Perplexity Signals
What leaders and teams rolling out Perplexity report, with every finding tied to its evidence.
Read the findingsWhat the evidence says
Practitioner sentiment, leaders and teams
−25 Negative from 20 sources that take a side
How this figure is madeNegative overall signal, net 25 points negative
Each from a different author and a different discussion. Small sample. Run a pilot on questions your team already asks to measure how often the answer is right and cites its source.
- Analyzed
- 89 total sources
- Top praise
- Cited answers (5 of 9 sources) See praise evidence
- Top concern
- Accuracy (5 of 6 sources) See concern evidence
Jan 2025 to Sep 2026.
73 sources from practitioners, by month and sentiment.
Research teams that audited Perplexity's answers report wrong answers and weak sources, while benchmark studies credit its citation accuracy. Browser security teams report that Comet's agent followed hidden instructions and phishing pages. Campus buyers split: Texas A&M approved Enterprise Pro for FERPA data, and Texas State ended its trial.
Positive, mixed and negative sources
The score counts each source once, as the split does. See How the numbers work.
See the evidenceBelow: 1 praise theme and 2 concerns.
Findings
Counted for Leaders and teams rolling it out, from 25 sources. Change group
What practitioners praise
Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.
Cited answers5 of 9 sources on it praise itInline citations and source links that let readers check each claim.
Counted for Leaders and teams rolling it out.
Positive, 5
Positive
Corvinus University of Budapest, university-wide Enterprise Pro trial
Perplexity Enterprise Pro trial at Corvinus: 16 March - 17 April 2026University news page
Corvinus opened Enterprise Pro to its students and staff for a trial from March 16 to April 17, 2026. The university presents it as an assistant for literature search and academic writing that answers with source citations drawn from several AI models.
Mixed
Ethan Smith, CEO of Graphite
Lenny's Podcast: Ethan Smith on AI search visibilityPodcast episode, Lenny's Podcast
Graphite's study of thousands of questions found Perplexity's citations overlap Google's about 70% of the time, against about 35% for ChatGPT.
"Perplexity was around 70%, but essentially they're all similar algorithms, but with very different citations and results."
Transcript excerpt, Ethan Smith, at 37:02 Mixed
Mingxuan Du and coauthors, University of Science and Technology of China and Metastone Technology
DeepResearch Bench: A Comprehensive Benchmark for Deep Research AgentsIndependent benchmark, arXiv
Perplexity Deep Research had the highest citation accuracy of four deep research agents at 90.24%, while its overall report score of 42.25 trailed Gemini at 48.88 and OpenAI at 46.98.
Positive
Texas A&M Technology Services, the university's central IT organization
Perplexity - Enterprise Pro Frequently Asked QuestionsUniversity IT service page
Texas A&M offers Enterprise Pro campus-wide at US$108 a year per license, approves it for public, internal, and FERPA data, and requires approval before HIPAA data.
Mixed
Tianze Xu and coauthors, Shanghai Jiao Tong University
ResearcherBench: Evaluating Deep AI Research Systems on the Frontiers of Scientific InquiryIndependent benchmark, arXiv
Perplexity Deep Research scored 0.48 on the rubric against 0.70 for OpenAI Deep Research, which the authors call moderate performance. Sonar Reasoning Pro had the highest groundedness score of any system tested, at 0.68.
Mixed or negative, 4
Negative
Haus Research, an independent research publisher
A third of Perplexity's citations don't contain the number they're cited forIndependent benchmark, Haus Research
Haus Research asked Sonar and Sonar Pro 310 questions about 210 technology companies. Of 1,826 citations behind a figure, 34.7% pointed at a page that would not open or held none of the sentence's figures.
Negative
Ilan Strauss, Jangho Yang, Tim O'Reilly, Sruly Rosenblat, and Isobel Moure, paper authors
The Attribution Crisis in LLM Search ResultsResearch paper, arXiv
Across about 14,000 LMArena conversations, Perplexity's Sonar visited about 10 relevant pages per query and cited three to four of them.
Negative
Klaudia Jaźwińska and Aisvarya Chandrasekar, Tow Center for Digital Journalism, Columbia Journalism School
AI Search Has a Citation ProblemIndependent benchmark, Columbia Journalism Review
The Tow Center ran 1,600 queries across eight AI search tools. Perplexity answered 37% of its queries incorrectly, against more than 60% across all eight, and Perplexity Pro cited syndicated copies of Texas Tribune articles in 3 of 10 queries.
