Automating Literature Searches: What Works Today

Software can take on four parts of a systematic review search today: suggesting candidate terms, rewriting syntax for another database, flagging duplicate records, and re-running a saved strategy as an alert. A person still designs the strategy, decides it is good enough, and has it peer reviewed. The Cochrane Handbook asks teams to peer review the strategy before it is run and to rerun it close to publication. The Campbell searching guide says text mining can suggest terms, and that the evidence artificial intelligence tools can design a comprehensive strategy is unclear.

Which search tasks can be automated

Each row names the task, the evidence, and the human check.

Task	What software can prepare	Evidence	Human check
Term discovery	Text mining can propose free-text terms from relevant papers, including terms a manual harvest missed, and can give the strategy a starting point.	Campbell section 6.4.3: text mining assists term selection and should not replace the information specialist in identifying keywords. The same section says the evidence that artificial intelligence tools design a comprehensive strategy is unclear, so use them cautiously. Scells, Zuccon, and Koopman (WWW 2019) refined Boolean queries already built for real systematic reviews. On six retrieval measures the refined queries were more effective: precision increased, and recall was maintained or increased.	Accept, edit, or drop every heading and free-text term. PRESS checks subject headings and the text-word search.
Syntax translation	A translator can rewrite field tags and Boolean structure into another platform's syntax.	PRESS 2015 removed "skilled translation of the search strategy to additional databases" from its checklist and left that work to the searcher. Handbook MECIR C33: each database needs its own strategy, and MeSH (MEDLINE) and Emtree (Embase) are not the same vocabulary.	Look up headings in the target thesaurus. Confirm fields, truncation, proximity, and the AND/OR structure before the translated strategy is run.
Deduplication	Reference managers and screening imports can flag records that look identical across databases.	Campbell section 7.3: automated deduplication can remove a unique reference or keep a true duplicate, because the same study is described differently in each database.	Review the suggested matches before anything is deleted. Keep the original export for each database and date.
Re-runs and alerts	A saved strategy on the database platform can send new matching records by email or RSS. PubMed can send that alert weekly, monthly, or in real time (Handbook section 4.4.9).	Handbook section 4.4.10 and MECIR C37, for a Cochrane intervention review: rerun or update searches for all relevant sources within 12 months before publication, and preferably within six months, then screen the new records. PRISMA-S item 12 asks you to report reruns and email alerts.	Screen the new records. Record the date of the last search and any change in syntax or sources.
Peer review	A second searcher reviews the strategy on screen.	PRESS 2015 (McGowan et al., Journal of Clinical Epidemiology, 2016;75:40-46). The evidence behind the guideline suggested that structured PRESS could identify search errors and improve the selection of search terms. Handbook section 4.4.8 strongly recommends peer review before the searches are run.	A second experienced searcher uses the six PRESS elements below, before the strategy is executed.

Khalil, Ameen, and Zarnegar (Journal of Clinical Epidemiology, 2022;144:22-42) reviewed 47 publications on automation in evidence synthesis. Abstract screening has reached maturity, data extraction is still an active area, and methods to semi-automate other steps remain an open research direction. Screening and extraction sit outside this table. Scells and colleagues refined queries an information specialist had already written. Handbook section 4.4.4 still asks Cochrane reviews to combine free text with controlled vocabulary, and it points to the Technical Supplement for text-mining methods.

Testing and peer review

PRESS 2015 kept six elements:

1. Translation of the research question.
2. Boolean and proximity operators.
3. Subject headings.
4. Text-word searching.
5. Spelling, syntax, and line numbers.
6. Limits and filters.

Most respondents in the PRESS survey wanted the review after the MEDLINE strategy was prepared and before translation.

Handbook section 4.4.11: "One simple test is to check whether the search is finding the publications that have been recommended as key publications or that have been included in other similar reviews." Finding only those records can mean the strategy is biased toward known studies. When citation searching or reference checking retrieves relevant papers the database search had already found, and the team had not listed them in advance, that is one sign the strategy is performing. If PRESS elements are missing, the search may not be finished. For a Cochrane review, section 4.4.8 says to note the peer reviewers' names, credentials, and institutions in the acknowledgments, with their permission.

Reporting the search

PRISMA-S (Rethlefsen et al., Systematic Reviews, 26 January 2021;10:39) is a 16-item reporting extension. A log that records the items below is aligned with PRISMA-S. The extension says it is guidance for transparent reporting, and that it is not a tool for judging search quality.

• Item 8. The full strategy for each database, copied as it was run.
• Item 9. Any limits, or a statement that none were used, with the reason for a limit.
• Item 12. How the search was updated (a rerun or an email alert) and any change in sources or syntax.
• Item 13. The date of the last search for each strategy.
• Item 14. Whether the strategy was peer reviewed, and the tool used, such as PRESS.
• Item 16. The deduplication process and any software. The total after deduplication belongs on the flow diagram.

