How to See Mentions in Perplexity When One Check Sees Half Its Sources

To see mentions in Perplexity, ask it your buyers' questions and read both the answer text (where your brand is named) and the numbered sources under it (where your site is cited), then repeat each question more than once. In the Machine Relations Index, across 434 questions each run three times over a median of two weeks, a single Perplexity run showed about half (48.5%) of the question-source pairs those three runs cited; two runs showed about four in five (78.8%).

That second sentence is the part most tracking advice skips. Nearly every Perplexity answer carries sources, which makes it the easiest engine to watch. It does not make one look representative. The leading source on a question is stable; most of the rest rotate between runs. About a third of the runs in that read (419 of 1,302) fall on or after October 7, 2026, when Perplexity's API began returning up to 10 sources per answer (4,225 of the 4,306 runs since carry exactly 10), so it blends two source-list sizes. Split at that date, prompts with four runs showed the same shape on both sides: before the cap (August 1 to October 6, 242 prompts) one run showed 47.5% of the question-source pairs and two runs 71.2%; after it (October 7 to November 9, 571 prompts), 46.8% and 70.3%.

The finding in one table #

The October 13, 2026 Machine Relations Index release, Perplexity only. The matched-count read takes the 434 monitored prompts that Perplexity answered exactly three times between 2026-09-14 and 2026-10-13 (median span between first and last run: 14 days), 1,302 answer runs in all.

Read it as two facts. The source Perplexity leans on most for a question is almost always there when you look: on 376 of the 434 prompts (86.6%) it was cited in every run. Everything below it is a rotation: of 12,275 prompt-source pairs, 7,792 (63.5%) appeared once in three runs.

The same read before and after the October 7 cap #

The matched-count window above straddles a change on Perplexity's side: 419 of its 1,302 runs fall on or after October 7, 2026, when its API began returning up to 10 sources per answer, where it had returned about 19 to 20 from July 31 to October 6. Its 13.73 average is a blend of the two (the ledger's citation count per run averages 20.0 before October 7 and 10.0 from then in that cohort). Of the 4,306 Perplexity runs since October 7, 4,305 carry 10 or fewer sources and one carries 18. The ledger records the same provider model (sonar) before and after. This table splits the read at the cap, with prompts that had exactly three or four Perplexity runs from August 1 to October 6, and from October 7 to November 9 (4,306 runs on 34 days, every one citing at least one source). The last column is the table above.

Compare columns with the same run count, not across them: the more runs a prompt has, the more distinct sources it collects in total, so a single run covers a smaller share. The same-run-count pairs are the first two columns (three runs each) and the third and fourth (four runs each). The post-cap three-run cohort has only 104 prompts, so it is described, not tested. Distinct domains per answer fell from about 16 to about 9. The share one run shows moved from 53.3% to 56.4% on three-run prompts and from 47.5% to 46.8% on four-run prompts. Pairs cited in only one run fell from 57.9% to 54.0% on three-run prompts and rose from 50.9% to 52.3% on four-run prompts. The most-cited source was in every run on 97.4% and then 98.1% of three-run prompts, and on 95.0% and then 93.0% of four-run prompts. After the cap, single runs showed about the same share of a prompt's sources (56.4% against 53.3% on three-run prompts, 46.8% against 47.5% on four-run prompts). The Sep 14 to Oct 13 column is a blend and is not a baseline. The runs before October 7 also span Perplexity's September 27 Sonar change, so the split is not a controlled comparison, and this page claims no cause for the differences between columns.

