Perplexity Citation Accuracy Test: What Scores Miss

I treat the Perplexity citation accuracy test as a measurement problem before I treat it as a trust problem: the same product can look near flawless on a controlled factual benchmark, yet still produce citation mismatches often enough to matter in publishing, research, policy and legal-adjacent work. The sharpest contradiction is this: Perplexity says one Deep Research mode has achieved 93.9% factuality, while the Tow Center found Perplexity answered 37% of its tested news-retrieval queries incorrectly, and those two numbers can both be useful if the methodology is clear.

That is the core answer. Perplexity’s citation accuracy is not a single stable percentage. It changes with query type, source availability, search mode, pricing tier, prompt clarity, language, topic sensitivity and the definition of accuracy being used. A citation can be present but unsupported. A source can be reputable but irrelevant to the exact claim. A benchmark can be impressive but too clean to represent messy live research.

During our 2026 editorial evaluation, I separated factual correctness from citation support, source quality and claim-level traceability. That distinction matters for anyone preparing a board report, academic literature scan, newsroom explainer, investor brief or technical comparison. This article explains why published accuracy figures diverge, how to build a repeatable verification workflow, what Perplexity’s plans and API costs imply for serious audits, and when a cited answer should be treated as a map to evidence rather than as finished evidence itself.

Why Perplexity Citation Accuracy Test Results Disagree

A useful Perplexity citation accuracy test has to answer four questions, not one. Is the generated answer factually correct? Does every material claim have a citation near it? Does the cited page actually contain the claim? And is that cited page the best available evidence for the claim? Many public discussions collapse these into one score, which is why accuracy figures drift from the mid-60s in stricter source-support reviews to the high-90s in easier factuality claims.

Benchmarks usually reward answer correctness under controlled conditions. Real-world citation audits reward evidence fit under changing web conditions. The Tow Center’s 2025 test, for example, was not a general intelligence benchmark. It asked AI search tools to identify news articles from excerpts, then judged whether the correct article, publisher and URL were retrieved. Under that test, all eight tools collectively returned incorrect answers more than 60% of the time, and Perplexity returned incorrect answers in 37% of its tested queries. That does not prove Perplexity is inaccurate for every task. It proves that a citation-rich interface can still fail when source retrieval is brittle.

This is where the Perplexity accuracy analysis becomes useful for readers who want the broader scorecard. The right interpretation is not that one number defeats another. It is that benchmark scores, spot checks, newsroom studies and API logs measure different layers of the system. A controlled factual answer can be right because the model knows the fact. A cited live answer can be wrong because retrieval landed on a syndicated copy, a stale article, a thin blog or a page that contains the right entity but not the right claim.

In practical terms, accuracy should be scored at the claim-source pair level. The answer may be 90% right overall, while one unsupported number is the only part a reader later repeats. That single unsupported number is the real risk.

Accuracy Definition	What It Checks	Typical Failure
Answer accuracy	Whether the generated statement is true	Correct answer with weak or absent citation
Citation presence	Whether the answer includes linked sources	Many links that do not prove the claim
Citation support	Whether the cited page contains the specific claim	Source mentions the topic but not the fact

The Benchmark Gap: SimpleQA, HLE, and Real Queries

The benchmark gap begins with task design. SimpleQA-style tests ask short factual questions with verifiable answers. The revised SimpleQA Verified paper argues that factuality benchmarks need better filtering, topic balancing and source reconciliation because noisy labels and duplicated prompts can distort conclusions. That warning matters for Perplexity because a headline figure such as 93.9% sounds precise, but the underlying benchmark still captures a narrower skill than claim-by-claim source verification.

Humanity’s Last Exam moves in the opposite direction. It contains 2,500 frontier questions across subjects such as mathematics, humanities and natural sciences, designed so that each solution is verifiable but not quickly answerable through internet retrieval. This means HLE is useful for measuring high-end reasoning and expert knowledge, but it is not a direct test of whether a live Perplexity citation supports a paragraph in a newsroom article.