Negative
Pete Archer, Programme Director for Generative AI, BBC
Groundbreaking BBC research shows issues with over half the answers from Artificial Intelligence (AI) assistantsResearch announcement, BBC Media Centre
BBC journalists reviewed answers from four AI assistants, Perplexity among them, and judged 51% of all answers to have significant issues. One Perplexity answer cited the BBC for adjectives about the Middle East conflict that BBC reporting never used.
What concerns them
Every concern practitioners raise, most raised first.
Accuracy5 of 6 sources on it raise it as a concernWrong answers and citations that do not support the sentences they mark.
Counted for Leaders and teams rolling it out.
Raised as a concern, 5
Negative
Haus Research, an independent research publisher
A third of Perplexity's citations don't contain the number they're cited forIndependent benchmark, Haus Research
Haus Research asked Sonar and Sonar Pro 310 questions about 210 technology companies. Of 1,826 citations behind a figure, 34.7% pointed at a page that would not open or held none of the sentence's figures.
Negative
Boyu Gou and coauthors, The Ohio State University
Mind2Web 2: Evaluating Agentic Search with Agent-as-a-JudgeIndependent benchmark, arXiv
Perplexity Pro Search fully solved 8% of the evaluated tasks and Perplexity Deep Research 15%, against 28% for OpenAI Deep Research. The authors found Pro Search struggled to synthesize accurate answers from many sources.
Negative
Klaudia Jaźwińska and Aisvarya Chandrasekar, Tow Center for Digital Journalism, Columbia Journalism School
AI Search Has a Citation ProblemIndependent benchmark, Columbia Journalism Review
The Tow Center ran 1,600 queries across eight AI search tools. Perplexity answered 37% of its queries incorrectly, against more than 60% across all eight, and Perplexity Pro cited syndicated copies of Texas Tribune articles in 3 of 10 queries.
Negative
Pete Archer, Programme Director for Generative AI, BBC
Groundbreaking BBC research shows issues with over half the answers from Artificial Intelligence (AI) assistantsResearch announcement, BBC Media Centre
BBC journalists reviewed answers from four AI assistants, Perplexity among them, and judged 51% of all answers to have significant issues. One Perplexity answer cited the BBC for adjectives about the Middle East conflict that BBC reporting never used.
Negative
Trellner Research, an independent research publisher
Three sites made 215,128 "best software" pages for AI. Perplexity cites themIndependent benchmark, Trellner
Trellner put 380 software categories to Sonar and Sonar Pro; 59.8% of 7,534 citations pointed at domains outside the 100,000 most-visited sites, and three related sites had published 215,128 generated best-software pages.
Discussed without concern, 1
Mixed
Jiayu Wang and coauthors, University of Wisconsin-Madison and Salesforce
LiveResearchBench: A Live Benchmark for User-Centric Deep Research in the WildIndependent benchmark, arXiv
Sonar Reasoning and Sonar Reasoning Pro scored just behind Gemini 2.5 Pro on factual and logical consistency, which the authors credit to concise reports. Both scored lowest of the single-agent web models on coverage.
GILD's fact checks, which don't count as sources
Checked Oct 12, 2026.Perplexity's documentation says a source label describes a whole website. It never vouches for a single claim or replaces reading the source. Model Council, on Max and Enterprise Max, runs one question across three models and shows where they disagree. Understanding source labels
Open them in the recordData & security4 of 5 sources on it raise it as a concernWhat happens to queries and files, whether work use is allowed, and agent and browser safety.
Counted for Leaders and teams rolling it out.
Raised as a concern, 4
Negative
Shaked Chen, Security Researcher, Guardio
"AgenticBlabbering" How AI Browsers' Verbose Reasoning Fuels the Ultimate Scamming MachineCompany research blog, Guardio Labs
Guardio captured the reasoning and tool calls Comet's agent streams while it browses, then used them to tune a phishing page offline until the guardrails stopped objecting. Comet fell for the tuned page after under four minutes of training.
Negative
Aviad Gispan, Senior Researcher, LayerX
Is Perplexity Comet Safe? LayerX Finds a Prompt Injection Attack VectorCompany research blog, LayerX
LayerX showed that a single crafted link could make Comet send email and meeting data it had seen to an attacker, with no malicious page content needed. LayerX reports that Perplexity replied it could find no security impact.
Negative
Paloma Perlov, Researcher, LayerX
LayerX Finds that Perplexity's Comet Browser is Up To 85% More Vulnerable to Phishing and Web Attacks Than ChromeCompany research blog, LayerX
LayerX ran the 100 most recent reported phishing attacks against five browsers. Comet stopped 7% of them out of the box, against 47% for Chrome and 54% for Edge.
