How I Check a Perplexity Mention

I check a Perplexity mention by running a fixed set of buyer questions and writing down two different events. One column is whether our name appears in the answer. The other is whether a URL of ours, or a URL about us, appears in the numbered sources. Perplexity has no Search Console and no native report of where a brand appeared this week. A one-off look is a sample of one. The weekly log is the measurement, and the repair is whatever source the log shows beating us.

The repeated-check method, with the collection rules I would want any tool to meet, is How to See Mentions in Perplexity Across Repeated Checks. That page is the deeper resource, the method. This essay is how I would use it as a founder: the columns I would keep, the two instrument results I would trust over a vendor blog, the capabilities I would demand before I paid, and the finding that would make me stop logging and start publishing.

Three appearances, three columns

I would not let a weekly total blend these, because they diagnose different holes.

An answer-text mention means the brand is named in the generated response. A source citation means a URL tied to the brand, or about the brand, sits in the numbered references. A comparative mention means the brand appears beside competitors in a category answer. A brand can be cited and unnamed. A brand can be named and uncited. The strongest cell is both.

What I log	The question on the row	What a yes is allowed to mean
Answer text	Is the brand named in the body?	The synthesis used the name.
Source citation	Does our domain, or a page about us, appear in the references?	Some authority source entered retrieval.
Comparative	Are we listed with the competitors a buyer would recognize?	We are inside that answer's consideration set.

A blended "mention count" destroys the diagnosis. I would rather have a short log with three columns than a dashboard with one score.

What I believe about the retrieval, and what I refuse to pretend

Nobody outside Perplexity has its ranking code. I would not hire anyone who claims to. What I will use is the public shape of retrieval-augmented systems and the behavior of the product from the outside.

These systems pull a wide candidate set first, with keyword matching beside dense vector retrieval, because that stage wants recall. The standard reference I would hand an engineer who wants a paper rather than a metaphor is Dense Passage Retrieval for Open-Domain Question Answering. A reranker then scores query-document pairs together, and shallow keyword matches fall out. The reference I would hand for that stage is Passage Re-ranking with BERT. A last selection produces the short, varied citation list on the screen.

For a founder the stage count matters less than the shape. Hundreds of candidates go in and a handful of citations come out. A slot we win is a slot someone else loses. The cut is on evidence the model can check, not on how many words we shipped. I would not brief the team on a secret three-layer formula. I would brief them on a scarce slot and a checkable page.

The two numbers I would rather use than a vendor's

Most Perplexity statistics in circulation come from companies that sell visibility tools and grade their own samples with unpublished methods. I would not build a plan on those. Two findings from the Answer-Source Fidelity instrument behind the Machine Relations Index are built the other way: they grade claim-and-citation pairs and publish coverage next to the rate.

A citation usually does not become a mention. Across 14,447 citations the instrument could read in full, from four answer engines, the cited source's own name reached the words of the answer 8.6% of the time. Perplexity was the lowest of the four: 440 of 7,501 read citations, or 5.9%, because it credits sources with numbered markers rather than naming them in the prose. On this engine, being cited and being named are close to separate outcomes. That is why the columns stay separate, and why a blended count is noise. One time in roughly seventeen is the coincidence rate I would remember: 440 divided by 7,501.

Half of what Perplexity cites does not support the claim it sits beside. On the frozen baseline, release 8509d376, generated 2026-09-28, 50.1% of the 1,298 resolvable Perplexity citations were Supported, and resolvable coverage was 47.6%, which is 1,298 of 2,726. I would not brief that as a unique indictment. Independent academic work audited generative search engines, Perplexity among them, and found 51.5% of generated sentences fully supported by their citations (Liu, Nelson, and Liang, Evaluating Verifiability in Generative Search Engines). Two methods, the same order of magnitude.

Read together, the slot is scarce, it usually will not carry our name, and it is awarded to a page that can be checked against the specific claim. A page that states a claim and makes the evidence legible earns slots. A page that only asserts is a coin flip. I would rather publish the first kind of page than argue with the 5.9%.

The manual log I would start this week

I would open Perplexity and type a real buyer question. For each of 15 to 20 queries I would log the exact question, whether we were named (yes, no, or indirect), whether we were cited and at which URL, which competitors were named, which competitor URLs were cited, and whether we were recommended, merely listed, compared, or absent.

The set would mix category discovery ("best tools in the category"), comparisons ("alternatives to the incumbent"), problem questions ("how do I solve this"), and procurement questions ("what should I evaluate"). The manual pass is free, and it forces someone to read the answer, which is where the repair lives. It does not scale to 20 queries across Perplexity, ChatGPT, Gemini, and Google AI Mode every week. That is the moment I would buy software, and not before I knew what the columns had to contain.

What I would require of a tool

The monitoring market is crowded enough that a short list from me would be a recommendation I cannot stand behind for someone else's queries. I would buy on five capabilities, and I would walk away from a trial that cannot show them on our questions.