A real Perplexity citation accuracy test sits between these worlds. It should include simple factual questions, messy news queries, academic claims, product pricing checks and multi-hop prompts where the answer depends on more than one source. During our 2026 evaluation, the most revealing questions were not the hardest questions in the abstract. They were ordinary professional questions with one trap: a source that looked credible but did not actually contain the exact number or legal wording being repeated.

That is why I would not publish one blended accuracy percentage without a methodology appendix. Perplexity may perform strongly on structured current facts, especially when primary sources are indexed and the question is unambiguous. It will face more risk on contested news, niche technical claims, recent product limits, healthcare guidance, politics, litigation and subjects where reputable sites disagree. The published evidence points to a simple editorial rule: benchmark performance tells you whether a system can answer; citation audit performance tells you whether you can safely publish the answer.

Evidence Type	Useful For	Not Enough For
SimpleQA or SimpleQA Verified	Short-form factuality and benchmark comparison	Citation support or source authority
Humanity’s Last Exam	Expert reasoning and difficult closed-ended questions	Live web retrieval accuracy
Tow Center news citation test	Article attribution and retrieval behaviour	General factual accuracy on all topics
Claim-level editorial audit	Publishable support for each claim	Fast leaderboard-style comparison

What Our 2026 Citation Spot Check Looked For

Perplexity Citation Accuracy Test Scoring Rubric

A useful audit starts by shrinking the unit of analysis. Instead of asking whether a Perplexity answer is broadly right, I split each response into atomic claims. An atomic claim is a statement that can be verified or falsified without relying on the rest of the paragraph. Pricing, model names, usage caps, dates, named executives, benchmark scores and regulatory assertions all become separate test items.

The next step is to grade each citation against the exact claim it sits beside. A citation earns full support only if the linked source contains the relevant fact in a way that matches the wording, scope and time period. Partial support applies when the source supports the general idea but not the precise number, geography or limitation. No support applies when the source is merely topically related. Broken support applies when the link fails, redirects to a generic page or points to a source that cannot be inspected.

This is also where the Perplexity APA citation guide becomes more than a formatting article. APA, MLA and Chicago styles are downstream tasks. Before formatting a citation, the researcher must decide whether the source deserves to be cited at all. In our testing workflow, a beautifully formatted reference to the wrong source scored worse than a plain source note to the correct primary document.

The most important edge case is the paraphrase gap. Perplexity may cite a page that contains all the ingredients needed to infer a claim, while the answer states the inference as fact. For internal research, that may be acceptable if flagged. For publication, it is not. The audit should mark the claim as inferred, not directly supported. This creates a defensible paper trail when editors, clients or reviewers ask why one source was accepted and another rejected.

Where Citations Break: Four Failure Modes

Citation failures are not all hallucinations. The scary version is a source that does not exist, but the more common version is quieter: a real source that does not support the attached claim. That distinction is important because users often relax once a link opens. The link opening is only the first test.

The first failure mode is source mismatch. The answer cites a page about the same topic, company or paper, but the specific statistic is absent. The second is source drift. A live page changes after Perplexity indexed or retrieved it, leaving the answer tied to a moving target. The third is authority inversion. A blog post, SEO page or syndicated copy is cited where the official document, original paper or regulator page should have been used. The fourth is synthesis overreach. Several sources each support one piece of a conclusion, but the answer compresses them into a stronger claim than any source actually makes.

The best AI citation tools market now exists because these errors are expensive to catch manually at scale. Yet automated checkers still need human review for nuanced claims. A checker can confirm that a page mentions a phrase. It cannot always judge whether the wording preserves causality, population, sample size and confidence level.

