How our composition data is built

Why ranges, not a single number

The same oil varies with cultivar, origin, soil, season and distillation. A single "linalool: 27%" hides that reality. For every constituent we publish the median across all samples, the typical range (the middle 50% — 25th to 75th percentile), the maximum seen, and the detection rate (how many papers reported it at all). That lets you judge a batch's lab report against what the literature actually shows.

Method

1. Search. One Europe PMC query per botanical species, restricted to open-access articles with full text. Thousands of articles were retrieved and screened down to those carrying real composition tables.
2. Table extraction. Full-text XML was parsed and every table scored for a compound-name column and percentage columns. Retention-index and retention-time columns were excluded; candidate percentage columns had to sum to a plausible total.
3. Species attribution. A sample is kept only on positive evidence — the table caption or its own column header names exactly one species. Tables naming two or more candidates are discarded, not guessed.
4. Name normalisation. Synonyms are unified against PubChem (eucalyptol and 1,8-cineole are one compound; so are d-limonene and limonene; so are 4-terpineol and terpinen-4-ol).
5. Pooling. Study-level medians first, then the median across studies — so one paper publishing many columns cannot dominate an oil's profile.

What we exclude

Tables with very few compounds, columns summing far outside a plausible total, duplicate profiles, molecular-docking and bioassay tables that only resemble composition tables, and non-oil matrices (juices, pectins, nanoemulsions, solvent extracts, biological fluids).

Limitations — read these

• Percentages are GC peak-area normalisations, not mass fractions. They are not concentrations and should not be treated as such.
• This is a literature sample, not a market survey. Oils are represented by research interest, not commercial volume.
• Plant part is often unstated. Where authors specified it, it is recorded — citrus leaf and peel oils differ completely and can share an oil name.
• Absence is recorded as zero. Some authors truncate tables at the top peaks, so a zero can mean "not reported" rather than "not present" — which is exactly what the detection-rate column makes visible.
• Variants are representative. Where one of our oils is a specific origin or cultivar of a species, its page says so: the figures describe the species across the literature, and that batch can sit anywhere within the range.
• No health claims. This is analytical chemistry. We publish what oils are made of, not what they treat.

Provenance & corrections

Every figure traces to a named table in a named paper. Each oil page lists its own sources on-page, and the full machine-readable provenance is open:

• source_articles.csv — every paper with its DOI, licence and Europe PMC link
• evidence_trail.csv — every sample linked to its exact table caption and column header

Measured values are facts and are not themselves subject to copyright; each source's Creative Commons licence is recorded so reuse can be assessed. If a value doesn't match its cited table, the evidence trail links it to the exact source so it can be fixed. Corrections are welcome. Dataset snapshot: 2026-09-19.