From literature search to a defensible thesis chapter — every number traceable.

Six stages, in the order a thesis actually moves through them. Each one hands the next a version, a setting, or a decision it can point back to.

1. 01

   Search & import

   Pull records from OpenAlex and Crossref inside the app, or import a file export from Scopus (CSV), Web of Science, PubMed, RIS or BibTeX. Combine several sources into one dataset version with cross-source deduplication, so the same paper found twice is counted once.

2. 02

   Check data quality

   A 0–100 corpus score from a published formula — completeness, keyword harmony, affiliation quality, reference integrity and author disambiguation — with a reversible repair ledger for cleaning.

   Q = 0.40 · Completeness + 0.20 · Keywords + 0.15 · Affiliation + 0.15 · References + 0.10 · Disambiguation

3. 03

   Screen with PRISMA 2020

   AI suggests a decision on each record; a person confirms it. Two reviewers can screen independently, with Cohen's kappa reporting how much they agree, and every stage feeds a real PRISMA 2020 flow diagram.

4. 04

   Analyse

   Performance, Lotka, Bradford, annual trends, co-authorship, co-occurrence, co-citation and bibliographic-coupling networks, and the thematic map. For panel data, an estimator advisor reads the diagnostics and recommends a specification instead of leaving the choice to habit.

5. 05

   Write & check

   Manuscript Studio reads a draft chapter for drafting residue and reference problems, and proposes edits you approve or reject one at a time — it never rewrites silently.

6. 06

   Export

   References in BibTeX, RIS or CSV; the PRISMA diagram as SVG, PNG or XLSX; and a methods section written from the settings the study actually used, in English or Arabic, as a .docx.

Claims we can point to code and tests for, and one habit that costs us a cleaner-looking screen.

• Panel-econometrics output has been checked by hand against R's plm reference values on the public Grunfeld benchmark panel.
• Co-citation and bibliographic-coupling networks are checked against hand-computed fixtures with an exact edge-weight match, not eyeballed against a picture.
• AI only suggests: a person confirms every screening decision and every manuscript edit. The only changes made automatically are rule-based repairs with an exact match, such as an author's ORCID or Scopus ID or an institution's ROR record, and each one is recorded.
• Every analysis run stores the dataset version, parameters and engine version that produced it, so a result traces back to exactly what made it.
• When a count cannot be determined, the interface says "not recorded." It does not substitute zero, and it does not substitute a plausible-looking number.

Bibliometra is self-hosted by the organisation running this deployment. A platform administrator chooses one AI provider for everyone and switches on each AI feature individually with its own disclosure. Export or request deletion of your account data from Settings at any time.