The scariest part of writing a paper isn’t a lack of literature; it’s “thinking you know the literature.” You list a few papers from memory, no one can vouch for whether the citations are right, and related work published in the last two years is easy to miss. Paper Search in the MuchSci app suite puts this step before you start writing: search first, then write, and what comes back from the search has to pass through a series of filters before it makes it into the manuscript.
Let’s put one boundary up front: a search tool can’t change your skill at reading the literature; it changes where you start reading. This article breaks down two things: how a single search actually proceeds, and how search results are filtered step by step until they are “usable.”
01 Search first, then write: how is that different from just starting to write?
▲ Figure 1 | Both produce text; the difference is whether citations have a grounded basis
The implicit assumption behind just starting to write is “I have enough literature in my head,” and this assumption almost always breaks down somewhere: you misremember the year of a citation, miss new work, or can’t assemble a complete, closed set of references. Searching first and writing afterward means acknowledging that memory is unreliable and building the candidate pool on paper first.
Search doesn’t write for you; it gives your writing a literature base you can check—every citation can be traced back to a hit record, which is far more solid than “I remember there was a paper like this.” What the writing stage gets is a small pile of already-filtered literature, not a vague impression.
This approach follows the same logic as doing experiments: first get the data, then draw conclusions. In writing, the “data” is the literature; gather it and filter it first, and only then can your argument have something to stand on.
02 How a single search runs from start to finish
▲ Figure 2 | Step 2 calls the tool to bring back candidates; the remaining steps are all about narrowing down
In the workflow, only step 1 is truly hands-on: figure out what type of literature you need and what period to cover. The rest is calling the tool to bring back candidates, filtering by criteria, reading each paper closely, and putting the usable ones into the library to hand off to writing.The center of gravity of the whole chain is not “finding” but “filtering”—what you find are candidates; only what has been filtered is usable.
The tool-call step is transparent: you can see on the interface what search source was used and how many records were returned. Search is not a black box; afterward you can review why you searched this way and why you filtered this way.
03 The window and the inventory are two different numbers
▲ Figure 3 | The cap and the inventory are on two different scales; a small window does not mean a small library
Let’s get the numbers straight first.What is searchable is the Crossref scholarly metadata database, with about 180 million records, including journal articles, book chapters, and more, covering the 19th century through 2025.The other number is:a single search returns at most 2,000 records.These two numbers are not the same thing—2,000 is the window cap for one search, not the library’s inventory.
Below is the Paper Search interface in MuchSci; you can see that the PaperMcp_paper_search tool call has completed, and candidates are waiting to be filtered:
▲ Figure 4 | The search process: tool call, filtering, hits—every step can be expanded and viewed
Why distinguish these two things? Because your expectations for results should be set by them: an inventory of 180 million means “there is a high probability that relevant literature exists,” while a 2,000-record window means “ranking and filtering strategy matter a lot.” The window is narrowed by query conditions, not by refreshing repeatedly—running the same query over and over returns the same window.
There is also a practical issue: if the literature you are looking for happens to be outside Crossref’s coverage—for example, certain conference papers or preprints—an empty query hit does not mean the literature doesn’t exist; it only means this database doesn’t have it. Search that knows its boundaries is what keeps you from mistaking “not found” for “nonexistent.”
04 From 2,000 records to the few usable ones
▲ Figure 5 | Every step narrows the scope; 2,000 is the starting point, not the endpoint
Narrowing mainly relies on two dimensions:year range, which cuts out literature unrelated to your research period—old 19th-century literature and new 2025 papers are both in the library, but your topic usually covers only one segment of that;citation count, which provides a preliminary ranking of influence so that highly cited papers enter the close-reading pool first.
After these two filters, what remains is at a scale a person can actually read paper by paper. Literature filtered to this point then goes into close reading, enters the library, and is handed off to the writing stage for citation—this is the difference between “literature selected through filtering” and “literature that just turned up”: for the former, you know why every item is there.
The order of filtering also matters: filtering by year first and then by citation count is usually easier than the reverse, because year cuts out irrelevant literature for an entire period, producing the largest change in scale. Literature that enters the library is remembered together with its hit records, so which paragraph you wrote and which paper you cited can both be matched up.
05 Boundaries
As usual, let’s put the boundaries out in the open.
① This article discusses where the feature sits and how it works; numbers are subject to what the interface actually returns. 180 million and 2,000 are figures used to convey scale and do not constitute a promise about the quality of hits in any single search.
② A search hit does not mean the literature is usable. A hit only means “it exists and is relevant”; whether to cite it and where to cite it must be decided after reading each original paper; the author is responsible for the authenticity of citations.
③ Academic norms are subject to the current regulations of your institution. Search tools do not change citation norms themselves, and judgments involving academic integrity cannot be outsourced to a tool.