The 2026 synthetic-source audit by Mowafak Allaham and Nicholas Diakopoulos adds another layer. Their paper reports evidence that roughly 16% of cited sources across four generative search engines were AI-generated. The problem is not simply that AI-generated pages are always false. The problem is provenance. A source can be machine-written, derivative and unreviewed while still looking polished enough to be retrieved and cited.

Failure Mode	What Happens	Editorial Fix
Source mismatch	The page is real but does not prove the claim	Search within the page for the exact fact and mark unsupported if absent
Source drift	The cited page changes after retrieval	Capture publication date, access date and archive copy
Authority inversion	A secondary page replaces the primary source	Prefer official filings, papers, vendor docs and regulator pages
Synthesis overreach	The answer states a stronger conclusion than sources support	Downgrade language and identify the inference

Pricing, Limits, and the Cost of Verification

Pricing matters because citation verification is not just a research habit. It is a workflow cost. Official Perplexity pages reviewed for this article show Pro at $17 per month when billed annually, Enterprise Pro at $34 per month per seat when billed annually and Enterprise Max at $271 per month per seat when billed annually. The same enterprise matrix says Pro offers up to 200 Pro queries per week and up to 20 Deep Research queries per month, while enterprise tiers multiply those limits.

The help centre adds useful operational caps. Free users get 3 Pro Searches per day and 1 Research query per month. Enterprise Pro is listed with extended limits of 400 Pro Searches per week, 50 Research queries per month and 80 Comet Assistant queries per month. Enterprise Max is listed with 4,000 Pro Searches per week, 500 Research queries per month and 800 Comet Assistant queries per month. Those are exactly the kinds of hidden limits a research lead needs before promising an always-on citation audit process.

API pricing is a separate line item. Perplexity’s Search API is $5 per 1,000 requests with no token charges. Sonar, Sonar Pro and Sonar Reasoning Pro combine token pricing with search-context request fees. Sonar Deep Research is more complex: $2 per 1 million input tokens, $8 per 1 million output tokens, $2 per 1 million citation tokens, $5 per 1,000 search queries and $3 per 1 million reasoning tokens. Perplexity’s own cost example shows that one Deep Research query can be materially more expensive when it performs many searches and reasoning steps.

The practical lesson is simple. For publication verification, budget by claim volume, not by article count. A 1,000-word article with 40 factual claims can cost more to audit than a 3,000-word essay with only ten factual claims.

Plan Or API	Current Public Price or Cost	Relevant Limit or Cap	Verification Implication
Free	$0 public tier	3 Pro Searches daily and 1 Research query monthly	Useful for casual checks, not enough for editorial operations
Pro	$17 per month when billed annually	Up to 200 Pro queries weekly and 20 Deep Research queries monthly	Adequate for individual researchers with manual review
Enterprise Pro	$34 per seat monthly when billed annually	400 Pro Searches weekly and 50 Research queries monthly in help centre matrix	Better for teams validating recurring reports
Enterprise Max	$271 per seat monthly when billed annually	4,000 Pro Searches weekly and 500 Research queries monthly in help centre matrix	Designed for high-volume research and audit teams
Search API	$5 per 1,000 requests	Request-based, no token costs	Useful for raw retrieval checks
Sonar Deep Research API	Token, citation, reasoning and search-query billing	Search count is model-determined, influenced by reasoning effort	Cost can rise sharply on complex audits

Perplexity Features That Affect Citation Quality

Perplexity’s citation quality is shaped by product architecture as much as by model intelligence. The March 2026 changelog says Perplexity Computer gives Pro users access to 20+ advanced models, prebuilt and custom skills and hundreds of connectors. For enterprise users, it says Computer routes tasks across 20 specialised models and connects to more than 400 applications, including Snowflake, Salesforce and HubSpot. That matters because citations are no longer only web links. In enterprise settings, evidence may come from internal files, apps, databases and proprietary data providers.

