Tools Bibliometric Analysis
Bibliometric Analysis

How to run a bibliometric analysis from your database export spreadsheet

Bring the spreadsheet from Merge Databases, choose which analyses to generate, and read the keyword clusters the HySCo-KC method builds.

Version
2026.08
Last updated
Available since
Create free account How it works inside

Free credits · No credit card · Instant access What changed recently

What this solves

The file this analysis reads is the spreadsheet Merge Databases gives you, and that is the normal way in. You send your exports there first, from Scopus, Web of Science, PubMed, Dimensions, Lens, RIS or BibTeX, and when the merge finishes a button sends the result straight here, with nothing to download and upload again. A Scopus export on its own also works: it only has to be saved as .xlsx first, because Scopus already names its columns the way the analysis expects. That is the rule behind all of it, the column names in the header rather than the database the file came from. A Web of Science Excel file writes AU and TC where the analysis looks for Authors and Cited by, which is why it goes through Merge Databases too. From there you tick the analyses you want and get the charts, the bibliometric laws and the keyword network. The clustering runs on HySCo-KC, a hybrid method that combines co-occurrence, TF-IDF semantics and topic modelling, with every weight visible and adjustable.

How it works

  1. Bring the spreadsheet from Merge Databases

    The merge result page has a button that sends the spreadsheet straight into the analysis. If you would rather upload by hand, the file has to be .xlsx and its header has to carry the Scopus column names, which a Scopus export already does once you save it as .xlsx. Before starting, we tell you which required columns are missing.

  2. Choose the analyses and the parameters

    Pick from basic charts, Lotka, Bradford and Zipf, the keyword network and the AI readings. Set clusters, keyword count and co-occurrence thresholds.

  3. Read the results and recalculate

    Open each chart, download it as PNG or CSV, and recalculate the network with new weights without uploading the file again.

Method

What happens to your file, step by step, and where the result stops being a decision made by a model.

How it works inside

Reads one .xlsx spreadsheet and canonicalizes the header against 12 fixed column names, correcting case and stray spaces only. Each selected component then runs isolated from the others, so one failure costs its own chart and nothing else. The keyword network follows HySCo-KC: keywords are split on ';', lowercased, hyphens become spaces; the most frequent ones, five times the Keywords setting, go to the model for normalization; the survivors, up to the Keywords count, build three matrices. Co-occurrence becomes Jaccard similarity (an edge only when two keywords appear together at least as often as Minimum Co-occurrences, default 5), TF-IDF over Title plus Abstract (a vocabulary of up to 500 terms of one or two words, each present in at least 2 documents) becomes cosine similarity with up to 0.30 added for shared documents, and an LDA topic distribution is also compared by cosine. The three fuse at alpha 0.25, beta 0.50, gamma 0.25 after being normalized to sum 1; edges below Connection Strength (default 0.15) are cut, and if the graph is still over 80 percent dense the threshold rises to the 75th percentile of the edge weights or the mean plus half a standard deviation, whichever is higher. K-means, with a fixed seed of 42 and the best of 10 starts, partitions the fused matrix; clusters under 3 nodes are absorbed by the neighbor with the highest summed similarity, modularity is computed, and a spring layout runs 150 iterations. Lotka is a frequency-of-frequencies count over accent-stripped author names, Bradford splits ranked journals into 3 zones of roughly one third of the articles each, and Zipf is a table of rank and frequency carrying the log10 of both, not a chart.

