AI for Researchers · Visual Deck

Session 2: Literature Discovery & Synthesis

Citation-grounded tools, deep-research agents, and how to check either one

{{Presenter Name}}

Landscape as of August 2026

AI for ResearchersLandscape as of August 20261 / 30

2-Minute Recap

The Series So Far

flowchart LR
  a1(["Capability/Failure Map
Safe · Unreliable · Dangerous"]):::art --> T(["Today — Session 2
we live inside Tier 2 and Tier 3,
and learn to verify what each hands us"]):::today
  a2(["2026 Tool Taxonomy
Tier 1 chatbots · Tier 2 research-specific
· Tier 3 deep-research agents"]):::art --> T
  a3(["CRIT
Context · Role & register ·
Instructions · Task — then iterate"]):::art --> T
  classDef art fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:18px
  classDef today fill:#ab7d22,stroke:#d9b36c,color:#14132b,font-size:19px

Hallucination is structural, not an occasional glitch —
and "ask an AI for references from memory" sits in Dangerous [26]

AI for ResearchersLandscape as of August 20262 / 30

Learning Objectives

By the End of This Session…

1Explain semantic vs keyword searchand why each surfaces literature the other misses
2Compare citation-grounded discovery toolscoverage · grounding · fit — Elicit, Consensus, Semantic Scholar, SCiNiTO, Scite, ResearchRabbit
3Run one question through both kinds of toola citation-grounded discovery tool and a deep-research agent

Objectives 4–6 next · full verbatim wording in the reference deck and curriculum

AI for ResearchersLandscape as of August 20263 / 30

Learning Objectives

…And You Will Also Be Able To

4Evaluate deep-research agent outputcoverage bias · missing paywalled literature · unverified claims
5Apply a citation-verification workflowconfirm a claimed source actually supports the claim made from it
6Identify the caution zonessystematic reviews · coverage bias · paywalled literature

Full verbatim wording in the reference deck and curriculum

AI for ResearchersLandscape as of August 20264 / 30

Section 01

How Search Finds — and Misses — Papers

Two different matching methods, two different blind spots. Knowing which one you are using tells you what you have not seen.

AI for ResearchersLandscape as of August 20265 / 30

Literacy Foundation

Two Ways a Search Engine Can Match

Keyword (lexical)

flowchart LR
  W["the words
you typed"]:::io --> B["BM25
exact matching"]:::blue --> R["same query,
same results, next year"]:::norm
  classDef io fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:19px
  classDef blue fill:#1c1a3f,stroke:#5f82c8,color:#9db4e8,font-size:19px
  classDef norm fill:#23204c,stroke:#8f8cb8,color:#b9b7d6,font-size:19px

classic method: BM25 [1] · Boolean AND / OR / NOT, field limits, exact phrases · deterministic

Semantic (embedding)

flowchart LR
  M["the meaning
of your question"]:::io --> V["vectors — nearest
papers win"]:::hot --> R2["a ranked list —
changes with the model"]:::norm
  classDef io fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:19px
  classDef hot fill:#1c1a3f,stroke:#d9b36c,color:#d9b36c,font-size:19px
  classDef norm fill:#23204c,stroke:#8f8cb8,color:#b9b7d6,font-size:19px

query and papers become vectors [1] · finds synonyms and paraphrases you never thought of · ranked, opaque

AI for ResearchersLandscape as of August 20266 / 30

Literacy Foundation

The Good Tools Run Both, Then Re-Rank

flowchart TB
  H["hybrid search — embeddings + BM25
Consensus' 2025 pipeline"]:::wide --> T1["top 1,500 re-ranked
recency · citation count · journal impact"]:::mid --> T2["top 20 re-scored"]:::narrow --> S(["synthesis"]):::hot
  classDef wide fill:#23204c,stroke:#5f82c8,color:#eceaf8,font-size:19px
  classDef mid fill:#23204c,stroke:#8f8cb8,color:#b9b7d6,font-size:19px
  classDef narrow fill:#1c1a3f,stroke:#8f8cb8,color:#b9b7d6,font-size:19px
  classDef hot fill:#1c1a3f,stroke:#d9b36c,color:#d9b36c,font-size:19px
"a hybrid search approach"Consensus, on its 2025 pipeline — steps 2 and 3 as drawn [1]; in 2026 it rebuilt around "Scholar Agent", a GPT-5-based multi-agent system [35]. The architecture changed; the lesson did not.
AND / OR / NOTSCiNiTO exposes Boolean alongside its semantic search [19]

"Is it semantic or keyword?" is the wrong question. Ask: what did the re-ranker throw away before I ever saw it?

AI for ResearchersLandscape as of August 20267 / 30

Literacy Foundation

What Each One Misses

Keyword misses…

papers using a different vocabulary for your concept
adjacent fields that named the same idea differently
anything past the result cap — only 46.4% of included studies sat in Google Scholar's first 1,000 hits [2]

Semantic misses…

exhaustiveness — it ranks, it does not enumerate
precise Boolean control over population, design or outcome
reproducibility — the ranking is not stable across model updates
AI for ResearchersLandscape as of August 20268 / 30

The Evidence

Measured Head-to-Head: Sensitivity vs. Precision

Missed most — found newit missed most of what belonged, and found included studies the original searches had not [3] — an adjunct, not a replacement [3]
Aug 2026Elicit now ships a different product (Research Agent) [33] — no independent evaluation of it exists yet. That gap is itself the finding.

Four completed evidence syntheses, re-run through Elicit Pro in Review mode, measured Feb–Mar 2025 [3]

AI for ResearchersLandscape as of August 20269 / 30

The Evidence

Before We Get Smug: Human Search Is Biased Too

2.5×humans were 2.5× more likely than the best AI re-ranker to cite a direct collaborator (co-authorship graph) [4]
Not a gold standarda reference list is a record of what an author found and chose [4]

The honest framing: both methods are biased. Use them against each other.

AI for ResearchersLandscape as of August 202610 / 30

Section 02

The Discovery Toolkit

Six citation-grounded tools, described from their own documentation — what each indexes, how each grounds an answer, and what it costs.

AI for ResearchersLandscape as of August 202611 / 30

Tier 2 — Research-Specific Tools

What "Citation-Grounded" Actually Buys You

flowchart LR
  S["SEARCH the scholarly
corpus first"]:::safe --> P["the papers exist
before the prose does"]:::norm --> G["generate second"]:::hot
  classDef safe fill:#1c1a3f,stroke:#2fae74,color:#7fd6a4,font-size:20px
  classDef norm fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:20px
  classDef hot fill:#23204c,stroke:#d9b36c,color:#d9b36c,font-size:20px
what it killsfabricated references — but not misread ones: a real paper can be cited for something it never said [1]
2023 baselineeven when a citation was real, 24–43% carried substantive errors (GPT-3.5 / GPT-4 era) [26]
grounded ≠ correctthe failure moves from "does it exist?" to "does it say that?" — now the dominant failure [31]

Consensus puts it plainly: "We only use AI after we search the scientific literature." [1]

AI for ResearchersLandscape as of August 202612 / 30

Series Artefact

The Literature-Discovery Tool Matrix (1 of 2): Corpus & Grounding

Elicit125M–138M papers [6] [7]
retrieval, then reports with sentence-level citations [6]
extraction tables across up to 1,000 papers [6]
Consensus200M+ documents [1]
hybrid BM25 + embedding search, quality re-rank, then synthesis [1]
fast question-shaped synthesis [1]
Semantic Scholar214M papers · 2.49B citations [12]
scholarly index, not a synthesiser; open API + SPECTER2 embeddings [12]
free non-profit infrastructure others build on [11]
Scite280M+ full-text articles [13]
Smart Citations: supporting · contrasting · mentioning [13]
licensed full text; 30+ publisher agreements [13]
ResearchRabbit310M+ papers [15]
citation-graph expansion from seed papers; no synthesis layer [15]
visual maps of how a literature connects [15]
SCiNiTO500M+ works, on OpenAlex [18]
Boolean + filters, then AI chat that cites its sources [18] [19]
open catalogue base + journal recommender [18]

Corpus figures are vendor-stated, not independently audited · row order is not a ranking — no endorsement implied · full cells in the reference deck

AI for ResearchersLandscape as of August 202613 / 30

Series Artefact

The Literature-Discovery Tool Matrix (2 of 2): Export, Cost & Fit

ElicitFree Basic · Pro $49/user/mo [7]
.bib and .ris; library itself not exportable [8]
extraction tables — measured sensitivity 39.5% (Feb–Mar 2025) [3]
ConsensusFree tier · Premium $11.99/mo [9]
CSV and RIS into EndNote / Zotero / Mendeley [10]
scoping a question fast — not an exhaustive search [1]
Semantic ScholarFree [11]
open API; most endpoints need no key [12]
reliable metadata and citation links — no synthesis [11]
SciteNo free tier · Basic $20/mo · Pro $50/mo [14]
Zotero plugin, API, MCP into chat tools [13]
"has this been contradicted?" — classification is automated [13]
ResearchRabbitFree forever tier · RR+ from $10/mo [16]
BibTeX out; Zotero import is one-way so far [17]
seed-paper expansion — needs a good seed [15]
SCiNiTOFree individual tier · institutional full text [18]
Bookmarks and Spaces; export formats not documented [18]
open-catalogue breadth — inherits OpenAlex's gaps [29]

