Your first search uses the words you already know.
Suppose you are looking for evidence about appeals against automated decisions. A search for “automated appeals” brings back product pages. An advice service, however, talks about “challenging a decision,” while an agency uses “reconsideration.” These are possible search terms to investigate, not interchangeable legal terms. If you search only with your original wording, you may miss the records you need.
AI browsing tools face a similar problem. BrowseComp tests their ability to find obscure facts that can require several searches.[1] OpenAI describes deep research as a system that plans and carries out searches, then responds to what it finds.[2] These examples help explain why access to a search engine is only part of the task: the next search has to make use of what the previous one revealed.
After opening a useful result, ask two questions: does this help answer my question, and does it show me a better place or term to search? A page that does not answer the question may still name the agency that holds the records or link to the original study. Keep those leads as well as the answers.
A useful result can teach you what to search for next.
List the words, places, and people your search needs to cover.
Before collecting many documents, make a short search plan. Start with words people might use for the issue, including older terms, abbreviations, and other languages. Then list places likely to hold useful evidence: research databases, official records, product documentation, or community organizations. Finally, note whose experience or which period your first searches might miss.
In the appeals example, a vendor can explain how its product works. An agency may publish processing figures. An advice service may describe problems people encounter when challenging a decision. These sources answer different questions. Looking across them helps you avoid mistaking a well-documented product claim for a complete account of the service.
For each part of the plan, write what you hope to find and where you expect to find it. “Look for processing times in agency annual reports” is easier to follow than “research efficiency.” If nothing turns up, you can check whether the record exists, whether access is restricted, or whether you searched under the wrong name.
Use an early search to learn how the topic is described.
At first, read across several kinds of source. Notice recurring terms, authors, organizations, dates, and references. Save a few useful starting documents and their details. You do not need to download every relevant-looking result before you understand which ones could help.
A broad search finds more potentially useful material, but also more irrelevant results. A narrow search reduces that reading burden, but can miss material described differently. Information specialists call these concerns recall and precision. Cochrane’s search guidance describes revising terms in light of the records already found.[5] For an exploratory project, I use the early pass to discover terms, then make later searches more specific. A systematic review still needs the broad coverage required by its protocol.
Write a brief handover note after this first pass. Name the terms you will use next, the sources worth examining, and the gaps you have noticed. For example: “Agency reports use reconsideration; advice pages use challenge. We have service statistics but no accounts from people who abandoned the process.” That note gives the next search a purpose.
Change the search when a result gives you a reason.
A paper may introduce a term that leads you to an archive, which then points to a different explanation. Marcia Bates described this evolving process as “berrypicking”: useful information is gathered in pieces, and each discovery can change the next search.[4] You do not always know the right query at the beginning.
Keep the reason for a change specific. If several relevant agency records use “reconsideration,” try that term alongside the original one. If several reports quote the same study, find that study. If the results concern the wrong country, add the jurisdiction or search its local records. Each change should respond to something you noticed.
Keep earlier search terms beside the new ones, with a sentence explaining the change. This also lets you catch a less useful habit: moving toward results because you agree with them. A good change helps you find missing material, check an original source, or examine an explanation that could challenge your first answer.
Changing a query is progress when you can explain what you learned.
Read the sources behind a useful document—and the work that followed it.
When a report makes an important claim, open the source it cites. Check what was actually studied and whether the report has kept the original limits. Then look for later work that cites the same source. It may repeat the finding, test it elsewhere, criticize it, or correct it. This can uncover material that uses different words from your query.
Following a document’s references is called backward citation searching; finding later documents that cite it is forward citation searching. The TARCiS guidance recommends using citation searching alongside other search methods in systematic reviews, rather than treating one chain of references as complete coverage.[7] A chain starting in one field or language can leave you inside that same field or language.
Choose several relevant starting documents where the question calls for different perspectives. Record the documents, the citation indexes you searched, the dates, and how many rounds you followed. Note why you stopped if further relevant records were still appearing. These are among the reporting details set out in TARCiS.[7] They let someone else see how far the search reached.
Divide up the searching, then review the results together.
When the question has separate parts, people or AI agents can search them at the same time. One might examine official records, another technical evaluations, and another accounts of using the service. Anthropic describes this division of work in its multi-agent research system.[3] It is less helpful when everyone first needs the answer to the same unresolved question, such as which service or definition is in scope.
Give each search a clear task, likely sources, time limit, and a shared way to return results. State any restrictions, such as which records may be accessed. Ask for source links, important passages, unsuccessful queries, and gaps as well as a summary. Otherwise, you may receive several polished answers without knowing what anyone actually checked.
When combining the results, inspect the passages that matter and compare definitions. Trace repeated claims back to their original sources. Three agents may all agree because they found three articles quoting one report. That is one underlying source, and it may still need an independent check.
Leave a short record of decisions that changed the search.
A browser history tells you which pages you opened. It rarely tells you why you changed a term or abandoned an archive. Keep a small decision log as you work: what you noticed, what you changed, and what happened next. A failed search is worth recording when it teaches you something another researcher would otherwise have to rediscover.
Record exact searches and restrictions while they are available; reconstructing them later is difficult.[5] For a formal literature search, ask someone to review the strategy too. The PRESS guideline covers questions such as whether the search matches the research question, whether AND and OR are used correctly, and whether spelling variants or filters accidentally exclude useful records.[6]
The amount of detail should fit the work. For an exploratory search, log decisions that changed what you could find: a new term, a date restriction, a new archive, or a reason to stop. For a systematic review, keep the full search details required by the reporting standard. A short decision log explains the reasoning, but does not replace those records.
Record why you changed direction, while you still remember.
An illustrative decision log. Record changes that affect what you can find; two unproductive rounds are a prompt to review progress, not a fixed stopping rule.
Check what another search is likely to add.
A folder containing twenty reports may hold twenty versions of one claim. Instead of counting documents, ask what the latest searches added. Did they reveal a new group, a different explanation, a contradiction, or evidence that changes your answer? If the results keep repeating what you have already checked, further searching may have less value.
Before stopping, check the places and terms in your plan and name the gaps that remain. Cochrane discusses possible stopping approaches, including whether new terms retrieve additional relevant records; the appropriate approach depends on the kind of review.[5] Repetition in an exploratory search is a reason to assess progress, not a universal stopping rule. Nor does exhausting accessible records make a weak answer strong.
Close with a short account of where you searched, the important changes in direction, and what you could not cover. Say what would reopen the work: access to an archive, a new study, or evidence from a missing group. A reader should be able to tell both why you stopped and what your search leaves unresolved.