Where AI is used, and where it is not

The numbers are deterministic; the model normalizes vocabulary and writes prose. Counts, Jaccard and cosine similarities, the alpha/beta/gamma fusion, K-means, modularity, Lotka, Bradford, Zipf and the research-frontier classification are deterministic arithmetic with a fixed seed of 42, with no model in the path. gpt-4.1-nano enters at four points: it rewrites keyword variants before counting (each proposal passes a sanity check that rejects rewrites too far from the original spelling, and rejected ones are replaced by the rule-based normalizer, and the result is cached for 7 days), it parses affiliation strings into institution, country and coordinates for the map (a rule-based parser stands by, and after a 1200 second budget the remaining affiliations go to that parser), it names and describes clusters that K-means has already formed, and it writes the narrative paragraphs for research frontiers and thematic evolution. gpt-5-nano writes the two optional texts, both built only from results already computed and both charged apart from the analysis: the full report (45 credits) and, on any single table or chart, a paragraph describing what it shows (10 credits each). The one place a model does reach the numbers is keyword normalization: merged variants change frequencies and with them node sizes and cluster membership. Whether a topic counts as hot, emerging, declining or stable is a rule, not a judgment: growth above 0.5 with recency above 0.55 is emerging, growth below -0.3 with recency below 0.35 is declining, and frequency at or above the 75th percentile is hot.

What it accepts

  • One .xlsx spreadsheet whose header uses the Scopus column names, the same layout Merge Databases writes for every other database. A Scopus export downloads as .csv and has to be saved as .xlsx first. Any other extension is refused before processing starts.
  • The header must carry at least 2 of these 9 columns: Authors, Author Keywords, Index Keywords, Title, Source title, Year, Cited by, Affiliations, Abstract. The check reads only the first 5 rows.
  • Each of the 14 components declares the columns it needs. A missing required column removes the component from the run and from the price; a missing optional column runs the component incomplete and charges in full, with a warning naming what was lost.
  • Parameters accepted from the form: Keywords sets how many of the most frequent keywords enter the network, 100 by default and clamped between 20 and 200; Target Clusters defaults to 8, Minimum Co-occurrences to 5 and Connection Strength to 0.15; Keyword Source picks Author Keywords, Index Keywords or both.
  • An upload is capped at 150 MB. There is no separate limit on the number of rows.

What you get back

  • A page with the selected charts, each downloadable as PNG. Lotka, Bradford and Zipf arrive as tables to copy, with no image. The keyword network also exports a CSV of nodes and edges, and the geographic panel exports XLSX.
  • The network returns nodes (label, frequency, cluster, x, y, size between 8 and 40), edges (fused weight plus the co-occurrence, network, semantic and topic components separately) and statistics including modularity and the normalized alpha, beta and gamma.
  • Frontiers returns up to 15 hot, 15 emerging, 15 declining and 10 stable keywords, each with its growth rate, recency and first and last year; thematic evolution returns 3 to 5 periods with Jaccard links between consecutive periods.
  • An optional written report built from the computed results, exportable as Markdown, PDF or Word, and an optional AI paragraph describing any single table or chart.
  • Every run is kept in your history with its own permanent link and can be reopened, or completed later with components not bought the first time.

What this tool does not do

  • It does not read .csv, .ris, .bib or PDF, and it does not fetch records from Scopus, Web of Science or any other bibliographic database. It analyzes only the spreadsheet you upload.
  • Column names are matched by case and padding only. TC, Times Cited or any other alias for Cited by is reported as a missing column, never guessed, because labelling the wrong column is worse than reporting a missing one.
  • Two runs of the same file can differ. Keyword normalization goes through a language model whose output is cached for 7 days; once that cache expires, a different normalization changes frequencies and cluster membership.
  • Research frontier detection needs a span of at least 3 years between the first and last year, and per-cluster trend detection needs 4. Below that the result is empty or everything is returned as stable.
  • Map coordinates come from the model or from a fixed table of country centroids, not from a geocoding service, so institution positions are approximate.
Plan
Pro
Cost
5 to 30 per component, 100 for the full package, 45 for the AI report

What you get

  • HySCo-KC, a hybrid clustering method

    Keyword clusters come from three combined views: co-occurrence network, TF-IDF semantics and LDA topics. You control the weight of each one.

  • Nothing is a black box

    Target clusters, keyword count, minimum co-occurrence, connection strength and keyword source are all editable, with the defaults shown as recommendations.

  • Missing columns are declared upfront

    If your file lacks a column an analysis needs, we say so before starting. You decide whether to continue and pay only for what is generated.