Rethlefsen and colleagues (Journal of Clinical Epidemiology, 2024;166:111229) took a random sample of 100 biomedical systematic reviews indexed in MEDLINE in November 2021, containing 453 database searches. They treated six PRISMA-S items as the reproducibility set: database name, multi-database searching, the full strategy, limits and restrictions, the search date, and the total records. Only 22 searches (4.9%) reported all six. Forty-seven (10.4%) could be reproduced within 10% of the original result count. Six differed by more than 1,000% from the reported count. One review article had the detail needed for every database search to be fully reproducible.

Lin and colleagues (BMJ Open, 30 September 2026;16:e123460) propose a minimum documentation set as a practice proposal, separate from a new reporting standard. Archive the exact strategy and the unedited export straight after each search, before deduplication. After combining records, keep the software, version, and rules, the number removed, and, where feasible, a map from each duplicate to the record kept. For an AI-assisted task, record the tool and model version, the date, the prompts or decision rules, the input fields, any thresholds, the validation sample and its metrics, and every human override.

A worked path, without a counted result

On 2 November 2026 this page was not run against a live database, so it reports no hit count, duplicate count, or translated string.

1. Frame the concepts you will search, using Handbook section 4.4.2: condition, intervention, and study design for a typical intervention review. Eligibility criteria still decide which headings and free-text terms belong.
2. Enter the question in Search Builder. It suggests MeSH terms and title/abstract synonyms. Review each suggestion, then export the PubMed strategy. A curated block can start a concept after you check its scope (Search Block Library). The Search Builder worked example shows an illustrative PICO and labels its strings as examples to adapt.
3. Paste the accepted PubMed strategy into the Search Translator. That page lists PubMed, Embase, Web of Science, Scopus, PsycINFO, CINAHL, Cochrane Library, ERIC, and OVID Medline. Then check headings and fields in the PubMed block translation guide. Save the string that was run, with the platform and the date.
4. Export each database to its own file. Deduplication takes RIS or PubMed/MEDLINE text and returns a deduplicated RIS file with CSV and JSON audit reports. Keep the pre-deduplication archive that page asks for, and read the matches before you accept them (Campbell section 7.3). How match order changes the counts is in deduplicating multi-database search exports.
5. Peer review the strategy with PRESS before you rely on it. Save an alert from that strategy, and rerun close to publication, as in the table above.

Where this fits

The wider pipeline is systematic review technology. Sources outside bibliographic databases need their own log: finding grey literature. Two earlier notes stay opinion: what AI systematic reviews get wrong about automation and where the line is drawn with AI.

Title-and-abstract screening comes later: Study Screener's screening note. Data extraction is Evidence Table Builder's note on automated extraction.

References

1. Lefebvre C, Glanville J, Briscoe S, et al. Chapter 4: Searching for and selecting studies. Cochrane Handbook for Systematic Reviews of Interventions. Version 6.5.1, last updated March 2025. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04
2. MacDonald H, Comer C, Foster M, et al. Searching for studies: a guide to information retrieval for Campbell systematic reviews. Campbell Systematic Reviews. 2024;20(3):e1433. Published 10 September 2024. https://doi.org/10.1002/cl2.1433
3. Khalil H, Ameen D, Zarnegar A. Tools to support the automation of systematic reviews: a scoping review. Journal of Clinical Epidemiology. 2022;144:22-42. https://doi.org/10.1016/j.jclinepi.2021.12.005
4. Scells H, Zuccon G, Koopman B. Automatic Boolean query refinement for systematic review literature search. The World Wide Web Conference (WWW '19). 13 May 2019. https://doi.org/10.1145/3308558.3313544
5. McGowan J, Sampson M, Salzwedel DM, Cogo E, Foerster V, Lefebvre C. PRESS Peer Review of Electronic Search Strategies: 2015 guideline statement. Journal of Clinical Epidemiology. 2016;75:40-46. https://doi.org/10.1016/j.jclinepi.2016.01.021
6. Rethlefsen ML, Kirtley S, Waffenschmidt S, et al. PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Systematic Reviews. 2021;10:39. Published 26 January 2021. https://doi.org/10.1186/s13643-020-01542-z
7. Rethlefsen ML, Brigham TJ, Price C, et al. Systematic review search strategies are poorly reported and not reproducible: a cross-sectional metaresearch study. Journal of Clinical Epidemiology. 2024;166:111229. https://doi.org/10.1016/j.jclinepi.2023.111229
8. Lin L, Xing X, Wu C, et al. Reproducible study identification and selection in systematic reviews: barriers and paths forward. BMJ Open. 2026;16:e123460. Published 30 September 2026. https://doi.org/10.1136/bmjopen-2026-123460