How keyword co-occurrence analysis works
Co-word analysis counts how often two terms appear together in the same document. Every pair becomes a link in a network where nodes are keywords and link weight reflects how frequently they co-occur. A clustering algorithm then groups tightly connected keywords, and each group usually corresponds to a topic or research stream.
Co-citation looks backwards through reference lists; keyword co-occurrence looks at what authors are writing about now. That is why it is the most common technique for describing the conceptual structure of a field and tracking how its topics shift, the approach that Callon and colleagues developed as co-word analysis in the 1980s. It is also the easiest map for non-specialist readers to understand, which makes it a natural centrepiece for a review article.
The technique has a known weakness, though. Keywords are chosen by people and by indexing systems, so they carry inconsistency, fashion and vocabulary drift. A sound analysis spends as much effort on preparing terms as it does on running the network, and that preparation is where most of our time goes.
Which terms should feed the network?
The choice of term source is the single biggest decision in a co-occurrence study. Each option has trade-offs:
| Term source | Strengths | Limitations |
|---|---|---|
| Author keywords | Reflect how authors position their own work | Many variants; missing in some older records and journals |
| Indexer keywords (Scopus Index Keywords, Web of Science Keywords Plus) | Broad coverage, assigned systematically | Can be generic; Keywords Plus is derived from titles of cited references |
| Terms extracted from titles and abstracts | Available for nearly every record and capture new vocabulary | Noisy; requires stop-term filtering and a relevance cut-off |
When we use VOSviewer's text-mining mode, we decide between binary and full counting and review the relevance-score cut-off instead of accepting the default mechanically. Where coverage allows, we sometimes build two networks from different sources so you can show that the main themes are robust to that choice.
Term cleaning makes or breaks the map
An uncleaned keyword network almost always splits single concepts across several nodes. We build a dedicated thesaurus for every project and record each merge so you can audit it. Typical operations, shown here as illustrative examples:
- Spelling variants: e-learning, elearning and electronic learning merged into one term.
- Singular and plural: supply chain and supply chains.
- Acronyms: SEM and structural equation modeling, after checking context so the acronym is not confused with another meaning.
- Generic terms: study, research, model, or country names used only as study settings, removed or retained depending on the research question.
Our co-occurrence workflow
Data preparation
We retrieve or receive the records and check what share of them carry author keywords before settling on a term source.
Thesaurus
The full term list is reviewed by frequency, variants are merged and contentless terms removed.
Threshold and counting
Several occurrence thresholds are tested with full and fractional counting, and cluster stability is compared.
Networks and clusters
Network, overlay by average publication year and density views in VOSviewer; trend topics and a thematic map in Bibliometrix.
Cluster tables
Each cluster's keywords are listed with occurrences, total link strength and average publication year.
Reading the indicators and spotting emerging topics
Three indicators appear in every results table: occurrences, link strength to each other keyword and total link strength. The overlay map, coloured by average publication year, separates established vocabulary from newer terms, while the Bibliometrix trend topics chart shows the window in which each keyword was used most heavily.
Very recent keywords deserve caution. A low count might mean a genuinely new topic, or simply a phrase favoured by one small group. We flag such cases so you can weigh them when writing, rather than presenting them as established trends.
Cluster boundaries also depend on two settings that are easy to overlook. Normalisation, by default association strength in VOSviewer, corrects for the fact that very frequent keywords co-occur with almost everything. Clustering resolution controls how finely the network is divided: a higher value yields more, smaller clusters. We test a small range of resolutions and choose the one where clusters are both stable and substantively distinct, then record that choice so it can be reported and reproduced.
What goes into your methods section
Report the term source, how terms were cleaned (with the thesaurus in supplementary material), the threshold and how many terms met it, the counting method, the normalisation and the software version. Our technical report supplies all of it, and in the Advanced package an English methods description is included for you to adapt.
A hypothetical example of the level of detail reviewers expect: author keywords were harmonised using a project thesaurus of 214 merge rules, 96 keywords met a threshold of ten occurrences, and the network was built with fractional counting and association strength normalisation. Every figure in a sentence like this comes straight from our parameter log, so the text you write will match the files you deposit. Interpreting what each cluster means for theory and practice remains your task as the author.
Integrity limits
If a reviewer later questions a merge, you can point to the exact rule and rerun the map without it. Learn the steps in our keyword map workflow, see pricing or order now.