  • Exports ready for the manuscript

    Charts download as PNG, the network exports CSV and the geographic panel XLSX, and the AI report comes as Markdown, PDF or Word. Lotka, Bradford and Zipf are tables you copy, with no image.

Changelog

Every line below is a change that actually shipped, dated by the day it went out.

  1. Latest
    • The AI report is only delivered after the charge succeeds, and nothing is debited on failure.
    • Clicking twice on the same report reuses the first charge instead of paying for it twice.
    • AI cluster naming now requires being logged in with an active plan.
    • The statement line that read 'AI Diarization' now names what it does: descriptions of tables and charts.
Show 5 earlier updates
    • Before running, the tool shows which analyses your spreadsheet cannot support and asks whether to continue.
    • You pay only for the analyses actually produced, and blocked components are no longer offered or charged.
    • One failing analysis no longer ends the whole job, and rankings run without the Cited by column.
    • The result page states which analyses were skipped and what was charged for the run.
    • Spreadsheets whose file name has accents now upload instead of failing after the analysis was charged.
    • History now says when the source spreadsheet was not stored, before you choose a complement.
    • Provider errors no longer print on screen, and you get a short message about what to do.
    • The analysis can run over the screened corpus instead of the whole merged base.
    • Each analysis records whether it ran on the full base or the screened corpus, with counts.
    • Results saved while inside a research project are captured into that project.
    • A failed analysis stops showing as still processing in the project timeline.
    • A long analysis is no longer marked as failed at thirty minutes while it is still running.
    • An analysis stuck in the geographic step now finishes with real data instead of spinning forever.
    • Countries stopped being attributed to the wrong article when the AI response came back truncated.
    • Analyses interrupted by a deploy are recovered, and upload errors show as messages instead of redirects.
    • The AI report follows a scientific article structure and can be downloaded as PDF or Word.
    • The Regenerate button now shows the report's real cost of 45 credits, not 30.
    • Authors written with a hyphen or a space are now counted as the same person.

Questions and answers

Does the AI decide the clusters?

No. Clusters come from K-means over a hybrid similarity matrix, which is arithmetic, not a language model. The AI only writes labels and summaries for clusters that already exist, and normalizes keyword spelling conservatively.

What file does it accept?

An .xlsx spreadsheet, and the easiest one to get is the file Merge Databases produces. The analyses read ten columns: Authors, Title, Year, Source title, Cited by, DOI, Affiliations, Abstract, Author Keywords and Index Keywords, and the names have to match apart from capitals and extra spaces. Document Type and Author full names are accepted in the header, but no analysis uses them today. A .csv is refused, and so are the two-letter tags in a Web of Science Excel file: both go through Merge Databases first.

Does it replace my reading of the corpus?

No. The tool counts, groups and plots what is in the spreadsheet. Which cluster matters, whether a boundary makes sense and what the map means for your question stay with you. The parameters exist so you can disagree.

Is my data used to train models?

No. The AI steps run on paid API tiers whose published policies state that content sent through the API is not used for training and is retained only briefly for abuse monitoring. Your files and results stay in your history.

How to cite this tool

Used it in your research? Here is the reference, already filled in with the version you are looking at and today's access date.

ABNT (NBR 6023)

LESSA, P. W. B. Bibliometric Analysis. Versão 2026.08. [S. l.]: Xplore Dados, 2026. Disponível em: https://xploredados.com/en/tool/bibliometric-analysis. Acesso em: 21 ago. 2026.

APA 7

Lessa, P. W. B. (2026). Bibliometric Analysis (Version 2026.08) [Computer software]. Xplore Dados. https://xploredados.com/en/tool/bibliometric-analysis

BibTeX

@software{xploredados_bibliometric_analysis_2026,
  author  = {Lessa, Patrick Wendell Barbosa},
  title   = {Bibliometric Analysis},
  organization = {Xplore Dados},
  version = {2026.08},
  year    = {2026},
  url     = {https://xploredados.com/en/tool/bibliometric-analysis},
  urldate = {2026-08-21}
}

Start with the free credits

Create an account and test the tools before deciding on a plan.

Create free account