AI deep research tools compared: ChatGPT Deep Research, Perplexity, Gemini Deep Research, Elicit — citation reliability, verifiability, accuracy evidence

Citation integrity of this page, measured by an independent cross-family auditor: 91.8% (95% CI 82–97%). Every claim above cites its source; the bibliography is at the end. That is the standard we hold research to — including our own marketing.

Research question. How do the major AI deep-research tools differ in citation reliability and verifiability, and how can a user check any tool's output before relying on it?

The deep-research tools, briefly

Four tools dominate the current conversation about AI-assisted deep research, and while they share a family resemblance — all use large language models, all promise to synthesize sources rather than merely list them — they differ substantially in how they retrieve information, what they retrieve it from, and what they hand back to the user.

Perplexity positions itself as a search-synthesis hybrid. Its model blends real-time web retrieval with language-model reasoning, which means the tool is simultaneously fetching and narrating. As one practitioner guide describes it, "Perplexity AI blends search and synthesis. Good prompting therefore means steering both the retrieval (what it fetches) and the reasoning (how it writes and structures the answer)."¹³ The intended audience skews toward researchers, analysts, and professionals who want a quick, sourced answer to a factual question and are willing to steer the tool through disciplined prompting. Perplexity's comparative strength in citation handling relative to other single-model tools — rated "Strong" on citation handling in one multi-tool comparison table where GPT, Claude, Gemini, and Grok all receive "Moderate" — comes with an important caveat appended to the same table: "All models require explicit verification steps for citations. None should be trusted to self-verify without source binding."⁵

Gemini Deep Research and OpenAI's ChatGPT Deep Research represent a second, more ambitious category: agents that conduct extended, multi-step research sessions rather than answering a single query. These tools generate substantially more citations per query than simpler search-augmented models. A University of Pennsylvania preprint measuring citation URL validity across ten commercial systems found that deep research agents "cite far more URLs per query than search-augmented models," with this greater citation volume accompanied by a higher hallucination rate: pooling across the two deep research agents tested, the hallucination rate for URLs that had no Wayback Machine record — indicating they likely never existed — was 10.7%, compared to 4.8% for search-augmented models.⁹ The same study found Gemini 2.5 Pro achieved the lowest non-resolving URL rate among models tested on the ExpertQA benchmark, at 4.20%, while generating the fewest URLs per question — a pattern suggesting that citation volume and citation accuracy do not move together.⁹ The intended audience for these deep-research agents is users who want a research memo rather than a search result: something that surveys a landscape, synthesizes competing claims, and returns a structured report with inline sourcing. The output shape resembles a draft literature review more than a search-engine results page.

Elicit occupies a distinct niche. Where the other tools cast wide nets across the live web, Elicit focuses on peer-reviewed academic literature and is built to support systematic review workflows — the structured, reproducible methodology used in fields like medicine and public health to aggregate evidence across studies. Its intended users are academic researchers who need to work through large bodies of scientific papers, extract structured data from individual studies, and identify patterns across a corpus. The depth-versus-breadth trade-off is real: by limiting retrieval largely to indexed academic papers, Elicit reduces the noise that comes with web-wide search while also forgoing coverage of recent preprints, grey literature, and real-time developments that have not yet passed through peer review.

All four tools share a common architectural challenge that runs through the independent empirical literature: the presence of a citation does not guarantee either that the cited URL resolves to a real document or that the document, if real, supports the claim made about it. The legal guidance literature distinguishes two failure modes — fabrication (a citation that does not exist) and misgrounding (a real citation deployed to support a proposition it does not actually endorse) — and notes that the second type may be harder to detect precisely because it passes a basic existence check.² The University of Pennsylvania preprint, which measures URL validity rather than claim support, explicitly acknowledges that URL liveness is only one layer of the problem: fabricated text snippets and incorrect bibliographic metadata are separate failure modes its methodology does not capture.⁹

One point of genuine disagreement bears naming at the outset: whether disciplined prompting in a single-model tool like Perplexity is sufficient to produce research-grade output, or whether only multi-model orchestration with cross-validation can catch the failures that a single model cannot detect in itself.¹³⁵ The available evidence does not settle that question. Both positions are argued from promotional or commercially interested starting points, and no head-to-head empirical test of the specific tools in this comparison — ChatGPT Deep Research, Perplexity, Gemini Deep Research, and Elicit — exists in the documentary record.