Pricing and features change monthly — re-check before relying on a cell · undocumented capabilities are stated as such, not guessed

AI for ResearchersLandscape as of August 202614 / 30

Critical Literacy

Read the Corpus Number Critically

Nobody is lying: "works", "papers", "documents" and "sources" count different things. A bigger index is not a better search — the independent head-to-head measured 39.5% sensitivity (Feb–Mar 2025) [3]; vendors' own newer benchmarks are self-graded [33]. All figures vendor self-reported; none independently audited.

AI for ResearchersLandscape as of August 202615 / 30

Section 03

Tier 3: Deep-Research Agents

They plan, browse, read and return a long report with citations. Three audits tell us exactly how far to trust that report.

AI for ResearchersLandscape as of August 202616 / 30

Tier 3 — Deep-Research Agents

What the Agent Actually Does

flowchart LR
  P["plan"]:::n --> S["search"]:::n --> R["read"]:::n --> I["iterate…"]:::n
  I --> S
  I --> REP["cited report
minutes, not seconds"]:::hot
  classDef n fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:19px
  classDef hot fill:#1c1a3f,stroke:#d9b36c,color:#d9b36c,font-size:19px
GoogleDeep Research and Deep Research Max, on Gemini 3.1 Pro — "autonomously plans, executes, and synthesizes multi-step research tasks" [22] [32]
OpenAI"creates a proposed research plan. You can review and modify it" [21] — a GPT-5.2-based model since Feb 2026, not the chat flagship [34]
AnthropicClaude Research (paid plans; model not publicly named) — "multiple searches that build on each other" [23]

The Tier 2 / Tier 3 line is blurring: agents search the open web — and via MCP can call Scite, Elicit and Consensus directly [13] [33] [36]

AI for ResearchersLandscape as of August 202617 / 30

Tier 3 — Deep-Research Agents

Where Agents Genuinely Win

grey literaturethey cross into policy documents and preprints a scholarly index will not surface [21] — and excel at orientation: "what are the camps in this debate?"
0.314Recall@100 for the best agent on ScholarQuest — the ceiling is real [5]
your planthey will read a plan you wrote and follow it — collaborative planning is a real lever [21] [22]
AI for ResearchersLandscape as of August 202618 / 30

The Evidence

Documented Failure Modes

> 94% links validon 14 current models — yet only 39–77% factual accuracy: the failure moved from fake URL to real-URL-unsupported-claim [31]
~42% dropin fact-check accuracy as tool calls scaled 2 → 150 — more retrieval, not more accuracy [31]; agents hallucinate URLs at 10.7% pooled vs 4.8% for search-augmented models [25]
one-sideddeep-research modes stayed "highly one-sided on debate queries" (2025 audit; no newer re-measurement) [24]
AI for ResearchersLandscape as of August 202619 / 30

Critical Literacy

The Same Tool Scored 25% and 69%

The 39.5% headline is the mean of four syntheses — it describes none of them [3]. What counts as a "hit" depends on the review's criteria and the query, not the tool alone [3]. Rule: before believing a headline number, ask what counted as a hit — and how many topics it was averaged over. Every performance figure in this deck is from independent evaluation [3] [24] [25] [31].

AI for ResearchersLandscape as of August 202620 / 30

Section 04

Three Caution Zones

Systematic reviews, coverage bias, and paywalled literature — where these tools need more scrutiny than the interface suggests.

AI for ResearchersLandscape as of August 202621 / 30

Caution Zone 1

Systematic Reviews Have Rules — and They Now Cover AI

§Joint 2025 position statement — Cochrane · Campbell · JBI · CEEuse AI "as long as you can demonstrate that it will not compromise the methodological rigour or integrity" of the synthesis [27]
👁"…should be used with human oversight"and any AI use that "makes or suggests judgements" must be fully and transparently reported [27]
✓The scoping evidence agreesLLMs "should only be used with caution and under human supervision" [28]
AI for ResearchersLandscape as of August 202622 / 30

Caution Zone 2

Coverage Bias: The Index Shapes the Answer

Africa is underrepresented in all three; high-income economies over-represented in Scopus and WoS [29]. Open is not simply worse — OpenAlex's reference coverage is comparable on shared publications [30]. Every AI tool inherits its index's shape. Ask which index, then ask who it leaves out.

AI for ResearchersLandscape as of August 202623 / 30

Caution Zone 3

Paywalls: Most Tools Are Reading Abstracts

4 in 5for the rest, an open tool sees metadata and an abstract — not the methods section. Effect sizes, limitations and the sample live in the full text you were not shown.
licensingScite: agreements with 30+ publishers to read behind the paywall [13] · SCiNiTO's institutional tier: full text via your library [18]

Practical move: route the shortlist through your institutional access before you summarise it.

AI for ResearchersLandscape as of August 202624 / 30

Caution Zone 3

And No, Google Scholar Is Not the Safe Harbour

"Neither GS nor one of the other databases investigated, is on its own, an acceptable database to support systematic review searching." [2] — holding a paper and surfacing it are different problems, for every tool in the matrix.

AI for ResearchersLandscape as of August 202625 / 30

Series Artefact

The Citation-Verification Workflow

flowchart LR
  A["1 · Resolve it
necessary —
no longer sufficient"]:::warn -->
  B["2 · Match it
authors · year ·
journal · title — all four"]:::n -->
  C["3 · Open it
find the sentence
that carries the claim"]:::n -->
  D["4 · Compare it
weaker, narrower
or opposite?"]:::n -->
  E["5 · Check the direction
later contradicted?"]:::n
  classDef n fill:#23204c,stroke:#d9b36c,color:#eceaf8,font-size:18px
  classDef warn fill:#1c1a3f,stroke:#c14b58,color:#ef9a9a,font-size:18px

Step 1: frontier systems now keep link validity above 94% [31] — in 2025 audits 5–18% still failed here [25]. Steps 1–2 catch fabrication; steps 3–5 catch misreading — now the dominant failure: with links valid, factual support runs 39–77% [31], and grounded tools do not remove it [1] [26]. Direction check: [13] [24]

AI for ResearchersLandscape as of August 202626 / 30

Systematic Prompting

CRIT, Applied to Discovery

A discovery tool

One answerable question, with the population, comparison and outcome named. Then constrain with the interface, not the prose.

C — the question, in one sentence  ·  I — year, design and open-access filters  ·  T — one claim at a time, then re-ask

A deep-research agent

Your field, what counts as evidence, a source per claim, and an explicit list of what it could not reach.

C+R — disciplines, audience, register  ·  I — demand disconfirming evidence: audited systems did not volunteer it [24]  ·  T — edit the plan before it runs [21]

— Context  ·  — Role & register  ·  — Instructions & constraints  ·  — Task

AI for ResearchersLandscape as of August 202627 / 30

Live Demo & Takeaways

What We Do Next — and What You Take Home

AOne real question, two tools, side by sidea citation-grounded tool and a deep-research agent
BThree citations pulled from the agent's report — checked liveevery one, through the verification workflow
⌂You leave with: the Tool Matrix · the verification checklist · prompt templates for both tool typesnext session: you have a shortlist — Session 3 turns it into structured notes you can compare
AI for ResearchersLandscape as of August 202628 / 30

References

References (1–18)

  1. Consensus (2025). Welcome to Consensus.consensus.app/home/blog/welcome-to-consensus/ — accessed 2026-07-27
  2. Bramer, Giustini & Kramer (2016). Comparing the coverage, recall, and precision of searches for 120 systematic reviews… Systematic Reviews 5, 39.pmc.ncbi.nlm.nih.gov/articles/PMC4772334/ — accessed 2026-07-27
  3. Lau & Golder (2025). Comparison of Elicit AI and Traditional Literature Searching in Evidence Syntheses Using Four Case Studies. Cochrane Evidence Synthesis and Methods 3(5), e70050.doi.org/10.1002/cesm.70050 — accessed 2026-07-27
  4. Sahu, Charlin & Pal (2026). Rethinking Literature Search Evaluation… arXiv:2605.29234.arxiv.org/abs/2605.29234 — accessed 2026-07-27
  5. Pan et al. (2026). ScholarQuest: A Taxonomy-Guided Benchmark for Agentic Academic Paper Search… arXiv:2606.20235.arxiv.org/abs/2606.20235 — accessed 2026-07-27
  6. Elicit (2026). Elicit: AI for scientific research.elicit.com/ — accessed 2026-07-27
  7. Elicit (2026). Pricing | Elicit.elicit.com/pricing — accessed 2026-07-27
  8. Elicit (2026). Export your data from Elicit.support.elicit.com/en/articles/1153857 — accessed 2026-07-27
  9. Consensus (2026). Pricing — Consensus.consensus.app/home/pricing/ — accessed 2026-07-27
  10. Consensus (2026). How to Export Consensus Results to Reference Managers.help.consensus.app/en/articles/9922811 — accessed 2026-07-27
  11. Semantic Scholar / Ai2 (2026). About Semantic Scholar.semanticscholar.org/about — accessed 2026-07-27
  12. Semantic Scholar (2026). Semantic Scholar Academic Graph API — Overview.semanticscholar.org/product/api — accessed 2026-07-27
  13. Scite (2026). AI for Research | Scite.scite.ai/ — accessed 2026-07-27
  14. Scite (2026). Scite Pricing.scite.ai/pricing — accessed 2026-07-27
  15. ResearchRabbit (2026). ResearchRabbit.researchrabbit.ai/ — accessed 2026-07-27
  16. ResearchRabbit (2026). Pricing | ResearchRabbit.researchrabbit.ai/pricing — accessed 2026-07-27
  17. ResearchRabbit (2026). Using the Zotero Importer.learn.researchrabbit.ai/en/articles/12796541 — accessed 2026-07-27
  18. SCiNiTO (2026). What is SCiNiTO?scinito.ai/help/getting-started/what-is-scinito — accessed 2026-07-27
AI for ResearchersLandscape as of August 202629 / 30

References

References (19–36)

  1. SCiNiTO (2026). Boolean operators (AND, OR, NOT).scinito.ai/help/smart-search/boolean-operators — accessed 2026-07-27
  2. OpenAlex (2026). About OpenAlex.openalex.org/about — accessed 2026-07-27
  3. OpenAI (2026). Deep research in ChatGPT.help.openai.com/en/articles/10500283 — accessed 2026-07-27
  4. Google (2026). Gemini Deep Research Agent — Gemini API.ai.google.dev/gemini-api/docs/interactions/deep-research — accessed 2026-07-27
  5. Anthropic (2026). Using Research on Claude.support.claude.com/en/articles/11088861 — accessed 2026-07-27
  6. Venkit et al. (2025). DeepTRACE: Auditing Deep Research AI Systems… arXiv:2509.04499.arxiv.org/abs/2509.04499 — accessed 2026-07-27
  7. Rao et al. (2026). Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents. arXiv:2604.03173.arxiv.org/abs/2604.03173 — accessed 2026-09-02
  8. Walters & Wilder (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports 13, 14045.nature.com/articles/s41598-023-41032-5 — accessed 2026-07-27
  9. Flemyng et al. (2025). Position statement on AI use in evidence synthesis across Cochrane, the Campbell Collaboration, JBI and CEE 2025. Cochrane Database Syst Rev, ED000178.cochranelibrary.com/cdsr/doi/10.1002/14651858.ED000178/full — accessed 2026-07-27
  10. Lieberum et al. (2025). Large language models for conducting systematic reviews: on the rise, but not yet ready for use. J Clin Epidemiol 181, 111746.jclinepi.com/article/S0895-4356(25)00079-4/fulltext — accessed 2026-07-27
  11. Maddi et al. (2025). Geographical and disciplinary coverage of open access journals: OpenAlex, Scopus, and WoS. PLOS ONE 20(3), e0320347.doi.org/10.1371/journal.pone.0320347 — accessed 2026-07-27
  12. Culbert et al. (2025). Reference coverage analysis of OpenAlex compared to Web of Science and Scopus. Scientometrics.doi.org/10.1007/s11192-025-05293-3 — accessed 2026-07-27
  13. Onweller, H., et al. (2026). Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents. Preprint, arXiv:2605.06635.arxiv.org/abs/2605.06635 — accessed 2026-09-02
  14. Google (2026). Next-generation Gemini Deep Research.blog.google/…/next-generation-gemini-deep-research/ — accessed 2026-09-02
  15. Elicit (2026). Elicit Blog — Research Agent and API/MCP release notes.elicit.com/blog — accessed 2026-09-02
  16. Wikipedia (2026). Deep research (secondary; cites OpenAI documentation).en.wikipedia.org/wiki/Deep_research — accessed 2026-09-02
  17. OpenAI (2026). Consensus: building Scholar Agent on GPT-5 (case study).openai.com/index/consensus/ — accessed 2026-09-02
  18. Consensus (2026). Consensus MCP Server (documentation).docs.consensus.app/docs/mcp — accessed 2026-09-02
1 / 30