How to see mentions in Perplexity, step by step #

1. Write the questions your buyers ask, without your brand in them. A question that names you will produce an answer that names you. The useful read is whether Perplexity brings you up when nobody asked about you.
2. Run each question at least twice in the period you report on. In the Index sample, one run showed about half of the sources three runs cited on the same questions; two runs showed about four in five. A weekly single check is a sample, not a census.
3. Log the answer text and the source list separately. A brand named in the answer is a mention. Your URL in the numbered sources is a citation. They are different events, and a tool or spreadsheet that merges them hides which one you are missing.
4. Log the sources that win when you do not. The source list is a free map of what Perplexity trusts on that question. The Machine Relations Index publishes the same map across categories and question types, ranked by citation rate.
5. Report presence as a rate, not a yes or no. "Cited in 2 of 3 runs" carries more information than "cited", and it survives the next run.
6. Treat a missing leader differently from a missing tail source. If the source that was cited in every earlier run disappears, something probably changed. If a source that appeared once disappears, that is what Perplexity normally does.

Why one check misleads #

Most of a Perplexity answer's sources rotate between runs. Across the full release window (2026-05-10 to 2026-10-13), the Index recorded 5,620 Perplexity answer runs on 1,143 prompts, and 5,619 of them cited at least one source. Those answers cited 12.47 distinct domains on average (median 11), against 4.17 for ChatGPT and 4.39 for Claude in the same Index window. That window pools several Perplexity source-list sizes: per-answer source counts that averaged about 8 before mid-July and about 14 from July 14 to 30, about 19 to 20 from July 31 to October 6 and up to 10 from October 7. From October 7 to November 9, all 4,306 Perplexity runs cited at least one source, at 8.63 distinct domains per answer on average (median 9), against 4.66 for ChatGPT, 5.22 for Claude, 6.28 for Google AI Overviews and 2.65 for Google AI Mode (averages over answers with at least one citation). Gemini's 9.48 in that window rests on 327 of its 3,079 runs, because most Gemini answers cited no source in this window (see the Gemini model-change study), so it is not a like-for-like figure. The shorter post-cap list still rotates, as the table above shows.

The same pattern holds on every engine the Index monitors, and Perplexity is the steadiest of the six. Over the full window, which pools the October 7 Perplexity cap and the Gemini model changes, on prompts each engine answered at least five times:

Runs in this table are spread across the window (a median of 95 days between first and last run for prompts with six or more Perplexity runs), so the full-window rates mix run-to-run variation with real change over time. The matched-count 30-day read shows a higher single-run share (48.5%) than the full window (35.3%), but the two samples differ in run count (exactly three, against five or more) and in prompts, so those shares are not comparable with each other. Each shows that a single run covers a minority of the sources a prompt's runs cite.

Restricted to October 7 to November 9, the window after the cap, with one source-list size, Perplexity is still first on every column shown. ChatGPT is still ahead of Claude on the share cited in only one run and on the share one run shows, and Google AI Mode and Google AI Overviews swap places on the cited-once, one-run and leader columns. These are prompts each engine answered at least five times in that window and that drew at least one citation:

Gemini is left out of this table: only 3 prompts had five or more Gemini runs and a citation in the window, so it cannot be compared. Google AI Overviews has 15 prompts, so its row is thin. The consecutive-run overlap column is not recomputed for this window.

Overlap is the share of domains two consecutive runs of the same prompt have in common, out of all domains either run cited. It uses a wider cohort than the other columns: every prompt with at least two answers that cited sources, each compared with the previous one. A median of 0.353 on Perplexity means two back-to-back runs of one question typically share about a third of their combined sources.

Mentions and citations are not the same event #

A mention is your brand named in the answer. A citation is your URL in the source list. Perplexity can do either without the other: it can name a brand while citing a review site, a forum thread or a publication about it, and it can cite a brand's documentation without naming the brand in the prose.

The Index counts citations. Counting mentions needs a reviewed list of the names each brand goes by, and the Index's list covers only the brands its questions name. In a test read over that list, the largest platform names (Google, Microsoft, OpenAI) dominated the counts because answers name them as products and technologies, not as recommendations. That read is not published here as a rate. For your own brand, where you know your names, log both events by hand or with a tool that separates them.

