What Did PubMed Actually Search? Review Your Research Agent's Query

A cleaner query can answer a different question

A research agent returns a tidy bibliography and says it improved the search by adding quotation marks. The result count fell, and the titles look more focused. That is not enough evidence to accept the revision. The agent may have changed how the database interprets the topic, not merely removed noise.

This guide is for researchers and developers using agents to assist with literature discovery in PubMed. Its output is a reviewable search record, not a clinical recommendation or a claim of exhaustive coverage. Before accepting the bibliography, ask for both the text the agent submitted and the search PubMed actually performed.

Methodology and the boundary of this guide

We checked the PubMed User Guide, NLM's phrase-search training and its search-history update notice on October 14, 2026. These are primary sources for search behavior and record keeping. The comparison exercise and acceptance rules below are our proposed editorial workflow; we did not run a retrieval benchmark or measure recall.

Our Crossref and OpenAlex retrieval guide discusses choosing discovery services and obtaining documents. This article addresses a narrower problem inside one service: preserving and reviewing the interpretation of each query revision.

Keep entered text separate from translated search

The PubMed User Guide explains that Search Details shows how a query was translated. It also documents an important phrase-search exception: if a quoted phrase is absent from the phrase index, PubMed can ignore the quotes and apply automatic term mapping, displaying a warning. Active sidebar filters can carry into subsequent searches.

The practical consequence is not that quotation marks are bad. It is that punctuation is a search decision with observable effects. A saved prompt saying “search this exact phrase” does not demonstrate that exact-phrase retrieval happened.

Our recommended record gives each attempt its own identifier. Attach the research question, submitted query, displayed translation, visible warnings, active filters, execution time and result count. Keep a copy of the returned identifiers when permitted and within the agreed retrieval scope. Do not overwrite an earlier attempt when the agent proposes a correction.

If Search Details cannot be obtained, record that limitation. An agent should not reconstruct a plausible translation from memory and label it as database output.

Revise one search decision at a time

NLM's phrase-search tutorial covers phrase interpretation and techniques for spelling variations, word order and proximity. Its starting point is that an unqualified multiword search can involve more than the literal phrase. Different syntax serves different retrieval intentions.

For a proposed exploratory search about research-data sharing, ask the agent to separate three intentions before choosing syntax:

• Find discussion of the broader concept, including alternative terminology.
• Find a particular expression used by authors.
• Restrict a term to a specified bibliographic field.

These are proposed search goals, not three interchangeable ways to write the same request. The person responsible for the research should decide which goal serves the question.

Use a revision table with columns for attempt ID, changed decision, reason, expected effect and observed difference. An agent changing quotation, field restriction and date scope simultaneously creates a harder review problem. Prefer one change at a time during diagnosis; retain the final combined strategy separately.

Do not select the winner solely by the smallest count. A narrow search can look reassuring while excluding relevant vocabulary. Conversely, a large result set can contain substantial irrelevant material. The decision requires inspection, not just a numerical comparison.

Compare candidate identities before writing the summary

Here is a proposed acceptance exercise, not a reported experiment. Choose a small set of independently identified papers that the reviewer considers relevant to the research question. Explain why each belongs; do not choose them only from the agent's current results.

Run the proposed query versions under the same declared filter state. For each known paper, record found, not found or not checked. Inspect a bounded sample of records that appear in only one version and write the likely reason for the difference.

Finding every known paper does not prove completeness: the reference set may itself be biased. Its value is diagnostic. If an intended paper disappears after a change, the agent must explain that disappearance before calling the change an improvement.

Keep screening separate from retrieval. “Returned by the query” is not “included in the review.” Save exclusions with reasons, and distinguish an irrelevant record from a relevant record whose full text is not available. Do not make download convenience an unannounced inclusion rule.

Export evidence before the session disappears

NLM's 2025 search-history update documents date stamps in downloaded search histories and notes that display formatting can depend on the software opening the CSV. Preserve the original export rather than relying only on a reformatted spreadsheet screenshot.

Our suggested handoff is a small evidence folder containing the search-history export, a query-revision note, the retained candidate identifiers and the screening decisions. Give each file a clear purpose. The narrative should point to the accepted attempt ID instead of saying only “PubMed was searched.”

Treat a later rerun as a new observation. Preserve the earlier evidence and describe differences; do not quietly replace the old count in the published methods section. A search URL is useful for rerunning a request, but the original record explains what the team observed at the time.

Limitations and the next step

This workflow does not establish that PubMed is sufficient for a particular review, that an inclusion strategy is scientifically sound, or that a retrieved paper supports a clinical claim. Formal evidence reviews may need a librarian, domain specialist and additional databases.

No search examples here have measured precision or recall. The proposed comparison is a way to expose changes for review, not a validated search strategy.

Browse the OpenAgentSkill directory for research workflows, then ask a candidate agent for a query-revision record before accepting its bibliography. The useful deliverable is not just a list of papers. It is enough evidence to explain why this search produced this list.