How to Screen 200 Papers Fast Without Missing Key Studies

November 8, 2026

Speed is easy to buy. Recall is not. The risk with any AI screening tool is that you finish in a day and never learn which five papers you skipped.

This protocol is for people who need defensible coverage, not just a fast table. It uses Elicit for the bulk pass and a small manual process for the parts a model handles badly.

Why do AI literature searches miss papers?

The index is the first limit. Elicit draws on Semantic Scholar, roughly 125 million records, which is deep but not complete — book chapters, conference proceedings and non-English journals are thinner than in Scopus or Web of Science.

The second limit is query phrasing. Semantic search matches meaning, so an unusual field term used by one research community may not surface when you use the neighbouring field’s term.

Third: papers with no abstract. Semantic Scholar includes them, but the model can’t judge relevance from a title alone and tends to drop them.

Before trusting any AI result set, build a list of papers you already know are relevant. Five to ten.

  1. Write those papers down before searching.
  2. Run your Elicit question.
  3. Check how many of the known papers appear in the top 50 results.

If you’re at 6 out of 10, your query is missing the vocabulary the field uses. Add the terminology from the missing papers’ titles and rerun. Repeat until you’re at 9 or 10 before doing any extraction.

This takes twenty minutes and is the single best defense against a confidently incomplete review.

How do I use citation chasing alongside Elicit?

Pair semantic search with the graph.

  • Forward chasing — take your three best-scoring papers and look at who cited them. Elicit’s paper view links to citing works in Semantic Scholar.
  • Backward chasing — read the reference lists of two recent reviews in your field.

AI search is good at finding things you didn’t know to look for. Citation graphs are good at finding the canonical papers everyone cites. You need both; neither replaces the other.

How do I document the process for a reviewer?

A supervisor or journal will ask three questions: what did you search, when, and how many records did you screen.

Keep a screening log with one row per query:

Field Example
Date run 2026-10-14
Query text “RCT social media adolescent sleep duration after 2020”
Records returned 312
Screened 120 (first 120 by relevance)
Included 34
Recall check 9/10 known papers found

That table is what makes the process auditable. Reproducibility in semantic search is imperfect — results shift as the index updates — so record the date and the exact query string.

Where must a human take over?

  • Inclusion decisions on borderline papers. A model reading an abstract cannot apply a nuanced eligibility criterion reliably.
  • Effect size extraction. Errors here propagate into meta-analysis and are hard to reverse.
  • Anything published in the last 30–60 days. Indexing lag is real; run a separate Google Scholar alert.

Cost note: Plus at about $12/month covers a moderate screening workload; heavy extraction across hundreds of papers pushes into Pro near $49/month. Budget credits before a deadline, not during one.

Quick Summary

  • Semantic search has recall gaps; run a 5–10 paper recall check before trusting any result set.
  • Combine Elicit with forward and backward citation chasing to catch canonical papers.
  • Log query text, date, counts and recall result for every search — reviewers will ask.
  • Reject model-extracted effect sizes; read those from the paper.
  • Watch for indexing lag of roughly 30–60 days on new publications.
Tags: