity logged 78 requests from Perplexity’s bots in the 30 days to October 2, 2026. Being fetched is not the same as being cited, but a page Perplexity never fetches is a page it cannot cite.

4. Record sources at scale: the API or a tracker

Perplexity’s API returns the list of sources it used with every answer, which makes it the easiest of the major engines to measure programmatically, and it is how the measurements in the next section were taken. If you would rather not build anything, several AI visibility tools track Perplexity, but check the tier: as of October 2, 2026, Semrush’s AI Visibility toolkit (from $99 a month) includes it, Peec sells it as a paid add-on on its self-serve plans, Profound includes it on Enterprise, and AirOps on Pro, which has no published price. Details in the Profound, AirOps and Peec comparison.

02: Why one check means nothing

In September 2026 I ran the same contested buyer prompts through Perplexity and ChatGPT dozens of times each, in separate sessions days and hours apart, and recorded every source cited in every answer. Then I compared how often each domain was cited from one session to the next.

Some of it was stable. Some of it was not stable at all. On one prompt in Perplexity, a domain was cited in all 68 of 68 runs, and 40 hours later in 2 of 68. Another domain on the same prompt was cited in 10% of runs, then 97% forty hours later, then 21% eight hours after that. That is not a slow drift in one direction. It is oscillation.

Across the two full sessions, 11 of 109 measured citation rates moved by more than sampling error could explain, after a strict correction for testing many things at once. Eight of those eleven were Perplexity. A comparison eight hours apart on the same day showed as much movement as one 48 hours apart at the same clock time, which rules out a simple nightly refresh: the churn is continuous.

So if you check once and see yourself cited, you have learned that you can be cited. You have not learned how often, and you have not learned whether it will be true tomorrow.

03: How many runs a result actually needs

Treat each run as a yes or no: cited, or not. How many runs it takes to know the rate depends on the rate itself, and mid-range rates are the expensive ones. To pin a rate within plus or minus 10 percentage points at 90% confidence:

Binomial sample size, plus or minus 10 points at 90% confidence. Tested directly at 68 runs: every one of the 67 variable citation rates measured stabilised, including the 20 to 80% band that never stabilised at 15 runs.

There is a catch at the low end, which is where most businesses start. A plus or minus 10 point margin cannot tell 0% from 10%, so it cannot see you go from 4% to 8%, which is a doubling. Measuring the change relative to the rate itself, within 25% of its own value, costs far more:

Same measurement log, September 12, 2026.

The practical conclusion is not “run every prompt 800 times”. It is that a precise rate on a handful of prompts is the wrong thing to buy. Coverage across many prompts, each run enough times to be honest, tells you more for the same effort. More on why tools rarely say this in how to judge AI visibility tool accuracy.

04: A tracking routine that holds up

1. Fix the prompt set. Write down the 10 to 30 questions your buyers ask, in their words, and do not edit them between measurements. A changed prompt is a different measurement.
2. Run each prompt several times per session, not once, and record every source in every answer, not only yours. Competitors moving is half the story.
3. Date every number, and note the country you asked from. A rate from last Tuesday is a historical fact, not a current one.
4. Compare sessions with suspicion.A before-and-after difference can be real churn rather than the effect of anything you changed. Repeat the “after” session before believing it.
5. Check fetches alongside citations. If Perplexity-User never fetches a page, fix crawlability before anything else.

The full measurement method behind this page is on the methodology page.

05: Frequently asked questions

Hami Tahm is an AI visibility consultant based in Toronto.