← All projects
Demo

Case study

keyword search optimizer.

Helps content teams start keyword research from the gap. One seed keyword becomes forty intent-clustered ideas with real Google signals, shown next to what's already published.

Role
Product owner & builder
Type
Internal SEO tool
Companion to
SEO Content Engine
See the demo ↓
Keyword Search Optimizer

The spark

Built the day after the SEO Content Engine, as its research companion: the engine answers "how do we write this", but not "what should we write about". Six weeks later came the second question: are we writing about the right things? That led to a stats view that puts what we've published next to what people search.

The problems to solve

  1. P1

    The content engine could not say what to write about

    Drafting started with no research step to decide which topics were worth writing about.

  2. P2

    LLM brainstorming alone suggests keywords nobody searches

    An LLM on its own produces keywords that sound plausible but that nobody types into Google. Without real search signals, there is no way to tell which ones are worth writing.

  3. P3

    Search data alone is a flat list with no intent

    Raw search data gives a flat list of keywords. It does not say what each searcher wants, so it cannot guide what kind of article to write.

  4. P4

    No view of published work next to what people search

    Six weeks in, a second question came up: are we writing about the right things? Without what had already been published sitting next to what people search, the team could not see which keyword to write next.

How it works

  1. 01

    Brainstorm

    Claude returns 40 ideas per seed keyword, each with an intent label and a one-line rationale.

    SolvesP1

  2. 02

    Enrich with Google data

    Optional SerpAPI enrichment merges in autocomplete suggestions, related searches and People Also Ask, and scores each keyword 0–100: +40 if it appears in autocomplete, +30 if it's a People Also Ask question, up to +30 for how many related searches the seed has.

    SolvesP2

  3. 03

    Cluster & flag

    Keywords group into Informational / Commercial / Transactional / Navigational; anything phrased as a question (what, how, is, does…) is flagged as an FAQ candidate.

    SolvesP3

  4. 04

    Compare with what's published

    A stats tab reads the SEO Content Engine's article sheet, classifies each published focus keyword's intent, and charts intent mix, sources and trend, with caching so repeat runs don't re-classify.

    SolvesP4

Product decisions

SerpAPI is optional, on purpose

The tool has to be useful with just one API key. Search data makes it better but isn't required, and the free tier is only 100 searches a month.

SolvesP1P2

A score anyone can audit

Strength is a published formula over real Google signals. When someone asks "why is this a 70?", the answer is one sentence. The question flag is a cheap rule too, with its one known blind spot written down.

SolvesP2

Fail soft, per tab

If the published-stats sheet can't be read, only that tab shows an error. The research tab keeps working.

SolvesP4

Built with

App
PythonStreamlit
AI & data
Claude (Anthropic SDK)SerpAPIGoogle Sheets API
Output
CSV export (Excel-compatible)

Demo

The tool runs locally in Streamlit, so here is an illustrative slice of a research run: scores follow the real formula, with this seed having 6+ related searches.

Seed: "where to stay in seoul"
KeywordIntentStrengthQuestion?
best area to stay in seoulInformational100n/a
where to stay in seoul first timeInformational70Yes
is gangnam a good place to stay in seoulInformational60Yes
myeongdong vs hongdaeCommercial70n/a
seoul hotels near subway budgetTransactional30n/a
hongdae guesthouse bookingTransactional30n/a

Outcome

Keyword research went from a list to a gap analysis: what people search, what intent it serves, and whether we've already covered it. Export to CSV in one click.

Next projectSEO Content Engine

let's build something. ✦

Happy to talk about products, AI and the calls behind them.

Say hello