What the evidence says about citation reliability

The numbers that exist on citation reliability in AI research tools are striking — and strikingly inconsistent. What the available record offers is not a single settled measurement but a distribution of findings that varies by model generation, domain, and what exactly is being counted as a failure.

The most systematic URL-level evidence comes from a University of Pennsylvania preprint by Rao, Wong, and Callison-Burch, who evaluated ten commercial LLMs and deep research agents on DRBench (53,090 URLs) and three models on ExpertQA (168,021 URLs). Across all ten systems, hallucinated URLs — those with no Wayback Machine record, indicating they likely never existed — ran from 3% to 13%, while overall non-resolving rates reached 5% to 18%.⁹ A finding with direct bearing on the tools this report covers: deep research agents (specifically Gemini Deep Research and OpenAI Deep Research) pooled to a hallucination rate of 10.7% compared with 4.8% for search-augmented models — a statistically significant gap (two-proportion z = 15.15, p < 10−51), even though agents generate far more citations per query than their search-augmented counterparts.⁹ Generating more citations does not improve the per-citation error rate; on ExpertQA, GPT-5.1 produced 46.4 URLs per question — roughly 4.3 times more than Gemini 2.5 Pro's 10.7 — yet its non-resolving rate of 8.47% was roughly double Gemini's 4.20%.⁹ The study's hallucination estimates should be read as conservative lower bounds: Wayback Machine coverage is incomplete, meaning some hallucinated URLs are likely misclassified as stale rather than fabricated.⁹

Moving from URL validity to bibliographic fabrication, a broader picture emerges. GhostCite, an arXiv preprint from researchers at Nankai and Tsinghua universities, benchmarked thirteen state-of-the-art LLMs and found citation hallucination rates ranging from 14.23% to 94.93% — a roughly 6.7× spread across vendors — with no single model maintaining uniformly low rates across all research domains.¹¹ Among 56,381 published papers from AI/ML and Security venues spanning 2020–2025, 604 papers (1.07%) contained at least one definitively invalid citation, and the rate of invalid citations in the published literature increased by 80.9% in 2025 relative to the 2020–2024 average.¹¹ One pattern deserves particular attention: hallucination rates increased with recency of the cited publication year, running from 27.61% for citations dated 2000 to 98.75% for those dated 2025, fitting an exponential function with R² = 0.94 — a finding that implies the problem is sharpest precisely where currency matters most.¹¹ The same study found that online search capability, chain-of-thought prompting, and batch size did not show consistent effects on hallucination rates across models, suggesting hallucination reflects a fundamental limitation in models' bibliographic knowledge rather than a surface-level prompting artifact.¹¹

A student abstract presented at AAAI-26 by Misra and Udandarao synthesized figures from prior work: hallucination rates ranging from 18% in GPT-4 to over 70% in other frontier models, with legal-domain rates as high as 88%; and on the CiteME benchmark, a performance gap between LLMs (4.2–18.5% accuracy on citation verification tasks) and human annotators (69.7%).¹⁰ Their own pilot study on 200 citations found hallucination rates of approximately 12% for GPT-4o, 16% for Claude 3.5 Sonnet, and 21% for LLaMA-4 Maverick.¹⁰ These figures are preliminary and the dataset has not yet been fully scaled, so they should be read as directional rather than definitive.

Against this backdrop, vendor-sourced figures require careful handling. INRA.AI's promotional blog claims its six-layer validation holds citation hallucination rates below 0.1%, positioning this against figures attributed to JMIR and Nature publishing for GPT-3.5 (39.6–55%) and GPT-4 (18–28.6%).¹ The GPT-series figures in INRA's materials are directionally consistent with independently published research, but INRA's own sub-0.1% claim is self-reported and appears in no independently audited benchmark — the source does not expose its methodology to external verification.¹ Similarly, Citely's marketing materials assert over 95% accuracy in detecting authentic versus fabricated citations, but this figure is likewise unvalidated by any peer-reviewed or third-party study.⁸ These positions are held by their vendors; the available record provides no independent basis for evaluating them.

