How it works
Each step opens on an analysis screen: how many rows are affected, of what kind, and where. You read before you fix, and nothing is written to the corpus until you apply.
Your Scopus exportCSV, as downloadedWhere each answer comes from
GeoNames
You decide: every suggestion is approved or refused by you.
The DOI is what finds the metadata. The metadata is what tells whether a venue is worth looking up. The authors are what build the affiliation grid. And the references only reconcile against a corpus that is finally reliable. Reconciling first would spend API quota on documents that are still wrong.
Find the identifier of documents Scopus exported without one.
Everything downstream keys on the DOI: without it a document cannot be matched against Crossref or OpenAlex, and its references stay raw text.
Fill in year, citations, document type and language.
A missing year removes the document from every time series; a missing type lets an editorial count as a research paper.
Name the publication venues Scopus left blank.
A venue with no name cannot be ranked, compared or counted, and the documents it carries silently drop out of every source-level figure.
Restore authors and map each one to its affiliation.
Author positions are preserved across the three Scopus columns, so a collaboration network can be rebuilt without guessing who wrote with whom.
Resolve institution, city and country for every affiliation.
This is what lets the analysis reach the internal unit and the city, where a raw export stops at the institution name.
Restore titles, abstracts and keywords.
These are the only columns a thematic analysis reads. An empty abstract is a document that no topic model will ever see.
Reconcile every cited reference against Crossref and OpenAlex.
A reference is kept only when corroborated. A single unconfirmed source would fabricate co-citation links that do not exist.
A CSV in Scopus format, with two Bibliominer conventions: author positions preserved across the three author columns, and each affiliation segment carrying its own label.
Empty cells per column, before and after. Articles removed and why. Affiliations consolidated. References Scopus → Global. These are the numbers you quote in a Methods section.
One row per article, one row per cited reference with its provenance. They exist nowhere else, and the API calls behind them have already been spent.
Bibliometric analysis
Cleaning is the means; this is the point. Five families of indicators, each answering one question, and each stating what it does NOT let you say.
What does this corpus contain, and what can it support?
Documents, period, sources, annual growth rate, trajectory over time. The figures that tell whether the corpus is large enough for what you intend to ask of it.
Who produces, authors, institutions, journals, cities, countries
Rankings, output over time, documents by number of signatories. This is the level that only exists because affiliations were resolved down to the city.
What counts, and for whom
Global citations against local ones, the historiograph, RPYS spectroscopy. A paper with 800 global citations and none local matters elsewhere, not here.
What this corpus is about, and how it shifts
Word clouds and treemaps, topics and their position in time, term dynamics. Two raw materials that must never be confused: author keywords and indexed keywords.
Who works with whom, which ideas travel together
Six families of networks, co-authorship, co-citation, bibliographic coupling, co-word. Two safeguards, without which none of them can be read.
The most useful distinction in bibliometrics, and the most misread. Global counts what Scopus counts, all sources together. Local counts how many documents OF THIS CORPUS cite it. A paper with 800 global and 0 local is important elsewhere, not here, the reverse signals foundational work for the community you are studying.
Years, document types, countries, the filter bar applies to every figure on every page. A number read on a filtered page does not describe the whole corpus, and nothing on the figure itself will remind you of that.
The Quality page states, analysis by analysis, whether the field it depends on is filled enough. An h-index computed on a corpus missing 40% of its citations is a number, not a measurement. Start there.