What the diagram is doing
The PRISMA 2020 flow diagram is a running tally of references as they move from your searches to your final included set. Every box is a count, and the counts have to reconcile from top to bottom: what leaves one box has to be accounted for in the next. The two things people get wrong are which count goes in which box, and the switch from records to reports partway down. Both are fixable once you know what each label means.
Identification: databases, registers, and where duplicates go
The top of the diagram is where your raw search results land. PRISMA 2020 splits the sources: records identified from databases and registers on one side, and records identified from other methods (citation searching, hand searching, contact with authors) on the other. Keep those streams separate, because they flow through the diagram as parallel columns.
Duplicates are removed here, in identification, before anything is screened. This is the single most common ordering mistake. The box reads "records removed before screening," and it covers duplicate records plus anything flagged by automation tools or marked ineligible before a human looked. The number that carries down to screening is what is left after deduplication, not the raw hit count. If you searched five databases and got 4,000 hits with 1,200 duplicates, then 2,800 is the number that enters screening.
Screening: records screened, records excluded
Screening is the title and abstract stage. "Records screened" is the post-deduplication number from the box above. You review each record against your inclusion criteria, and "records excluded" is the count you drop at this stage. The subtraction has to hold: records screened minus records excluded equals the records you move forward to retrieve. If you ran dual screening with two reviewers, the diagram still reports one screened count and one excluded count, taken after conflicts are resolved, not two separate tallies.
Retrieval: reports sought, reports not retrieved
This is where the label changes from records to reports. Up to now you were screening records (a title and an abstract). Now you go after the full-text documents, so PRISMA calls them reports. "Reports sought for retrieval" is how many full texts you tried to obtain. "Reports not retrieved" is how many you could not get: no full text available, paywalled with no access, no response from authors. Those not retrieved leave the flow and are subtracted before the next box.
Eligibility: reports assessed, exclusions with counts
"Reports assessed for eligibility" is the full texts you actually read. Here PRISMA 2020 asks for more than a single excluded number: you list each reason a report was excluded, each with its own count. Wrong population, wrong study design, wrong outcome, no extractable data, conference abstract only. The reasons should add up to the total reports excluded at this stage. A reviewer who sees "reports excluded (n = 40)" with no breakdown, or a breakdown that sums to 37, will ask about it.
Included: studies versus reports
The bottom box carries the studies versus reports distinction that PRISMA 2020 added deliberately. One study can appear across several reports: a trial registration, the main results paper, a secondary analysis, a long-term follow-up. If you count reports as if they were studies, your included number is inflated. So the box asks for both: the number of studies included in the review, and the number of reports of those studies. Three reports describing one trial is one study and three reports. When you later run a meta-analysis, you pool at the study level, which is another reason to keep the two counts apart.
The mistakes reviewers actually flag
- Numbers that do not add up. Screened minus excluded should equal sought; sought minus not retrieved should equal assessed; assessed minus excluded should equal included. A break anywhere is the first thing a peer reviewer checks.
- Duplicates in the wrong box. Deduplication belongs in identification, before screening. Putting the duplicate count at the screening stage makes "records screened" wrong and throws off everything below it.
- Exclusion reasons missing or not summing. Listing "n = 40 excluded" without per-reason counts, or reasons that add to a different total, gets queried every time.
- Studies and reports mixed. Reporting the report count as the study count (or vice versa) at the included box, especially when one study spans multiple papers.
Let the counts follow the decisions
Most of these errors are arithmetic, and arithmetic is what software should handle. Covexe builds the PRISMA 2020 (and 2009) diagram straight from your screening decisions: deduplication is recorded in identification, the screened and excluded counts come from your include and exclude votes, retrieval status and eligibility exclusion reasons carry their own counts, and the included box separates studies from reports. A generator that counts the flow from your screening decisions removes the adding-up errors before a reviewer finds them, and you can still override any number by hand. It is free, and it sits in the same project as the screening and full-text stages that produce the counts.