Negative
Nati Tal, Head of Guardio Labs, Guardio
"Scamlexity" We Put Agentic AI Browsers to the Test - They Clicked, They Paid, They FailedCompany research blog, Guardio Labs
Guardio Labs made Comet its main test subject. In some runs Comet checked out on a fake Walmart store with the user's payment details. It also opened a live phishing link from a fake bank email and helped fill in the login.
Discussed without concern, 1
Positive
Texas A&M Technology Services, the university's central IT organization
Perplexity - Enterprise Pro Frequently Asked QuestionsUniversity IT service page
Texas A&M offers Enterprise Pro campus-wide at US$108 a year per license, approves it for public, internal, and FERPA data, and requires approval before HIPAA data.
GILD's fact checks, which don't count as sources
Checked Oct 12, 2026.Enterprise data is never used to train Perplexity's models, and agreements bar third-party providers from training on it. Files attached to sessions are deleted after seven days. Consumer Pro and Max users can opt out of training data collection. Data retention and privacy for Enterprise organizations
Open them in the record
Where the evidence agrees
Statements this group's sources agree on, then the ones close to agreement.
For leaders and teams rolling it out, no statement reaches 5 sources from 3 authors.
Where it disagrees
Questions the sources split on, with the count on each side.
For leaders and teams rolling it out, no question reaches 5 sources from 3 authors, with 2 on each side.
The evidence
Whose experience this is
The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.
Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on Leaders and teams rolling it out.
Leaders and teams rolling it out
25 sources. These accounts come from institutions that rolled out Perplexity and from research and security teams that tested it.
Individual users
48 sources. These accounts weigh speed, citations, accuracy, and plan value for people using Perplexity for their own work.
All users
73 sources. Most of these accounts come from individual users.
73 sources from practitioners, by the context they work in. Filled rows count for this group.
- Individual userOne person using it for their own work48 sources
Emphasizes Speed, citations, accuracy, and whether the paid plan earns its price
- Small team2 to 10 people sharing it1 sources, counted here
Emphasizes GILD found no public small-team account; ask Perplexity for a reference customer of your size
- DepartmentA department, or a named company's team16 sources, counted here
Emphasizes Research teams audit citations and benchmark reports, and browser security teams test Comet's agent. One agency runs a public Perplexity Room
- IT & leadershipIT and business leaders deciding the rollout3 sources, counted here
Emphasizes Security leaders at Brave and Guardio report that Comet's agent followed hidden instructions and scam pages, and the BBC's generative AI lead reports distorted news answers
- Organization-wideCompany-wide rollout5 sources, counted here
Emphasizes Campus-wide licenses and trials with approved data classes, and one campus trial that ended in a move to Gemini
What changes with scale. Individual users talk most about speed, citations, and accuracy, and many have moved some research back to a general assistant. Institutional accounts focus on who gets access and which data may go in, and research teams test whether citations support their sentences.
Where the sources come from
Where this group's sources were published.
Counted for Leaders and teams rolling it out, from 25 sources. Change group
41 sources for this group, 16 of them documentation
- Business blogs & reviews23, 56% of 41 sources
- Product documentation16, 39% of 41 sources
- Podcasts, Talks & Videos2, 5% of 41 sources
Older evidence stays when it still applies. Sources that depend on a retired model lineup, an earlier version of Deep Research, or a renamed feature stay in the history, marked Needs a recheck.
See sources by originThe record
Who each source is, how it knows, whether it still applies and what it counts toward.
Newest first, on its own page. Every count above opens its sources there.
Read all 89 sourcesAbout this read
What changed, and how we count
What changed
- Nov 1, 2026: since Oct 30 the page prints one vote per author, and a figure given on the weighted reading is the second view. Leaders and teams read −25, individual users +2 and all users −6.
- Oct 30, 2026: a second-model check removed a Business Brain host's airport test, a personal question with no buying stake, and the Changelog hosts' personal research errands, from its Operator Insight and the ChatGPT comparison. It removed Texas State's Operator Insight, since the IT page gives no start or publication date; the Texas State record now says the university encouraged users to move to Gemini, as the page states. On the weighted reading, individual builders move to +2 on 45 counted sources, and all practitioners to -12 on 65.
- The review removed 17 sources from the first build that did not meet the source standard, among them 11 pages published before 2025 and two institutional accounts that had been Operator Insights.
- A finding shows only when at least 5 sources from 3 authors discuss it. A disagreement needs at least 2 sources on each side.
- The review dropped three sets that fell short of those counts. Company Knowledge in Projects had 4 sources as a praise theme, Vendor Conduct had 2 as a concern, and a statement on group rollouts had 2.
