Keyword Research

Keyword Research

Check Google search trends and relative interest for keywords. Useful for SEO content strategy, blog-post prioritization, and deciding where to place marketing assets.

This skill uses a driver script at scripts/keyword-research.ts that wraps the google-trends-api npm package. A self-contained reference driver is bundled — adapt the tracked-keyword lists to your own site.

Requirements

• Node.js 20+
• tsx to run the TypeScript driver: npm install -g tsx
• google-trends-api npm package: npm install google-trends-api

Google Trends has no official API and no API key — the package scrapes the public Trends endpoints, so it is rate-limited. Wait a few seconds between requests to avoid temporary blocks.

The bundled driver is TypeScript/Node, but the same data is reachable from Python via the pytrends library if you prefer that ecosystem (pip install pytrends). The CLI surface and result interpretation below are framework-agnostic.

Immediate action

When this skill is invoked, run the keyword research script from the project root:

npx tsx scripts/keyword-research.ts

If the user provides arguments, pass them through:

npx tsx scripts/keyword-research.ts --keywords "AI writing tools, ChatGPT prompts, prompt engineering"
npx tsx scripts/keyword-research.ts --related "AI writing tools"
npx tsx scripts/keyword-research.ts --geo US --top 5

Commands

Compare all tracked keywords

npx tsx scripts/keyword-research.ts

Returns a ranked table of all keywords defined in the driver's tracked-keyword lists, sorted by average Google Trends interest over the last 12 months.

Compare specific keywords

npx tsx scripts/keyword-research.ts --keywords "keyword1, keyword2, keyword3"

Compare up to 15 custom keywords. Google Trends allows max 5 per batch, so the driver compares them in batches.

Related queries for a keyword

npx tsx scripts/keyword-research.ts --related "AI writing tools"

Returns:

• Top related queries: most commonly searched alongside this keyword
• Rising queries: fastest-growing related searches (breakout = >5000% growth)
• Top regions: countries with highest search interest

Filter by country

npx tsx scripts/keyword-research.ts --geo US
npx tsx scripts/keyword-research.ts --geo JP
npx tsx scripts/keyword-research.ts --related "AI writing" --geo US

Use ISO 3166-1 alpha-2 country codes. Omit --geo for worldwide.

Show top N only

npx tsx scripts/keyword-research.ts --top 5

Understanding results

• Average Interest: relative number (0–100) where 100 = peak search interest for that term in the period. NOT absolute search volume.
• 0 interest: the keyword has too little search volume for Google Trends to register. This doesn't mean zero searches — it's below Google Trends' threshold (roughly <1000 monthly searches).
• Comparing across batches: keywords in different batches are normalized to different baselines. The most reliable comparisons are within the same batch (up to 5 keywords).

Limitations

• Google Trends shows relative interest, not absolute search volume. A score of 50 means half the peak interest, not 50 searches.
• Very niche / long-tail keywords often show 0 because they fall below Google Trends' threshold.
• For actual monthly search-volume numbers, you need a paid tool (Ahrefs, SEMrush, DataForSEO).
• Rate limited: wait between requests to avoid being temporarily blocked.

Tracked keywords

Keywords are defined in scripts/keyword-research.ts in the BLOG_KEYWORDS and LANDING_KEYWORDS objects. Update these when adding new blog posts or landing pages so the default (no-args) run reflects your own content inventory.

Ad-hoc queries

For quick one-off checks, query Google Trends directly:

import googleTrends from "google-trends-api";

const result = await googleTrends.interestOverTime({
  keyword: ["ChatGPT custom instructions", "AI writing tools"],
  startTime: new Date(Date.now() - 365 * 24 * 60 * 60 * 1000),
  geo: "US",
});
const data = JSON.parse(result);
console.log(data.default.timelineData);

Known gotchas

• Rate limiting is silent. A blocked request often returns an HTML error page rather than JSON, which surfaces as a JSON-parse error. Back off and retry after a pause.
• Batch normalization confuses comparisons. Two keywords ranked across different batches of 5 are scored against different peaks — never read absolute rank across batch boundaries as exact.
• geo must be a valid ISO 3166-1 alpha-2 code. Invalid codes silently return empty timelines that look like "0 interest".