Answer-text mentions and source citations in separate fields. The 5.9% figure is the reason. A tool that blends them into one visibility score is destroying the signal.

The raw answer, stored and exportable. If I cannot read the response that produced a point, I cannot audit the tool or see why we lost.

Competitor citations at URL level, not only brand names. The useful question is which page beat ours.

Our query set, unmodified. If the product substitutes its own keyword list, we are measuring its benchmark.

A stated collection method: which model, which date, how many runs. Answers vary between runs. One sample per week per query is a headline.

No dashboard explains the absence. The explanation is the source the answer preferred, and that is a publishing and corroboration problem.

The three repairs I would look for in the log

Entity consistency. If the homepage says one category, the company profile says another, and the press says a third, the retrieval stack has three signals. The clearer entity wins the name. I would state the category in prose and in markup. The Organization type carries name, description, sameAs, and knowsAbout. Filling those consistently, everywhere we already appear, is an afternoon. I would do the afternoon before I bought another content sprint.

Third-party corroboration. A brand describing itself is weak retrieval evidence. A publication, an analyst, or a research domain describing the brand is stronger, because it lowers the chance the model is repeating an unchecked claim. I would rather have three outside pages saying the same clear thing than five outside pages saying five different things.

Recency with substance. Answer engines favor pages that were published or updated recently, which is why a dated page loses ground it once held. I would not attach a percentage to that preference. The percentages in circulation come from vendor blogs grading their own samples, and I have not measured a percentage I am willing to publish. The operating rule survives without one. Stale pages lose. Fresh thin pages do not automatically win. A comprehensive page updated last month with a real figure beats a weekly post that says nothing. Changing the year in a heading is not an update. Replacing a stale figure with a current one is.

When the log shows an absence, the cause is almost always one of those three. I would not start the repair by changing a prompt.

What I would not import from rank tracking

Rank tracking trained teams to ask where we sit for a keyword, on a page where position and traffic moved together. An answer engine asks a different question: which evidence is strong enough to name. A page can rank first in ordinary search and never be named in Perplexity if it is thin, self-referential, and unclear as an entity. I have no interest in pretending those are the same game.

The directional evidence I would allow into the memo, labeled as Google and not as Perplexity, is Pew's tracking of 900 US adults. Users who saw a Google AI summary clicked a result link on 8% of visits, against 15% for users who did not see one, and clicked a link inside the summary on 1% of visits. I would not transfer 8, 15, and 1 onto Perplexity. I would keep the structural point: when an answer sits in front of the results, the click we used to compete for often never happens, and what the buyer keeps is the sentence the answer said.

Google's own developer docs say there are no additional technical requirements for its AI features beyond being indexable with a snippet, and that specific optimization is not required for AI Overviews and AI Mode while the ordinary fundamentals still matter. Perplexity publishes less. The behavior I would plan for is the same shape: evidence and clarity, not a format trick. I would still log Perplexity separately, because its 5.9% name-in-the-answer rate is its own fact.

The weekly pass

Step	What I would do	What I would keep
1	Freeze 15 to 20 buyer queries	A stable set, written down
2	Run each one and log name, citation, and competitors apart	The raw rows
3	Compute a mention rate and a citation rate as separate percentages	Two rates
4	Put competitor rates next to ours	The gap, in the same units
5	For every absence, record the URL that was cited and why it was easier to check	The repair list

Step 5 is the work. An absence without the winning URL teaches nothing. The winning URL is the brief for the next placement or the next edit.

Category language has to match the language buyers use. A clever name for the category that buyers do not type is a translation the retrieval stack will not do for us. Repetition across trusted sources matters because a reranker trusts a pattern: three publications saying the same clear thing beat five publications saying five things. Freshness matters in the substance of the figure, not in the copyright line.

The reason a third-party sentence outperforms a page about ourselves is the mechanism I mean by Machine Relations. The term for the outcome the two columns are estimating, so a vendor cannot rename one of them, is AI visibility.

How I would read one week of the log

I would not wait for a quarter of rows before the log changed someone's week. Fifteen to twenty queries, read once, already separate three repairs, and I would assign the week from the pattern rather than from the scariest blank.

If we are cited and not named, I would not panic and I would not celebrate. The fidelity instrument says that is the ordinary Perplexity outcome: 440 of 7,501 read citations, 5.9%, carried the source's name into the prose, against 8.6% across 14,447 citations from four engines. A citation without a name means a URL entered retrieval and the synthesis did not pick the name up. The repair is a sentence on that URL, or on the page it supports, that states the claim the answer is making, so the next selection has a reason to say the name. It is not a reason to buy distribution of a page the answer already found.

If we are named and not cited, I would find out which numbered source the name is leaning on. A name with no URL of ours and no URL about us is a synthesis that may be borrowing a competitor's source or a stale training trace. I would not count it as a win in the citation column. The supported-claim rate is why: on release 8509d376, 50.1% of 1,298 resolvable Perplexity citations were Supported, coverage 47.6% of 2,726. An unsupported name is the coin flip. I would rather add a checkable page than frame a screenshot of the name.