- GILD checked each documentation page against Perplexity's pricing page, Help Center, and changelog.
- Oct 25, 2026: added the Feature gaps and Speed & reliability themes from what existing accounts state. Of those, 3 raise features they could not get, 4 raise slow answers or failed runs, and 1 finds it slower than Google but a time saver overall.
- Oct 25, 2026: adds 9 sources from independent citation audits, an arXiv paper, university pages, GitHub issues, and a video episode.
- Oct 26, 2026: adds browser security tests of Comet from Brave, Guardio, and LayerX and research benchmarks of Sonar and Deep Research. The BBC's news study and a university trial join them. Citation audits whose reports state a verdict now carry that sentiment, and one agency account now records the feature gaps it names.
- Oct 28, 2026: adds 6 sources from GILD's research ledger, each with the passage it quotes: podcast episodes from Graphite, Business Brain, Changelog and Tiny, and Cloudflare's report that Perplexity crawled sites that had blocked it. On the weighted reading, leaders and teams now reach 20 sources and print a score, Negative at -34; individual users move from 0 to +2 and all practitioners from -14 to -12.
Method
A running review of what people using Perplexity at work report, from individual users to company-wide rollouts. Each finding links to its sources, and each conflict shows both sides.
Counts cover only the sources GILD reviewed.
Who counts
- Leaders and teams rolling it out
- Small team, Department, IT & leadership, Organization-wide
- Individual users
- Individual user
- All users
- Everyone in the leaders and teams and individual users groups
Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.
How the numbers work
- Overall signal. Positive votes minus negative ones, per 100 votes from the sources that take a side. Each author counts once, with their latest view, whoever they are and wherever they posted. The older weighted reading, where accounts GILD matched to a named person count three times when GILD lists the author among IT and business leaders and twice otherwise, is shown beside it as a second view. Where an author wrote more than one source, the line beside the split gives the votes, one per author, so the score follows from it. When one author's sources from the same day disagree, that author counts as mixed, and a half point rounds away from zero. Sources with no opinion are listed and not counted. Strong positive is +50 or higher, positive +20 to +49, mixed −19 to +19, negative −20 to −49, strong negative −50 or lower.
- Weighted reading, second view. Counted the older way, where an account GILD matched to a named person counts two or three times, leaders and teams read −34 negative (method sentiment/S1). GILD printed that number as "GILD score" until October 29, 2026; most of those matches rest on a name or a self-described role.
- Method change, Oct 30, 2026. GILD changed how practitioner sentiment counts, to one vote per author. Weighted, it read −14 Mixed; with one vote per author it reads −8 Mixed. No evidence changed.
- Unscored. A group gets a score only with 20 sources from 15 authors. Below that, one source moves the score about 5 points, so the read says "Unscored until 20 sources from 15 authors" and shows its counts.
- Independence. A source whose author works for the vendor, for one of its partners or programs, or for a competitor, and a page the vendor published or ran, is listed with that tie and counts toward nothing. One author counts on at most two records per product. The line under the score says how many authors and how many separate discussions the counted sources come from: comments in one thread are one discussion.
- Praise and concerns. How many of the sources that discuss a theme praise it or raise it.
- Consensus. A statement joins when 20% of the group's sources support it and fewer contradict it. Contradicting sources are counted separately. Below that line a statement shows as building. At 40% or more it reads Strong agreement; below, Recurring point.
- Disagreements. The count on each side. GILD names no winner and gives no share of all users.
- Minimum evidence. A finding needs 5 sources from 3 authors; a disagreement also needs 2 on each side. Smaller counts stay in the table below, marked under the minimum.
Labels on a source
A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.
- Documented
- Stated in official documentation.
- Reported
- A practitioner's account of their own use.
- Measured
- Backed by the source's own measurements, such as a comparative test or tracked team metrics.
- GILD analysis
- GILD's check of a claim against official documentation.
- Current
- Describes the product as it is now.
- Still applicable
- Older, and the behavior or lesson still holds.
- Needs a recheck
- May depend on a plan, version or setting that changed since it was written. It keeps counting until GILD rechecks it, since the rest of the account usually holds, and its card says so.
- Out of date
- No longer accurate. Kept in the record; counts toward nothing current.
Flagged for a recheck, of the sources that take a side: leaders and teams 10 of 20, individual users 7 of 45, all users 17 of 65.
