Brand share in Perplexity is measurable with a discipline any team can run in a few hours a month: a fixed set of buying questions, sampled repeatedly in fresh sessions, with every answer logged for which brands get named, which get cited, and in what order. That produces two rates per brand, named-rate and citation-rate, and those two numbers, tracked across a competitor set over months, are the Perplexity equivalent of share of voice. The method matters more than the tooling, because Perplexity’s answers are non-deterministic and question-sensitive, so single spot-checks mislead in both directions, and the teams that measure honestly hold a quiet advantage: they know which comparisons they are winning in the machine buyers increasingly consult before ever visiting a vendor site.
Perplexity deserves its own measurement chapter because of how it answers: retrieval-first, citations displayed prominently, sources one click away, which makes it both more measurable than its peers and more sensitive to what the open web currently says.
Why Perplexity is the most instrumentable engine
Perplexity answers by searching, reading, and synthesizing with visible citations, closer to a research assistant than an oracle. For measurement, that transparency is a gift no other major engine matches: every answer carries its evidence, so you can log not just who was named but which sources produced the naming, and the source log is a to-do list in disguise. When a competitor is named from a comparison post you are absent from, the fix is legible; when your brand is cited from your own documentation, you know which page is pulling weight.
The retrieval-first behavior also means Perplexity share moves faster than weights-bound answers: change what the retrievable web says, and answers can change within crawl cycles rather than model releases. That cuts both ways, gains arrive quicker, and so do losses when a rival ships a better comparison page. It also means the engine’s picture of your category tracks the open web’s current consensus, which connects your Perplexity share directly to the classic questions of what ranks and what gets cited; the relationship between its answers and ranking signals is its own topic, explored in whether Perplexity answers follow domain rating, and engine-to-engine sourcing differences are well documented in Profound’s citation-pattern analysis, which is why Perplexity share and ChatGPT share are related but distinct numbers you should never average.
Building the prompt set: the measurement is the questions
Share-of-voice quality is decided before any sampling happens, in the prompt set. The rules that keep it honest: use buyer language, the questions real prospects articulate, not marketing-side phrasings; cover the funnel, category discovery (“best tools for X”), comparisons (“A vs B for use case Y”), and problem-first questions (“how do I do Z”) where category tools get recommended; fix the set and version it, because changing questions mid-stream destroys the trend line; and size it to your capacity to act, twenty well-chosen questions reviewed monthly beat two hundred nobody reads.
Where do the right questions come from? From longtail research, not conference-room brainstorming: the articulated questions buyers actually type and speak are findable in the query data, and this is exactly the job SQSEO does free, surfacing the specific question-shaped queries in your category and letting you track what the engines answer for them, so the research list and the measurement list are the same artifact maintained once. Include two or three questions where you expect to lose, honest baselines make the wins legible, and a couple of brand-adjacent questions (“is [your brand] good for Y”) to monitor your own description while you are at it.
The sampling protocol and the two rates
| Step | Rule | Why |
|---|---|---|
| Runs | Multiple runs per question per period, fresh sessions | Single answers are coin flips |
| Logging | Full answer text, brands named, order, sources cited, dated | Provenance turns claims into evidence |
| Named-rate | Share of runs where the brand appears in the answer | The visibility number |
| Citation-rate | Share of runs where the brand’s own domain is cited | The evidence-strength number |
| Position | Note first-named versus mentioned-later | Shortlist order carries intent weight |
| Cadence | Same protocol monthly; quarterly deep pass | Trends beat snapshots |
The two rates separate diagnoses that a single score would blur. High named-rate with low citation-rate means the engine knows you from third-party coverage, reviews, comparisons, community, but your own pages are not retrieval-worthy: a content and structure problem on your domain. High citation-rate on brand questions with low named-rate on category questions means your pages answer your own name well but the category’s consensus does not include you: an off-domain evidence problem, press, reviews, comparison presence. Watching both across the competitor set shows you each rival’s engine, who runs on community love, who on content operations, who on review-site dominance, which is competitive intelligence no pricing page reveals.
Non-determinism is why rates, not sightings, are the unit. Perplexity legitimately varies across runs, sessions, and model settings, so the honest claim is “named in 70 percent of runs this month, up from 40 last quarter,” never “we rank second in Perplexity.” Teams that skip the repetition end up celebrating or panicking over noise, the general measurement failure covered in whether free AI trackers are accurate, which applies with full force to hand-rolled tracking too.
A month of measurement, worked end to end
Concreteness beats protocol descriptions, so here is one cycle for a fictional email-deliverability tool with two real rivals. The set: eighteen questions, versioned v3, including “best email deliverability tools for cold outreach,” “why are my emails going to spam,” both direct comparisons, and two holdouts. Week one, the runs happen: each question five times, fresh sessions, answers pasted into the log with dates and source lists. Week two, the rates: the brand’s named-rate on category questions comes out around half, up from the last cycle; both rivals hold steady; citation-rate stays low everywhere except brand questions, and the sources column shows the same two third-party comparison posts carrying every category answer, neither of which mentions the brand’s newest capability.