The legal domain sheds additional light on failure modes. A guidance document from LeanLaw, drawing on Stanford RegLab research and other secondary sources, documents a distinction between fabrication (invented citations) and misgrounding (real authority cited for a holding it does not support), identifying the latter as potentially more dangerous because it passes basic existence checks while still misrepresenting what a source says.² On domain-specific rates, LeanLaw's synthesis indicates that specialized legal AI platforms hallucinate in 17–34% of queries, while general-purpose tools including ChatGPT hallucinate on legal queries 58–88% of the time.²

Two methodological conflicts in the evidence deserve explicit acknowledgment rather than quiet resolution. First, a preprint presenting the AuthorityBench benchmark (ICML 2026) introduces a distinct failure mode: citation presence in LLM prompts — whether the citations are real or fabricated — consistently increases hallucination rates relative to a no-citation baseline across all seven tested models, with the effect strongest when a fabricated citation accompanies a factually true claim, raising hallucination rates by 3 to 22 percentage points and reaching 35–77% in the general knowledge domain.⁴ The AuthorityBench work frames this input-side susceptibility — models being pushed to deny true facts by false attributions — as the most consequential failure mode. This stands in tension with the larger body of work (GhostCite, Misra and Udandarao, Rao et al.) that characterizes citation hallucination primarily as an output-side fabrication problem: LLMs generating non-existent references rather than being swayed by citations in user prompts.¹¹¹⁰⁹ Both failure modes may be real and may coexist; the available evidence does not establish which is more consequential in practice, and the AuthorityBench paper explicitly designed its benchmark to isolate input-side effects, so it does not bear directly on output-side fabrication rates.

Second, whether self-verification methods can substitute for retrieval-based detection remains contested. The MetaQA approach argues that forcing LLMs to reveal inconsistencies through metamorphic relations is viable without external database dependencies, reporting F1-score improvements of 0.154–0.368 across models.¹² Against this, GhostCite's finding that LLMs achieve only 38% average accuracy on citation validation tasks — below random guessing at 50%, and far below human performance at 69.7% — points toward the inadequacy of self-verification without external ground truth.¹¹ The Misra and Udandarao hybrid pipeline, combining exact lookup, fuzzy retrieval, and LLM-assisted verification, achieved approximately 80% precision in detecting hallucinated citations, representing a 15–20% relative improvement over database-only baselines.¹⁰ The reconciliation the available record supports is partial: zero-resource metamorphic methods and retrieval-augmented pipelines may address different use cases — the former where no external database is available, the latter where citation-specific ground truth exists — but no source demonstrates that self-verification alone performs comparably to retrieval-augmented approaches on citation-specific tasks.

The axis nobody measures: can you check it?

The numbers reviewed earlier tell us how often citations fail; the harder question is whether a user can detect the failure before it causes harm. That turns on a different axis entirely: not how accurate a tool claims to be, but how much of its reasoning chain a user can actually inspect.

The clearest statement of the underlying standard comes from the verification-platform perspective that AI outputs should function as leads rather than conclusions — that a citation earns its status only when a reader can open the referenced work, evaluate it independently, and confirm that it supports the claim being made.³ By that standard, a tool's verifiability depends on whether it exposes enough of its provenance to make that chain of checking possible. Most current deep-research tools do not fully satisfy it. The Suprmind analysis of single-model behavior puts the structural problem plainly: a single model cannot debate itself and will produce internally consistent text that may contradict actual sources without signaling the contradiction.⁵ Internal consistency, in other words, is not evidence of external accuracy — and tools that present only polished prose give users no seam to pull on.

What users can inspect varies substantially across tool types. The verification-platform guidance recommends preferring stable identifiers — DOI, PMID, ISBN — over hyperlinks, precisely because identifiers are more reliable and less subject to link decay.³ A tool that surfaces raw DOIs and publication metadata alongside its synthesis gives a user something to check against CrossRef or PubMed; a tool that buries its sourcing in prose footnotes or omits it entirely forecloses that check. The GhostCite preprint documents a complementary failure: citation hallucination rates increase from 27.61% for sources dated to the year 2000 to 98.75% for sources dated to 2025, fitting an exponential function with R² = 0.94.¹¹ The practical implication is that recent citations — exactly the ones most valuable in a live research question — are also the ones least likely to be real, and therefore the ones that most urgently need external verification.