Limits
- GILD collected public Perplexity accounts published in 2025 and 2026 through Hacker News's search API. Other sources came from podcast episodes, personal blogs and newsletters, university and institutional pages, independent citation audits and arXiv papers, company security research blogs, GitHub issues, and a video episode. These counts describe the reviewed sources. They do not represent all customers. Hacker News contributes more than half the sources, so individual users dominate. Institutions and research or security teams supply the accounts for leaders. G2 and Perplexity's customer stories blocked automated access or lacked publication dates.
- Sources name their plan, mode, and model inconsistently, and Perplexity has renamed features since many were written. Spaces became Projects in August 2026. Sources that depend on a retired model lineup, an earlier Deep Research, or a renamed feature are marked Needs a recheck.
Every count, and the sources behind it
18 findings for 3 groups
Each row opens its sources for the group you are reading.
| Finding | Leaders and teams rolling it out | Individual users | All users | Sources |
|---|---|---|---|---|
| Current readCurrent sentiment | 5 positive, 5 mixed, 10 negative | 17 positive, 12 mixed, 16 negative | 22 positive, 17 mixed, 26 negative | Sources |
| PraiseCited answers | 5 of 9 positive | 4 of 8 positive | 9 of 17 positive | Sources |
| PraiseResearch speed | 1 of 1 positive (under the minimum) | 10 of 10 positive | 11 of 11 positive | Sources |
| PraiseDeep Research & reports | 0 of 4 positive (under the minimum) | 7 of 9 positive | 7 of 13 positive | Sources |
| PraiseModel choice | 2 of 2 positive (under the minimum) | 4 of 6 positive | 6 of 8 positive | Sources |
| ConcernAccuracy | 5 of 6 raise it | 7 of 10 raise it | 12 of 16 raise it | Sources |
| ConcernSource quality | 3 of 4 raise it (under the minimum) | 6 of 7 raise it | 9 of 11 raise it | Sources |
| ConcernLimits & pricing | 0 of 1 raise it (under the minimum) | 3 of 5 raise it | 3 of 6 raise it | Sources |
| ConcernData & security | 4 of 5 raise it | 4 of 5 raise it | 8 of 10 raise it | Sources |
| ConcernFeature gaps | 2 of 2 raise it (under the minimum) | 4 of 4 raise it (under the minimum) | 6 of 6 raise it | Sources |
| ConcernSpeed & reliability | 0 of 0 raise it (under the minimum) | 5 of 6 raise it | 5 of 6 raise it | Sources |
| ConsensusUsers check the sources | 0 supporting (under the minimum) | 6 supporting (below threshold) | 6 supporting (below threshold) | Sources |
| ConsensusIt replaced web search | 0 supporting (under the minimum) | 5 supporting (below threshold) | 5 supporting (below threshold) | Sources |
| ConsensusA starting point for research | 0 supporting (under the minimum) | 6 supporting (below threshold) | 6 supporting (below threshold) | Sources |
| ConsensusUsed beside a general assistant | 2 supporting (under the minimum) | 8 supporting (below threshold) | 10 supporting (below threshold) | Sources |
| DisagreementDoes a team still need Perplexity once its general assistant can search the web? | Perplexity for research 0, The assistant covers it 1 (under the minimum) | Perplexity for research 3, The assistant covers it 7 | Perplexity for research 3, The assistant covers it 8 | Sources |
| DisagreementIs the paid plan worth keeping? | Worth paying for 0, Canceled or would not pay 0 (under the minimum) | Worth paying for 4, Canceled or would not pay 8 | Worth paying for 4, Canceled or would not pay 8 | Sources |
| Current readEvidence mix | 41 sources | 64 sources | 89 sources | Sources |
Fact checks
GILD's checks of claims against official documentation, kept apart from the source counts.
- Accuracy. Perplexity's documentation says a source label describes a whole website. It never vouches for a single claim or replaces reading the source. Model Council, on Max and Enterprise Max, runs one question across three models and shows where they disagree. Understanding source labels, checked Oct 12, 2026
- Source quality. Perplexity marks some domains as Government, Academic, or Trusted through its source review process. Partnerships and payments do not affect the labels, and most domains carry no label. Understanding source labels, checked Oct 12, 2026
- Limits & pricing. Enterprise Pro allows 400 Pro Searches a week and 50 Research queries a month; Enterprise Max allows 4,000 and 500. Computer runs on credits at 100 per US$1, with 500 a month included per Enterprise Pro seat. Which Perplexity Subscription Plan is right for you?, checked Oct 12, 2026
- Data & security. Enterprise data is never used to train Perplexity's models, and agreements bar third-party providers from training on it. Files attached to sessions are deleted after seven days. Consumer Pro and Max users can opt out of training data collection. Data retention and privacy for Enterprise organizations, checked Oct 12, 2026