The same changelog introduced useful implementation details: Computer integrates into Slack, MCP connectors can be added with OAuth, API key or open authentication, and Snowflake integration creates a Data Map, a semantic layer that learns schemas, tables, column relationships and historical query patterns. These are powerful features, but they complicate citation accuracy. A correct answer from Snowflake may need a query log, table definition and timestamp, not just a link. A correct answer from Slack may need channel permissions and message provenance. A correct answer from a custom MCP connector may need connector-level audit logs.

Perplexity’s Model Council accuracy mode points in the right direction conceptually because serious research benefits from disagreement. A multi-model comparison can surface uncertainty, but it still does not prove the cited evidence. Three models agreeing on a claim can still agree because they retrieved the same weak secondary source.

For researchers, the feature lesson is operational. More connectors increase coverage. More models increase alternative reasoning paths. Neither automatically increases citation support. Citation quality improves when the system records which source was retrieved, why it was selected, which claim it supports and whether a human reviewer accepted that support.

How Perplexity Compares With AI Search Rivals

Perplexity’s biggest strength is that citation is part of the interface rather than an afterthought. Compared with a general chatbot response, Perplexity makes source inspection easier because links are displayed near the answer. That design choice explains why many researchers still use it as a first stop even after reading negative citation studies. The problem is not that citations are useless. The problem is that citations can be mistaken for verification.

Recent academic work suggests the comparison is more nuanced than citation count. Zhang, He and Yao’s 2026 citation absorption framework found that Perplexity and Google cite more sources on average, while ChatGPT cites fewer sources but shows higher average citation influence among fetched pages in their dataset. In plain English, more links do not necessarily mean more evidence. A concise answer with two strong sources can beat a long answer with ten weak ones.

For readers comparing assistants, the AI answer tool rankings should be read through this lens. Perplexity is strong for sourced current answers, fast topic mapping and finding evidence trails. ChatGPT is often stronger for long-form reasoning, document work and coding workflows. Claude is often favoured for careful prose and dense document interpretation. Gemini can be useful when Google ecosystem integration matters. None removes the need for source checking.

Jean Philip De Tender, the EBU’s media director, captured the public-trust stakes after the EBU and BBC study: “When people don’t know what to trust, they end up trusting nothing at all.” That quote is not only about news. It applies to enterprise research too. If employees see confident but poorly sourced AI summaries, they may either overtrust them or reject useful AI tools entirely. The better approach is calibrated trust: use AI search where it is efficient, then verify where consequences are real.

Tool Category	Best Use	Citation Risk
Perplexity	Fast sourced answers and live research maps	Citation may be real but mismatched to claim
ChatGPT with search	Mixed reasoning, drafting and workflow support	Fewer links may hide source-selection gaps
Google AI Overviews or Gemini	Search-adjacent summaries and Google ecosystem context	Sourcing varies by query and region
Specialised academic tools	Paper discovery and literature triage	May overstate study strength without reviewer context

Verification Workflow for Researchers and Publishers

The safest workflow is boring, repeatable and documented. First, split the answer into claims. Second, label each claim by risk: low-risk background, numerical fact, legal or regulatory claim, health claim, pricing claim, quote, technical specification or analytical inference. Third, open the cited source. Fourth, find the passage that supports the claim. Fifth, decide whether the support is direct, partial, inferred, contradicted or absent. Sixth, replace weak citations with primary sources where possible.

For academic work, add one more step: check whether the source’s population, method and confidence level match the wording in your article. If a paper studies 127 respondents in one country, the sentence cannot become a global behavioural claim. If a benchmark tests short factual questions, it cannot become proof of broad research reliability. This is where the AI researcher tool stack is most valuable: combine Perplexity for discovery, scholarly databases for paper retrieval, reference managers for formatting and human review for interpretation.

Deborah Turness, BBC News CEO, warned that “Gen AI tools are playing with fire” after BBC researchers found significant issues in AI answers about current affairs. That warning should not lead professionals to stop using AI search. It should lead them to put firebreaks around it.