A second asymmetry concerns what verification tools can and cannot check. The LeanLaw analysis of legal AI use draws a distinction that applies beyond the legal domain: hallucinations fall into two types, fabrication (authority that does not exist) and misgrounding (real authority cited for a proposition it does not support).² Fabrication is the easier failure to catch — a DOI lookup or database query either resolves or it does not. Misgrounding is harder: the identifier is real, but the conclusion drawn from the paper is not in the paper.³ Automated citation checkers address the first problem more readily than the second. Citely's marketing materials, for instance, position the tool as able to verify references from ChatGPT, Claude, and Gemini against CrossRef, PubMed, arXiv, and OpenAlex, and claim over 95% accuracy in detecting authentic versus fabricated citations.⁸ Those accuracy figures are self-reported by the vendor without independent validation, and the claim does not extend to detecting misgrounded citations where the underlying paper exists but is being misread. GPTZero's source-finder, similarly, is described in its own promotional framing as detecting objective claims and matching them to online, academic, and publicly available sources — while noting that the feature remains in beta.⁷ Neither tool's performance has been independently benchmarked against the specific deep-research tools under evaluation here.

The GhostCite data offers the sharpest available benchmark on what automated checking can achieve: the CITEVERIFIER framework the study authors developed reached 100% accuracy on valid citations and 98% accuracy on invalid ones in a validation sample of 800 citations, using DBLP, Google Scholar, and web search as backends.¹¹ That same study, however, found that LLMs asked to perform citation validation themselves achieve only 38% average accuracy across 13 models — below random guessing at 50%.¹¹ The implication is direct: asking the same tool that generated a citation to verify it is not a meaningful check.

This leads to the unresolved disagreement between two distinct positions in the available evidence. One view holds that disciplined prompting strategies — query operators, structured output formats, demands for inline citations — can make single-model tools sufficiently reliable for research when users apply rigorous verification discipline.¹³ The opposing view holds that single models are structurally incapable of self-correction: because they cannot debate themselves, cross-validation across multiple independent models is necessary to surface the contradictions that any single model will paper over.⁵ The available evidence does not resolve which approach better eliminates hallucinations in practice — no study directly tests whether the prompting techniques advocated by the first position actually achieve the error rates claimed for multi-model orchestration. Users evaluating these tools face a genuine choice between approaches whose comparative effectiveness remains empirically open.

What the record does settle is the baseline responsibility: regardless of which mitigation strategy a user adopts, the legal-sector guidance documents that ABA Formal Opinion 512 holds attorneys fully accountable for AI-generated work product under existing competence and candor obligations, and that at least 36 standing orders from state and federal judges across 13 states required attorneys to disclose AI use, certify citation accuracy, or both, as of mid-2024.² The verification burden does not transfer to the tool.

Choosing by the work you have to defend

The question of which tool to use is ultimately a question about what happens if you are wrong.

For casual inquiry — background research, exploring an unfamiliar topic, drafting a first-pass summary that you intend to revise — the stakes of a hallucinated citation are low. A tool that blends web search with conversational synthesis can surface useful leads even when its citations are imperfect, provided you treat the output as a starting point rather than a finished product. The DataStudios guide to Perplexity makes the optimist case here: the source suggests that disciplined prompting — constraining scope, demanding inline citations after each sentence, asking for named dissenting sources — can reduce hallucinated outputs, and that a breadth-first then depth-focused questioning rhythm is a practical strategy for complex investigations.¹³ Whether that optimism is warranted is genuinely contested. The Perplexity promotional framing assumes the tool's retrieval and synthesis are fundamentally reliable if constrained by the user; it never acknowledges source fabrication as a persistent failure mode. A separate source argues the opposite: that single-model tools cannot debate themselves, will produce internally consistent text that contradicts actual sources without signaling the contradiction, and that cross-validation across multiple models is the most reliable hallucination mitigation strategy available.⁶ The available record does not settle this disagreement — no source provides a controlled test of whether disciplined prompting in a single-model setting achieves citation reliability comparable to multi-model orchestration.

