Examine Evidence

Guide

How to build an evidence map, step by step

16 October 2026 · Examine Evidence team · 5 min read

To build an evidence map, you set a question and a framework, search for studies, import and deduplicate the records, screen them against your criteria, classify each included study on the framework's dimensions, and then read and report the map together with its method.

1Question 2Search 3 · and dedupeImport 4Screen 5Classify 6 · and reportMap
The six steps. Steps 4 and 5 are where most of the work, and most of the judgement, lies.

1. Set the question and the framework

Start with what the map is for and who will use it. Then fix the scope: which populations, interventions or exposures, outcomes, settings and study designs count. PICO is one way to write this down; many maps use other dimensions, such as policy area or type of programme.

The framework is the heart of the map. It names the dimensions (the rows, columns and any filters) and the categories within each, with a one- or two-sentence definition for every category. 3ie, which has produced many maps, consults decision makers and other stakeholders at the start, often through an advisory group, to settle the scope and the framework. Do the same if you can. Write all of this into a protocol before you search.

Search the databases that cover your field, and decide whether you also need grey literature, trial registers or the websites of organisations. For each source, record the date, the exact query and the number of records. PRISMA-S (Rethlefsen and colleagues, 2021) lists what to report about a search.

Examine Evidence can search PubMed, OpenAlex and Semantic Scholar from inside the app and save the results to a library. For other databases, export the results and import the file.

3. Import and deduplicate

Export each search in a structured format such as RIS, with abstracts included. The same study often appears in several databases, so remove duplicates before screening and count how many you removed: the number belongs in your flow diagram.

Examine Evidence flags likely duplicates after an import, by DOI and by title similarity, and you decide which record to keep. Examine Evidence reads RIS, CSV, Excel, BibTeX and JSON files.

4. Screen

Screening decides which records are eligible. Most teams screen titles and abstracts first and, where the question needs it, full texts afterwards. Pilot the criteria on a small set, compare decisions, and sharpen the wording before screening everything.

For systematic reviews, Cochrane's MECIR standards require at least two people working independently on inclusion decisions, with a process for disagreements agreed in advance. Dual screening of titles and abstracts alone is desirable but not mandatory. Many map teams follow the same practice, at least on a sample. If you use AI to help, test it on your own judgements first. The post on validating AI screening shows how.

Screening with several reviewers is part of Examine Evidence's Studio workspace, not of the free quick map. There, a screening question holds the criteria and the exclusion reasons. Reviewers, people or AI agents, can answer blind, and disagreements go to a resolver. Rules on fields and keywords can also exclude records, and every exclusion keeps who or what made it.

5. Classify each study

Classification, often called coding, places every included study in the categories of each dimension. A study can belong to several categories, for example when it measures several outcomes. Use the definitions from the protocol, add examples of borderline cases as you meet them, and keep a log of changes to the codebook.

Check consistency: have a second person, or a second classifier, code a sample and compare. Where they disagree, the definitions usually need work.

In Examine Evidence, the AI suggests dimensions and categories, you edit them, and a decision model classifies each study from its title and abstract. When the decision model fails on a study, the failure is recorded and the study is flagged for review instead of being placed in a guessed category.

6. Build the map

Choose which dimensions form the rows and columns, and which act as filters or as a third dimension. Show counts in the cells, and let readers open a cell to see the studies in it. Keep the total number of included studies visible, since cells overlap.

Pick the layout that answers the question the map was built for. A funder may want interventions against outcomes. A team planning a review may prefer outcomes against study designs, to see where enough trials exist. If a cell holds too many studies to read, add a filter, such as region or population, instead of a new dimension.

7. Read and report

Describe what the map shows in plain terms: which areas are well studied, which are thin, and which are empty in this collection. For every empty cell you mention, say what the search covered. Report the method next to the map:

  • the question, criteria and framework, with category definitions;
  • sources, dates and queries;
  • a flow of records from identification to inclusion, for which the PRISMA 2020 flow diagram is the usual model;
  • how screening and classification were done, by whom, and where AI was used and how it was checked;
  • the date of the map.

Examine Evidence can share a map as a read-only link, a frozen snapshot of the map at the time you share it.

An evidence map is a way to screen information. It is not a systematic review. It shows where studies are and does not appraise or synthesise what they found, so state that in the report.

Sources

Start with a file of studies and a first set of dimensions

Make a map