A good firebreak is a citation worksheet. Each row contains the claim, cited source, source type, quote or page section, reviewer decision, replacement source and final publication wording. The worksheet turns citation checking from a vague act of caution into an auditable process. It also makes editors faster because they can review disputed claims without rerunning the whole Perplexity session.

API Implementation Notes for Citation Audits

Teams that want repeatable audits should avoid building the process only in the browser. The API gives better logging and reproducibility. A simple implementation starts with a prompt that requests cited claims in structured JSON: claim text, citation URL, source title, retrieval date and confidence. The audit layer then fetches each URL, extracts text, searches for supporting passages and stores the reviewer decision.

Perplexity’s API documentation matters here because cost and behaviour vary by component. The Agent API can use tools such as web_search, fetch_url, people_search, finance_search and sandbox. The Search API returns raw web search results at $5 per 1,000 requests. Sonar models add token and request fees, while Deep Research adds citation-token, reasoning-token and search-query charges. This means a verification product should use the cheapest retrieval layer that answers the question. Do not send every claim to Deep Research if a primary vendor page or filing can be fetched directly.

A practical Perplexity citation accuracy test in the API should include at least three passes. The first pass retrieves the answer and citations. The second pass independently fetches the cited pages and checks whether the claim text is present. The third pass searches for better primary support if the cited source fails. This design catches source mismatch without forcing the model to judge its own evidence.

For high-stakes topics, store source snapshots. Source drift is unavoidable on live pages, and a source that supports a claim today may not support it after a pricing page, support article or newsroom page is edited. Archiving is not bureaucracy. It is the only way to explain a citation decision months later.

Enterprise Constraints, Connectors, and Bottlenecks

Enterprise users face a different problem from individual researchers. They do not only ask whether a public URL supports a claim. They ask whether Perplexity can cite across permissioned work apps without leaking context, losing access-control boundaries or summarising a stale internal document. Official enterprise materials emphasise no training on customer data, SSO or SCIM provisioning, user management, permissioning, dedicated support, compliance claims and data retention features. Those features matter because citation accuracy is inseparable from data governance.

The biggest bottleneck is permission-aware retrieval. If an answer cites a Slack message, CRM field or warehouse table, the audit trail must show that the user had permission to view it, the source was current and the summary preserved the underlying record. A normal public citation standard is too weak. Internal citations need record IDs, timestamps, connector identity and, ideally, a replayable query or retrieval log.

The Perplexity ranking guide helps explain the public-web side of this problem: AI engines do not merely select sources, they absorb evidence, wording and structure into answers. In the enterprise, the same dynamic applies to internal knowledge. A well-structured source will be easier to retrieve and summarise than a poorly named PDF in a shared drive, even if the PDF is more authoritative.

Peter Archer, the BBC’s programme director for generative AI, said that “Our research can only scratch the surface of the issue.” That humility is useful in enterprise rollouts. A pilot should test not only whether Perplexity answers accurately, but whether it cites the right tier of evidence: source system over export, policy page over chat comment, signed contract over sales deck and primary data table over narrative summary.

What Low-Quality Sources Look Like in Perplexity

Low-quality sources are often easy to recognise once you know the pattern. They have no named author, no publication date, no methodology, no links to primary evidence and language that mirrors other AI-written pages. They may rank well because they are structured for search, not because they are authoritative. In citation audits, I treat these as discovery sources only. They can point towards a better source, but they should not usually be the final citation for a publishable claim.

The highest-risk sources are synthetic roundups, content farms, thin affiliate pages, copied press releases without context, scraped news pages, forum comments used as factual authority and stale product pages that have been superseded by current documentation. The 2026 Allaham and Diakopoulos audit gives this concern empirical weight by showing evidence of AI-generated sources inside generative-search citation sets. Their finding does not mean every synthetic source is wrong. It means provenance now deserves its own scoring column.

The Perplexity statistics reference is a good example of when readers should demand methodology clarity. Statistics articles can be useful entry points, but any number about users, revenue, valuation, accuracy or market share should be traced to filings, official statements, credible research reports or transparent third-party datasets. A statistic without lineage is not evidence. It is a rumour with formatting.

A simple source-quality checklist helps. Is the source primary? Is the date current for the claim? Does the author or institution have relevant expertise? Does the page show methodology? Does the cited sentence directly support the claim? Does another independent source corroborate it? If any answer is no, downgrade the citation or keep looking.

Perplexity Citation Accuracy Test Verdict

The most defensible verdict is conditional. Perplexity is one of the strongest mainstream tools for rapid source discovery and cited current answers, but the citation layer still needs verification before publication, academic submission, policy work or commercial decisions. Treat it as a research accelerator, not as a citation authority.

The positive case is real. Perplexity puts sources close to answers, exposes users to evidence trails, supports advanced search modes, offers Deep Research and now provides API layers that can be integrated into custom audit workflows. Its enterprise roadmap also shows serious attention to connectors, data sources and multi-model workflows. These are useful building blocks for trustworthy AI-assisted research.

The negative case is equally real. Independent newsroom and academic studies show that AI search systems can misattribute sources, overstate support, cite AI-generated pages and give confident answers when retrieval fails. The Tow Center’s 37% incorrect figure for Perplexity in a specific article-retrieval test is not a universal accuracy rate, but it is a warning against blind trust. Reuters’ coverage of the EBU and BBC study adds a broader news-context warning: 45% of studied AI responses contained at least one significant issue, and a third had serious sourcing errors.

My working scorecard therefore has three grades rather than one. For source discovery, Perplexity is high-value. For answer drafting, it is useful with review. For final citation authority, it is not sufficient on its own. That may sound cautious, but it is also generous. A tool does not need to be perfect to be valuable. It needs to be used at the right stage of the research process.

Takeaways

• Use Perplexity to find source trails quickly, not to replace claim-level verification.
• Score citation accuracy at the claim-source pair level, because answer-level accuracy hides unsupported facts.
• Treat the 93.9% factuality figure as a benchmark signal, not as proof that every live citation supports every generated claim.
• Budget verification by factual-claim volume, especially when API use involves Deep Research citation, reasoning and search-query billing.
• Prefer primary sources for pricing, regulations, benchmarks, scientific claims, corporate statements and technical limits.
• Flag source mismatch, source drift, authority inversion and synthesis overreach as separate failure modes.
• Archive or snapshot important cited pages when the work may be reviewed weeks or months later.
• Publish Perplexity-assisted research only after a human reviewer confirms the cited source contains the exact claim.

Our Editorial Verification Process

This article was built by cross-referencing Perplexity’s official enterprise pricing page, subscription help centre, API pricing documentation, Sonar Deep Research model documentation and March 2026 changelog with independent evidence from the Tow Center, Reuters, The Guardian and recent arXiv research on SimpleQA Verified, Humanity’s Last Exam, synthetic sources and citation absorption. I treated pricing, plan limits, API costs, benchmark scores and citation-risk figures as confirmed only when tied to a retrievable source. Where a widely repeated metric, such as a 78% citation spot-check figure, could not be verified from a primary or methodologically transparent source during research, I did not use it as a confirmed benchmark. The evaluation framework used claim-level support, source authority, source freshness, retrieval reproducibility and auditability as the core metrics, because those are the dimensions that determine whether a Perplexity citation can safely move from research notes into published work.

Conclusion

Perplexity is valuable because it makes research feel faster, more structured and more transparent than a blank search box. Its cited answers can shorten the distance between a question and a useful evidence trail. The danger is that the interface can also make verification feel finished before it has started.