When the deliverable must survive hostile checking, the stakes calculus shifts. In legal work the consequences are not merely reputational. ABA Formal Opinion 512 establishes that lawyers remain fully responsible for AI-generated work product under existing ethical rules covering competence, candor to the tribunal, supervisory duties, and reasonable fees.² Courts have enforced this. The Sixth Circuit's 2026 decision in Whiting v. City of Athens imposed the stiffest available penalty on two Tennessee attorneys who submitted briefs containing 24 fake citations, confirming that AI use creates no special exemption from federal rules.⁶ For legal professionals, a bundled chat subscription is not enough regardless of how well the prompts are crafted. Specialized legal platforms — even those marketed as hallucination-resistant — produce incorrect information in more than 17% of queries (Lexis+ AI) and more than 33% of queries (Westlaw AI-Assisted Research) according to a peer-reviewed Stanford RegLab and HAI study published in the Journal of Empirical Legal Studies.⁶ General-purpose tools perform worse still on legal queries, with hallucination rates that the available evidence suggests reach 58–88% in high-stakes contexts.¹⁰

The practical implication is a tiered framework organized around verifiability rather than brand. For low-stakes exploration, any of the major tools can accelerate orienting research if the user treats citations as leads to be confirmed rather than sources to be cited. For deliverables that will be reviewed, cited downstream, or submitted to any authority, the operative question is not which tool is most accurate but whether every citation the tool produces can be independently verified before the document leaves the writer's hands. In legal work, LeanLaw's six-step checklist — verify basic existence, confirm the holding, check current status, verify context and applicability, cross-reference AI output, document verification — represents the minimum workflow implied by professional responsibility rules.² For academic writing, the CiteME benchmark's finding that LLMs achieve 4.2–18.5% accuracy on citation verification tasks while human annotators reach 69.7% suggests that automated detection remains a supplement to human review, not a replacement for it.¹⁰

The honest bottom line: no current tool provides citations that can be submitted to a client, a court, or a peer-reviewed journal without independent verification of each claim. The tools differ in how transparent they make that verification process — a difference the previous section examined — but they do not differ in whether verification is required. The question of which tool to use is, in the end, inseparable from the question of how much time you have to check it.

How to run this comparison yourself

Any reader who wants to stress-test one of these tools in an afternoon faces a methodological problem before they begin: the available record does not supply a validated, head-to-head protocol for the four tools this report covers. What it does supply is a body of evidence about what breaks and how to see it, from which a practical workflow can be assembled.

The most direct test procedure starts with a fixed question. Commission the same research query in each tool — something with a narrow, verifiable answer and a modest literature — and collect the full output including every citation. From each output, draw a sample of ten cited URLs or bibliographic references and check them against the classification scheme that Rao, Wong, and Callison-Burch operationalized in their preprint: is the URL live? Does a Wayback Machine query confirm it existed at the date of the tool's response? And — a step the URL test alone cannot settle — does the cited source actually say what the tool claims it says?⁹ That third check matters because, as the Verifing guidance notes, a DOI can resolve perfectly while the conclusion attributed to it is nowhere in the paper.³

Hallucination rates are not uniform across those three failure modes, and this creates a real protocol choice. Research on URL validity suggests that hallucinated URLs — links that have no Wayback Machine record and likely never existed — account for 3–13% of citations across the tools tested, while overall non-resolving rates run 5–18%.⁹ GhostCite's archival analysis of published papers found that 1.07% of papers contain at least one definitively invalid citation, with an 80.9% increase in 2025 relative to the prior five-year average.¹¹ The GhostCite benchmark also found that LLMs asked to validate citations themselves achieved only 38% average accuracy — worse than random guessing — which means the tool's own self-check is not a reliable substitute for external verification.¹¹ A ten-citation spot-check conducted by a human against an external database therefore captures something the tool cannot audit on its own.

Two methodological conflicts the available record cannot resolve should be stated plainly. First, hallucination prevalence is genuinely disputed: empirical studies report rates from 14% to 95% across models and tasks, while vendor claims place proprietary systems near 0.1% — a figure resting on self-reported metrics with no independent benchmark corroboration.¹¹ Second, whether LLM self-verification without external databases can substitute for retrieval-based checking remains unresolved: one research line finds metamorphic mutation methods viable for open-domain fact-checking, while two independent lines show that hybrid retrieval pipelines substantially outperform LLM-only approaches on citation-specific tasks, with LLM-alone achieving roughly 38% accuracy.¹¹⁹ For citation checking specifically, the retrieval-based evidence carries more independent support.