Tools that track Perplexity mentions #

Several tools now watch Perplexity on a schedule. Trakkr tracks when a site is cited as a source beneath a Perplexity answer. TrackAIMentions publishes a repeatable monitoring workflow. Practitioner guides such as Rankability's recommend a fixed weekly prompt library and separate counts for mentions, citations and links.

The question to ask any tool, and any weekly spreadsheet, is how many runs stand behind each reported result. The research case for asking it is set out in "Don't Measure Once" (Schulte et al., arXiv:2604.07585), which argues that AI answers vary across runs, prompts and time, so visibility should be read as a distribution rather than a single-point outcome. The Index numbers above put a size on that for Perplexity: at one run per prompt per reporting period, about half of the sources that three runs of the same questions would have shown are missing from that period's report.

Engines also differ in which pages they cite at all. The GEO-16 study collected 1,702 citations from Brave Summary, Google AI Overviews and Perplexity on 70 product-intent prompts and audited 1,100 unique URLs, finding the engines differed in the quality of the pages they cited (GEO-16, arXiv:2509.10762). A Perplexity read does not stand in for the others; see how the AI search visibility measurement framework separates engines.

How to read a change in your Perplexity mentions #

For the metric that turns these reads into a number across engines, see Share of Citation. For which publications Perplexity and the other engines cite most often by category, see the publications AI search cites most and the Machine Relations Index.

Methodology and sources #

All Index figures come from the Machine Relations Index ledger behind the October 13, 2026 Machine Relations Index release (window 2026-05-10 to 2026-10-13). The view was rebuilt the way the release builds it: public-scoring prompts only, successful runs, the six monitored engines, engine-days with at least five successful runs but no citations excluded, and hosted-platform URLs resolved to their source domain where resolution confidence is at least 0.8. The rebuilt view matches the release's published totals of 24,496 answer runs, 186,563 citation events and 1,143 monitored prompts.

• Unit. A prompt-source pair is one monitored prompt and one domain Perplexity cited on it at least once. A pair's rate is the share of that prompt's runs that cited the domain.
• Share one run shows. For each pair, the probability that a run drawn at random from that prompt's runs cites the domain, averaged over pairs. "Two runs" draws two runs without replacement and counts the pair as seen if either cites it.
• Leader. The domain cited in the most runs of a prompt (ties broken alphabetically).
• Matched-count read. Prompts with exactly three Perplexity runs between 2026-09-14 and 2026-10-13, so every prompt contributes the same number of runs over a short span. The averages are pooled over prompt-source pairs, so a prompt with more cited domains weighs more than one with few.
• Engine access. These runs came through Perplexity's API with the sonar model, not the consumer app, which adds account, model and location settings. Perplexity ended Sonar Chat Completions support on September 27, 2026 and is reformulating those requests as Agent API requests, rolling out by model (migration overview), and on October 7, 2026 Perplexity's API began returning up to 10 sources per answer (4,225 of the 4,306 runs since carry exactly 10, 80 carry 7 to 9 and one carries 18), down from about 20, so the matched-count window spans that change. The figures describe observed run-to-run variation on the API, not a fixed-instrument experiment.
• Reads split at the October 7 cap (November 9, 2026). The table in "The same read before and after the October 7 cap" and the second engine table use the Index answer-run ledger export of 2026-11-09 08:10 UTC, successful runs only, with Perplexity run dates 2026-08-01 to 2026-10-06 and 2026-10-07 to 2026-11-09 (October 7 to November 9: 4,306 runs on 34 days, the same count the Index's Perplexity board shows since the cap; the last day is partial). Table prompts are those with exactly three or four Perplexity runs in the stated window; the engine table uses October 7 to November 9 prompts with at least five runs of the engine and at least one cited domain. Cited domains are the ledger's root domains. Hosted-platform resolution and the engine-day exclusion were not applied; the rebuild without them reproduces this page's September 14 to October 13 matched-count figures exactly (434 prompts, 1,302 runs, 12,275 pairs, 7,792 cited once, 13.73 domains per answer, 48.5% and 78.8%, 376 leaders in every run). Per-engine averages are over answers with at least one citation. The ledger's citation count per run averages 19.44 (median 20) from July 31 to October 6, 8.4 (median 9) from May 10 to July 13 and 13.7 (median 15) from July 14 to 30. The splits are descriptive and not a fixed-instrument experiment.
• Limits. These are citation counts, not prose mentions. The prompts are the market's buyer questions across the Index categories, not any one brand's. Variation is measured, not explained.