The best evidence in 2026 points to a balanced position. Perplexity’s benchmark and product claims show strong capability on many factual and research tasks. Independent studies show that citation accuracy still varies sharply when queries become messy, sources are blocked, pages are synthetic, news moves quickly or the answer compresses several sources into one confident statement. Those findings are not contradictory. They describe different layers of the same system.

The future likely belongs to workflows that combine retrieval, multi-model review, source snapshots, structured citation scoring and human judgement. Perplexity can be part of that workflow, especially for teams that need rapid discovery and broad source coverage. The open question is how well AI search systems will expose uncertainty, provenance and source-support strength without forcing users to build their own audit layers. Until that improves, the safest rule is simple: trust Perplexity enough to start faster, but verify enough to publish responsibly.

FAQs

What Is a Perplexity Citation Accuracy Test?

It is an audit that checks whether Perplexity’s cited sources actually support the claims in its answer. The best version scores answer correctness, citation presence, source support and source authority separately.

Is Perplexity More Accurate Than ChatGPT for Citations?

Perplexity is usually easier to inspect because citations are central to the interface. That does not automatically make every citation more accurate. Some studies show Perplexity cites more sources, while source-support quality still varies by query.

Why Do Perplexity Accuracy Scores Vary So Much?

Different tests measure different things. Benchmarks may measure short factual answers, while newsroom or research audits test whether citations retrieve the correct source and support the exact claim.

Can Perplexity Hallucinate Sources?

Yes, but hallucination is not the only risk. More often, the source exists but does not support the attached claim, points to a copied version, or reflects a page that changed after retrieval.

Should Academic Researchers Cite Perplexity Directly?

Usually no. Researchers should cite the primary source Perplexity helps uncover, not Perplexity’s generated answer, unless the research topic is the AI output itself.

How Many Sources Should I Check Before Trusting an Answer?

For casual use, check at least one cited source. For publication, check every material claim, and use at least two independent sources for contested, numerical or high-stakes claims.

Does Perplexity Pro Improve Citation Accuracy?

Paid tiers can improve access, models and research depth, but a stronger model does not guarantee citation support. The Tow Center found premium models could still be confidently wrong in its news-retrieval test.

What Is the Fastest Way to Detect a Bad Citation?

Open the cited page, search for the exact number, quote or claim, then check date and source authority. If the page only discusses the general topic, the citation is weak.

References

Allaham, M., & Diakopoulos, N. (2026). Synthetic sources?: Auditing generative search engine citations for evidence of AI-generated sources. arXiv. https://arxiv.org/abs/2605.23684

Haas, L., Yona, G., D’Antonio, G., Goldshtein, S., & Das, D. (2026). SimpleQA Verified: A reliable factuality benchmark to measure parametric knowledge. arXiv. https://arxiv.org/abs/2509.07968

Jaźwińska, K., & Chandrasekar, A. (2025, March 6). AI search has a citation problem. Columbia Journalism Review. https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php

Le Poidevin, O. (2025, October 21). AI assistants make widespread errors about the news, new research shows. Reuters. https://www.reuters.com/business/media-telecom/ai-assistants-make-widespread-errors-about-news-new-research-shows-2025-10-21/

Perplexity. (2026a). Enterprise pricing. https://www.perplexity.ai/enterprise/pricing

Perplexity. (2026b). Pricing. Perplexity API documentation. https://docs.perplexity.ai/docs/getting-started/pricing

Perplexity. (2026c). Sonar Deep Research. Perplexity API documentation. https://docs.perplexity.ai/docs/sonar/models/sonar-deep-research

Phan, L., Gatti, A., Han, Z., Li, N., Hu, J., Zhang, H., & others. (2026). Humanity’s Last Exam. arXiv. https://arxiv.org/abs/2501.14249

Zhang, K., He, X., & Yao, J. (2026). From citation selection to citation absorption: A measurement framework for generative engine optimization across AI search platforms. arXiv. https://arxiv.org/abs/2604.25707