Google Keyword Intelligence

Scoring, clustering, and niche analysis

Keyword scores are transparent weighted averages on a 0–100 scale. The four components are independent and may be unavailable.

Keyword formulas

  • Demand: 100 × log10(volume + 1) / log10(demand_reference + 1), capped at 100. Search demand spans orders of magnitude, so a logarithm preserves the meaningful difference between 100 and 1,000 without letting very large terms swamp everything else.
  • Trend: 50 + 50 × clamp(growth, -1, 1). Flat interest is 50, a doubling reaches 100, and a complete decline reaches 0.
  • Commercial: 100 × min(top_of_page_bid / bid_reference, 1). The high bid is preferred; the low bid is the fallback.
  • Opportunity: `100 × min(impressions / 1,000, 1) × clamp((position - 1) / 29, 0, 1)`. Many impressions combined with a weak position expose more room to improve.

The weights and references live in Settings and can be changed through .gkai.toml or GKAI_... environment variables. Defaults are demand 0.35, trend 0.20, commercial 0.20, opportunity 0.25, demand reference 100,000, and bid reference 5.0.

An unavailable component is excluded from both the weighted numerator and denominator. It is not treated as zero: missing Ads or Search Console access is not evidence of zero demand or zero value. With no available components the score is 0 and confidence is none.

Confidence reports coverage: high means four components, medium three, low one or two, and none zero. Every component also reports its raw value, normalized value, weight, contribution, availability, and explanation.

Trend growth compares the mean of the last quarter of a timeline with the previous quarter. It requires at least eight points and uses values[0] from one Trends response. Values are comparable only inside that response's single normalization_scope; values from different requests must never be compared. The two windows are cut from the timeline by position, before anything is dropped, because a position in this series is a week on the calendar. A week Google marks as having no data arrives with a value of zero; averaging it in would report a collapse that never happened, but discarding it before the windows are cut is worse still — both windows slide backwards and the answer describes an older period while claiming to be the recent trend. A window therefore has to be measured in full to stand for its quarter, and a year whose last months could not be measured reports no trend at all.

The trend component belongs to the run, not to the keyword next to it. Trends is queried once per run, for a single series, and that one growth figure scores every keyword — so each explanation names the series it came from.

Niche factors

Niche analysis separately reports total measurable demand, significant-keyword count, long-tail depth, trend direction, commercial value, the share of demand held by the top five keywords, cluster diversity, and existing-site coverage. Its score is the arithmetic mean of available factors. The factor breakdown is always shown because a lone aggregate would hide which sources were missing and which market characteristic drove the result.

Clustering is deterministic and lexical. It uses normalized-token Jaccard similarity and keeps the implementation boundary open for future embedding, SERP, or LLM clusterers without pretending those exist today.

Limitations that must remain visible

  • Google Trends values are relative interest on a 0-100 scale, not search volume.
  • Google Ads competition describes advertiser demand, not SEO difficulty.
  • A site seed returns keyword ideas Google associates with the site, not the queries the site actually ranks for.

When search-result-page data is absent, SEO difficulty remains unknown. Advertising competition is never substituted for it.

Edit on GitHub · 2026-09-02