If a competitor is named and cited, step 5 of the weekly pass is the whole assignment. I write down their URL, I write down the claim it makes in the first screen, and I write down whether our page says the same thing with a source or says something vaguer. Entity mismatch is an afternoon with Organization markup and one category sentence. A missing third party is a placement. A stale figure is an edit. I would not open a new content series until I had named which of the three the winning URL actually won on.

If the same query flips between runs, I would not let one run set the repair. The reason I demand a stated run count from any tool is that a single sample is a headline. Two weeks of the same 15 to 20 questions is the minimum I would let into a planning meeting. A tool that stores the raw answer lets me see whether the flip is a different source or the same source worded differently. A tool that stores only a score hides that, which is the failure mode the 5.9% figure exists to prevent.

I would also keep the architecture papers in the appendix of the log, not in the repair. Dense passage retrieval and BERT re-ranking describe why a handful of citations fall out of a wide candidate set. They do not tell me which of our pages lost. The log tells me that. Liu, Nelson, and Liang's 51.5% fully supported sentences tell me the bar is support, and our own 50.1% tells me Perplexity sits on that bar rather than above it. I would brief both so nobody treats a footnote as proof the claim is true.

Pew stays in a different appendix. Nine hundred US adults, 8% result clicks with a Google AI summary against 15% without one, and 1% clicking inside the summary, is an argument that the sentence in the answer is what the buyer keeps. It is not a Perplexity click model, and I would correct anyone who pasted it into the Perplexity column. Google's developer docs, which decline to offer a special optimization for AI Overviews and AI Mode beyond an indexable snippet, are the reason I would not add a format hack to the repair list. Perplexity publishes less. The behavior I have is the separate columns and the support rate. That is enough to spend against.

The decision

I would run the 15-to-20 query log by hand for two weeks before I bought a tool, then buy only a tool that passes the five tests on those same queries. I would repair entity language first when the columns disagree across our own surfaces, third-party corroboration when the winning URLs are all someone else's publications, and freshness when our page is the right URL and the figure on it is stale. I would not repair by adding prompts, and I would not repair by rewriting for a trick the Google docs already decline to offer.

Scope: Perplexity mention tracking for a company that sells into questions a buyer actually asks. The instrument numbers are release 8509d376 and the 14,447-citation read. The Pew figures stay labeled as Google. The two arXiv papers are the public architecture, not Perplexity's source code. This does not authorize a claim about Perplexity's undisclosed ranker, and it does not transfer the 5.9% onto ChatGPT.

What would change my mind: a later fidelity release in which Perplexity's name-in-answer rate rose near the other engines and the 5.9% separation disappeared, because that is the reason the columns must stay apart. I would also collapse to a single weekly sample if a disclosed variance study showed Perplexity answers stable enough that one run per query per week matched seven. I would not collapse the columns because a tool's default score is more convenient to chart. Convenience is how the 440-of-7,501 fact gets thrown away.

FAQ

How do I see whether Perplexity mentioned us?

I run 15 to 20 buyer questions every week and log, for each, whether the name is in the answer, whether a URL is in the sources, and which competitors occupied either column. There is no native dashboard. A monitoring product is acceptable when it stores the raw answer and keeps the two events in separate fields, on our unmodified queries, with the model, the date, and the number of runs stated.

What is the difference between a citation and a mention here?

A citation is a numbered source URL tied to the brand or about the brand. A mention is the name in the generated prose. On the fidelity instrument, Perplexity named the cited source in the answer for 440 of 7,501 read citations, 5.9%, against 8.6% across 14,447 citations from four engines. The strongest row is a yes in both columns. A single score that adds them together is the metric I would not pay for.

Why can a competitor be named when our Google rank is better?

Perplexity is assembling evidence. Entity clarity, third-party corroboration, and a page that can be checked against the claim decide the slot. A thinner, self-referential page can rank in ordinary search and still lose the answer. I would log the competitor's URL and ask why that page was easier to verify. I would not ask how to "rank" inside Perplexity as if it were a results page.

What do I fix first if the name never appears?

Entity consistency, if our own surfaces disagree: one category sentence on the homepage, the company profile, and the coverage we already have, including Organization markup. Then third-party corroboration, if the winning sources are publications describing someone else. Then a real update to the figure on the page we want cited. I would not start with a content trick. Google's AI-feature docs say ordinary indexable pages are the requirement on its surfaces, and I would not invent a stricter Perplexity hack to fill the silence.

Which outside statistic would I refuse to paste onto Perplexity?

Pew's. Among 900 US adults, an AI summary on Google cut result clicks from 15% of visits to 8%, and clicks inside the summary were 1%. I cite it for the structural point that an answer absorbs the session. I do not cite it as Perplexity's click rate. I also refuse any freshness percentage that comes from a vendor grading its own sample. The rule I will operate without a percentage is that a substantive update beats a cosmetic one.