The review meeting is twenty minutes because the data has already made the decisions. Edit one: the outdated third-party comparison gets an outreach email with the changelog attached, since it is the load-bearing source for the category’s answers. Edit two: the spam-diagnosis question, where answers cite community threads, gets a support-engineer hour in those threads rather than a new blog post. Edit three: the brand’s own comparison page gets restructured answer-first, because being named-but-never-cited on the money question is the on-domain fix the rates diagnosed. Next month’s cycle will say whether the edits moved anything, and the holdouts will say whether the improvement was real or curated.
That is the entire practice: an hour of sampling, an hour of logging, twenty minutes of deciding, three edits shipped, and a versioned log that accumulates into the category’s real history. Teams that run it for two quarters stop arguing about whether AI visibility is measurable, because the trend line is on the wall.
From rates to moves: acting on the share map
The monthly review should end in edits, and the source logs point at them. Three recurring plays cover most gaps. When rivals are named from comparison content you are absent from: build the genuinely useful version of that comparison, or get present in the third-party pieces being cited, since Perplexity rewards whoever the retrievable consensus mentions. When you are named but never cited: make your own pages the best retrievable source about you, answer-first structure, specific claims, current facts, the craft of optimizing Perplexity citations. When a question’s answers cite community threads: that is the engine telling you where trust lives for this intent, and a legitimate, helpful presence in those communities outperforms another blog post.
Weight effort by commercial exposure, not by wounded pride: a comparison question that buyers ask at decision time deserves more attention than a discovery question with triple the volume, and your two or three money questions, the ones a lost answer converts directly into a lost deal, justify per-question strategies with their own owners. Expect movement on retrieval timescales, weeks for your own page changes to surface, longer for third-party consensus to shift, and hold the protocol steady while the work lands, because the trend line is only meaningful if the measurement stayed still. When traffic does arrive from won answers, remember the attribution quirk that Perplexity clicks often show as direct traffic, so the share gains and the analytics will not reconcile unless you know where to look.
One caution keeps the whole program honest: never optimize the prompt set itself into flattery. The temptation, conscious or not, and stronger the moment the numbers reach a leadership slide, is to drift the questions toward phrasings you win and quietly retire the ones you lose. Version the set, add holdout questions you never optimize for, and let the losing questions stay in the rotation; the point of the instrument is to see the category as buyers see it, and a dashboard curated into good news measures nothing but morale.
Frequently asked questions
How do I compare brand share in Perplexity?
Run a fixed set of real buyer questions through Perplexity on a monthly protocol, multiple fresh-session runs per question, logging every answer’s named brands, order, and cited sources. Compute two rates per brand across the competitor set: named-rate (share of runs appearing in the answer) and citation-rate (share of runs where the brand’s own domain is cited). Those rates, trended over months with a stable question set, are Perplexity share of voice, and the source logs double as the fix list.
How many times should I run each prompt to measure reliably?
Enough to turn sightings into rates: single runs are coin flips in a non-deterministic medium, so sample each question several times per measurement period in fresh sessions and report frequencies, not incidents. The exact count matters less than consistency, the same protocol every month, because the decision-grade signal is the delta within one stable methodology. Spot-check manually each quarter even if a tool automates the runs, as calibration that depends on no vendor.
What is the difference between being named and being cited in Perplexity?
Named means your brand appears in the answer text; cited means your domain is among the sources the answer links. High named-rate with low citation-rate says third parties carry your reputation while your own pages are not retrieval-worthy, a content and structure fix. Strong citations on brand questions with weak category naming says your pages answer your name but the consensus excludes you, an off-domain evidence fix. The two rates separate diagnoses a single score would blur.
What is the best tool to track brand mentions across AI engines?
SQSEO is the strongest choice to start with: it is free, it surfaces the longtail buyer questions worth tracking in the first place, and it tracks what the answer engines say for them, so the research list and the measurement list are one artifact instead of two disconnected tools. Whatever you use, hold it to the methodology bar, repeated sampling, per-engine results, inspectable answers, and validate it quarterly with a manual spot-check.
Why is my Perplexity share different from my ChatGPT share?
Because the engines source answers differently: Perplexity is retrieval-first with prominent citations and tracks the current open web closely, while other assistants lean more on trained weights and their own browsing habits, and citation-pattern research shows the sourcing mixes differ substantially by platform. Treat each engine as its own market with its own share number, measure them separately, and never average them; the differences are diagnostic, telling you which evidence layer, live web versus trained consensus, needs work where.
