consumer-category-intelligence

An open, evolving experiment in building a reusable human + AI category-intelligence workflow from public data — turning fragmented public signals about a consumer product category into evidence-based marketing hypotheses, with the AI/human split, the evaluation, and the data-safety rules made explicit.

Status: v0.1 — experimental. Two lightweight demos, published to be run, challenged, and improved. Not a finished product. What this is: an open, reusable, still-iterating method experiment, built and shared in public. Public data only, no confidential or company information. What this is not: a market-sizing model, an automated truth engine, or formal investment / product / medical advice.

What's in here (proof, not buried)

• The method → method/ — the 5-phase pipeline.
• Two demos, two jobs → demo/:
  • Robot vacuums = a transferability test (a category I have no expertise in — proves the method is systematic, not knowledge-dependent).
  • Scalp-health shampoo = a source-quality stress test (what happens when public sources are mostly brand blogs and affiliate content — including a real failure the run exposed).
• Evaluation notes → evaluation/ — how I decide what to trust, and a worked example of a failure I kept on purpose. (The most important file here.)

---

TL;DR / So what

Most "AI for marketing" work shows that a person can make AI produce content. This repo shows something rarer: how a marketing operator turns messy public signals into a decision, decides what AI does versus what stays human, and grades how much to trust the result — including where the AI quietly gets it wrong. The method is reusable across categories; the judgment and the evaluation are the point.

1. The business problem

Before a launch, a retailer conversation, or an annual plan, a marketing team has to answer: what is actually happening in this category, and what should we do about it? The raw material — reviews, forums, search signals, AI answers, market reports — is fragmented, uneven, and easy to mistake for fact. This is a structured way to turn fragmented public signal into evidence-based hypotheses and marketing decisions, quickly, without pretending digital signals are market truth.

2. Who this is for

• A brand or channel marketer who needs a category view for a real decision.
• As a reader: anyone evaluating how I think — how I frame a problem, split work between AI and human, and control quality and data risk.

3. How the workflow works

A 5-phase pipeline, driven by one config so every category is analysed the same way and stays comparable over time: 0. Pre-flight — lock the category definition, audience, and key decisions before collecting anything.

1. Collect — several independent public-signal workstreams (search trends, community discussion, expert reviews, AI-answer sources).
2. Build V1 — AI drafts the brief from the collected data.
3. Iterate — human editorial and commercial judgment.
4. QA + ship — fact / integrity check, then data-safety check, then deliver.

Full detail: method/category-intelligence-workflow.md.

4. Demos

• Robot vacuums — transferability test. A category I have no expertise in. Proves the method is a system that runs on anything.
• Scalp-health shampoo — source-quality stress test. A category where public sources are promotional and health claims are unverified. It shows the workflow's limits honestly, and surfaced a real failure (see §9).

5. Where AI helps

AI is strong in the middle of the workflow: gathering and parsing public data, structuring messy notes, clustering themes, producing a fast first draft, flagging unsupported claims, and formatting. It compresses hours of collection into minutes.

6. Where human judgment is required

This is the point of the project. AI does not decide:

• which business question matters, and what "good" looks like;
• what to leave out — the diplomatic and commercial edit;
• correcting local facts the model can't know;
• catching where a fluent AI answer is quietly wrong (see the failure in §9);
• the final sign-off on what to trust.

7. Evaluation method

Output is treated as hypotheses to be graded, not facts: confidence levels (High / Medium / Low), source type (independent vs promotional), triangulation (agreement vs the same source amplified), and evidence gaps (what public data can't show). Worked example, including a real failure: evaluation/evaluation-notes.md.

8. Limitations and data-safety

• Public data only. No confidential, company, retailer, or personal information. Categories and examples are generic and public.
• Not a market-sizing model, a POS/sales replacement, an automated truth engine, or medical/product advice.
• Human-in-the-loop is mandatory. AI runs the checklists; the final integrity and data-safety sign-off is always human.

9. What I learned and improved

Running this for real changed the method:

• Search interest is not market growth. I re-scoped the search-trend question from "is the category growing?" to "how is search interest changing?", and added a separate market-evidence question.
• Market-size numbers can't be taken at face value. On one category, public "2026" figures ranged ~50x, because each source defined the boundary differently. That's why Phase 0 locks the definition first.
• Overlap is amplification, not validation. When AI engines and reviewers cite the same source, that's one source repeated, not two confirmations.
• The AI treated partial matches as full matches. In the shampoo run, with a locked scope (silicone-free · oily · daily · 200–300, local currency), it listed products meeting only some criteria. This exposed the need for a config-validation step before examples are shown. Kept visible on purpose — it's the clearest proof of why the human layer matters.
• Source quality varies wildly by category. Electronics has authoritative testers; personal care is dominated by promotional content. The workflow has to lower its confidence accordingly, not treat every category the same.
• Ship beats polish. The method only got sharper once I actually ran it.

---

Feedback welcome

This is v0.1, published to be challenged. Open a GitHub Issue to challenge an assumption, flag a weak or overconfident claim, suggest a better source, or share a result from running it yourself.

What feedback would be most useful:

• Did the workflow surface something genuinely useful?
• Which claims feel unsupported or overconfident?
• Which workstream is missing?
• Did the method transfer to a category you tried?
• Where did human judgment matter most?

Reuse it for your own category

The method and config are category-agnostic. Copy the config, fill in your category / competitors / markets / sources, follow the 5 phases, and keep the data-safety rules — public signals only, human sign-off on what to trust.

Roadmap (deliberately not in v0.1)

Parked on purpose, to be added through real use, not upfront: a fuller claim-to-source evaluation rubric, a market-and-industry-evidence workstream, structured Google Trends + Reddit collection, a config-validation step, more category demos, and eventual light automation.

Repo map

consumer-category-intelligence/
├── README.md          (this file)
├── method/            the reusable 5-phase workflow
├── demo/              two demos: a transferability test + a source-quality stress test
└── evaluation/        how I decide what to trust, with a worked failure