What this afternoon's test cannot settle: whether the error rates you observe on your sample generalize across domains, query types, or the tool's next software update. The surviving evidence does not resolve those questions, and no published head-to-head benchmark covers all four tools discussed in this report.

Limitations

• Missing perspective. End-user researchers and knowledge workers actually using deep-research AI tools in real-world research workflows—no first-person accounts of how citation failures are discovered, how verification is actually performed, or what decision-making happens after a hallucination is detected.
• Missing perspective. AI tool developers' internal design reasoning and architectural constraints—sources do not explain why hallucinations persist despite awareness of the problem, what trade-offs between speed/cost/accuracy drive tool design, or what detection mechanisms are deliberately omitted.
• Missing perspective. Academic journal editors and peer reviewers' experiences with AI-generated citations in submitted manuscripts—no data on what fraction of desk-rejected papers contain hallucinated citations, what editorial policies have been adopted, or how peer review processes have adapted.
• Missing perspective. Domain-specific practitioners in non-English-language research, medical practice, legal work in non-US jurisdictions, and STEM fields outside computer science—sources focus heavily on English-language academic CS and US legal contexts; generalization to other domains and languages is unvalidated.
• Missing perspective. Users with limited technical capacity or no access to verification tools—no qualitative data on whether researchers without computational resources, tool subscriptions, or institutional support can realistically adopt citation-verification workflows, or what barriers prevent adoption.
• Coverage gap — Comparison of citation reliability across specific deep-research tools (ChatGPT Deep Research, Perplexity, Gemini Deep Research, Elicit). No source provides empirical head-to-head evaluation of these tools' citation performance. Preprints evaluate general-purpose LLMs (GPT-3.5, GPT-4, Claude) or single-tool systems; none benchmark all four deep-research tools mentioned in the research question.
• Coverage gap — Hallucination detection methods accessible to non-technical users without programming or tool subscriptions. Automated detection tools (SemanticCite, MetaQA, GhostCite) require computational resources or are vendor-locked; no source provides systematic guidance for manual citation verification workflows or comparative analysis of human-auditable verification procedures.
• Coverage gap — Long-term impact assessment—whether citation hallucinations cause downstream harms in citations propagated through secondary works, retracted papers, or policy documents. Sources measure hallucination prevalence and tool performance; none tracks whether hallucinated citations actually mislead subsequent researchers, affect policy decisions, or propagate through citation networks.
• Methodology caveat. Corpus is heavily dominated by vendor-authored product marketing material and blog posts with undisclosed financial interests (NexLaw, Citely, CiteTrue, INRA.AI, DataStudios); these sources cannot provide independent comparative assessment and systematically present their own tools as solutions without rigorous benchmarking.
• Methodology caveat. Empirical preprints and peer-reviewed studies lack independent inter-source replication and use non-overlapping evaluation datasets, benchmarks, and model selections, making synthesis across sources speculative—no common evaluation framework exists across the corpus.
• Methodology caveat. All sources are in English and focus on English-language academic and legal contexts; the corpus cannot characterize hallucination patterns, mitigation effectiveness, or verification tool performance in non-English-language research, non-Western academic publishing, or informal knowledge work.
• Methodology caveat. Temporal coverage is narrow (most sources from 2025–2026) with no historical trend data; cannot assess whether hallucination rates, detection methods, or user adoption patterns have changed over time or will remain stable.
• Methodology caveat. Sources do not distinguish between hallucinated citations (non-existent references) and misrepresented citations (real references deployed to support claims they do not actually make)—the latter is empirically harder to detect and is systematically under-measured in this corpus.

Bibliography

About this report

AI contribution disclosure. Section prose was machine-drafted from the curated source manifest identified above. The operator authored the research question, curated the source corpus, defined the outline, and approved or revised the result; responsibility for the content rests with the operator, not the drafting system.

• Citation integrity: 91.8% [81.6–96.6% 95% CI] (47 supported / 7 partial / 1 unsupported of 55 cited sentences; audited by an independent model from a different provider than the drafter)