The Semantic Evaluation CSV is RankGear’s uncapped, saved-to-disk export of every keyword variation it evaluated, with the classifier components, evidence, sources, and family assignment behind each one, plus blank columns left for human review.
| Term | Semantic Evaluation CSV |
|---|---|
| Category | Content and Topical Relevance |
| Also known as | Variations Evaluation Export, Labels CSV |
| Where it appears | Variant keywords, and optional Bulk Run exports |
What it means in RankGear
When RankGear works through keyword variations, it does more than list candidate phrases: it classifies each one and groups related phrases into families. The Semantic Evaluation CSV is the complete, uncapped record of that pass. Unlike the trimmed view you see on screen, it keeps every candidate the tool considered, and for each row it carries the classifier component signals, the evidence and sources that supported the label, and the family the variation was assigned to. It is produced from the Variant keywords workflow and can also fall out of a Bulk Run export. The file is written locally, so it stays on your machine as a working artifact rather than a cloud report.
How to interpret it
Read the CSV as an audit trail, not a scoreboard. The point of the export is to show why a variation was labeled or grouped the way it was, so lean on the classifier components and the cited evidence rather than the final label alone. The family assignment tells you which phrases RankGear treats as the same underlying intent, which is what you use to avoid building near-duplicate coverage. The blank human-review fields are there on purpose: they are where an editor records a keep, drop, or merge decision, so the file is meant to be marked up, not just read. Because it is uncapped, expect volume and triage from the top down.
| Column group | What it tells you |
|---|---|
| Variation candidates | Every phrase RankGear evaluated, including ones not surfaced in the UI |
| Classifier components | The sub-signals behind each label, so you can see the reasoning, not just the verdict |
| Evidence and sources | What supported the classification, and where it came from |
| Family assignment | Which intent cluster a variation was grouped into |
| Human-review fields | Blank columns for your keep/drop/merge decisions |
Example
Say you export the CSV for a page targeting “running shoes for flat feet.” The file comes back with several hundred rows: “best running shoes flat feet,” “stability shoes overpronation,” “arch support running shoes,” and many more. Each row shows the component signals that led to its label, the source evidence RankGear drew on, and a family tag such as “flat-feet-stability” that clusters the intent-equivalent phrases together. The Reviewer and Decision columns are empty, so your editor works down the list, merging the near-duplicates and dropping the off-intent phrases before any of it reaches the brief.
Important considerations
- The classifier components and evidence are comparative indicators of semantic relatedness, not Google ranking scores. A strong component value means a variation looks closely related to the target intent, not that including it makes a page rank.
- The export is uncapped and local. Files can be large, and they live on your machine rather than syncing anywhere, so treat them as working artifacts and manage them accordingly.
- The blank human-review fields are intentional. The CSV is designed to be annotated during triage, and its value comes from the decisions your team records in it.
- Use related wording that preserves search intent. Repetition, keyword stuffing, and unsupported claims do not create topical authority, so add genuinely useful coverage rather than forcing terms to hit a density target.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.