Third-party sources: arXiv:2604.07585, arXiv:2509.10762, Trakkr, TrackAIMentions, Rankability, and Perplexity's 2025 launch of free Deep Research as reported by TechCrunch.

FAQ #

How do I see my brand's mentions in Perplexity? #

Ask Perplexity the questions your buyers ask, without your brand in them, and read two places: the answer text for your brand name and the numbered sources for your URL. Run each question at least twice per reporting period. In the Machine Relations Index, one run showed about half the sources that three runs of the same questions cited, and two runs about four in five.

How many times should I run a question in Perplexity? #

At least twice per period, and more for questions that matter. Across 434 prompts with three Perplexity runs in 30 days, one run showed 48.5% of the sources cited across all three on average, and two runs showed 78.8%. The most-cited source was in all three runs on 86.6% of those prompts. After Perplexity's API began returning up to 10 sources on October 7, a four-run read on 571 prompts showed 46.8% for one run and 70.3% for two, with the most-cited source in every run on 93.0%.

Does Perplexity always cite its sources? #

Almost always. Of 5,620 Perplexity runs in the Index window, 5,619 cited at least one source, with 12.47 distinct domains per answer on average. From October 7 to November 9, after the API began returning up to 10 sources per answer (4,225 of the 4,306 runs carry exactly 10), all 4,306 runs cited at least one, at 8.63 distinct domains per answer on average (median 9).

Why was my site cited last week and not this week? #

Most likely rotation. In the 30-day read, 63.5% of the sources Perplexity cited on a prompt appeared in only one of three runs; on the 104 prompts with three runs since the October 7 cap, 54.0% did. A disappearance means more when the missing source had been cited in every earlier run.

Is Perplexity more consistent than ChatGPT or Gemini? #

In the Index sample, yes, on every stability measure in the engine table, though each engine answered a different set of prompts: Perplexity had the lowest share of sources cited only once (50.1%) and the highest median overlap between consecutive runs (0.353). It still rotates most of its sources, so a single check is not a census on any engine. From October 7 to November 9 Perplexity was still lowest on that measure: 47.0% of its sources were cited in one run, against 55.7% for ChatGPT and 61.8% for Claude. Gemini could not be compared in that window.

Can I see every Perplexity query that mentions my brand? #

No. Perplexity does not publish a query log the way Google Search Console reports impressions. You can only sample the questions you choose, which is why the number of runs per question matters.

Correction, November 8, 2026 #

The one- and two-run figures are now described as what they are: averages pooled over prompt-source pairs, not per-question shares. The overlap column is labeled with its wider cohort, the excluded engine-days rule is stated as the release applies it, and the September 27 Sonar change and the October 7 drop to at most 10 sources per Perplexity answer are noted. Every count is unchanged.

Correction, November 9, 2026 #

This page said Perplexity's answers carry more distinct sources on average than any other engine's, and tied the rotation to that breadth. That rested on a window that pools the October 7 cap, after which Perplexity returns up to 10 sources per answer. The claim is removed, and the causal link between source-list length and rotation is dropped. The 12.47 average is labeled as pooling several source-list sizes and the 13.73 average as a blend of two (419 of the 1,302 runs in the 30-day read fall on or after October 7). The November 8 correction's phrase "at most 10 sources" is superseded: the API returns up to 10. The one- and two-run read is repeated before and after the cap. The September 14 to October 13 counts are unchanged.