← Work
// product · self-hosted, rebrandable

A three-language SEO engine, as a product

The internal version wrote authority content in German, English and French for one company. I rebuilt it, generic, so any small business could point it at their own brand and run it for the cost of a Gemini call.

1config block sets the voice
€0keyword data, by design
3languages, 3 personas

The problem

The internal engine (the case study is here) was tuned to one brand's voice, one domain, one set of internal links. To become a product it had to forget all of that and learn it from a config file instead.

There was also a positioning trap. The market for “AI writes your blog” is saturated and mostly junk. If the product was going to be worth anything, its pitch had to be something the junk couldn't claim.

What I built

A clean-room rebuild driven by a single CONFIG block. The buyer writes their brand voice, three reader personas and their domain once; every prompt, every internal-link lookup and every piece of on-page markup reads from it.

  • One keyword goes in; a German, English and French article comes out, each adapted to a different reader, not translated.
  • Keyword research runs on Google's autocomplete endpoint, which is free and needs no key, so the buyer's only cost is the Gemini call.
  • A bring-your-own-keyword mode, for buyers who already know what they want to rank for.
  • Full on-page package per article: meta description, JSON-LD schema, internal links from the sitemap, and ready-to-paste HTML.
LIVE PRODUCT CONFIG & SCHEMA ADAPTATION SIMULATOR
Test White-Label Brand Voice, 3 Personas & JSON-LD Generation

Select an industry profile below to see how a single config block shapes the Gemini editorial tone, persona focus angles, and automated Schema.org markup generation.

SELECT BUYER INDUSTRY VERTICAL:
// TOP-LEVEL EDITORIAL CONFIG NODE

      
GENERATED ARTIFACT & SCHEMA GEMINI COST: ~0.28 €

// GENERATED JSON-LD SCHEMA.ORG

          
KEYWORD RESEARCH Google Autocomplete (€0 / Unlimited)
OUTPUT ASSETS Markdown, Clean HTML, JSON-LD Schema
LANGUAGES Native DE / EN / FR (3 Personas)
REVIEW GATE Mandatory human editor sign-off

How it works

Operator topic · language KEYWORD API Autocomplete free · zero key ✦ AI ENGINE (GEMINI) Master article Gemini 2.5 Pro voice from CONFIG Adapt EN · FR 3 reader personas Docs & HTML ready-to-paste CONFIG block brand voice · personas On-page package meta · schema · links Human review gate no blind auto-publishing

The buyer edits one block. Everything downstream reads their brand voice, personas and domain from it.

Under the hood

DECISION

One config block owns the whole personality

Brand voice, three personas, domain, internal-link rules: all in one node. The prompts stopped hard-coding a company and started reading from it. That single change is what turned a bespoke workflow into something a buyer can make their own.

TRADE-OFF Rewiring every prompt and every downstream node to a central config was tedious and invisible work. It is also the entire reason the product exists.

DECISION

Free keyword data, on purpose

The keyword step calls Google's public autocomplete endpoint, not a paid SEO API. It returns real search suggestions, for nothing. So the buyer's running cost is just the Gemini call, and the pitch is honest: no hidden subscription bolted on.

TRADE-OFF No search-volume numbers, which a paid API would give. For a small business publishing a handful of articles a month, that data was never worth the monthly bill.

DECISION

Not promising to beat AI detectors

Every competitor in this space promises “100% human, undetectable”. I had already proven, on the internal version, that the detectors contradict each other, the same text scored 82% AI on one and 27% on another. So the product ships a strong, configurable brand voice and human review, and refuses to sell a number I know is noise.

// EMPIRICAL BENCHMARK · THE AI-DETECTOR CONTRACTION
Detector A (Copyleaks)
82% AI PROBABILITY
⚠️ Classifies technical terminology as statistical AI text
Detector B (ZeroGPT / Writer)
27% HUMAN SCORE
✅ Same article, same paragraphs: flagged as authentically human
COMMERCIAL RULE: Promising '100% undetectable AI' is deceptive engineering. Our product optimizes for search intent, factual density, and Schema.org rankability instead.

TRADE-OFF A less flashy sales line. But a claim that falls apart the first time a buyer tests it is worse than no claim.

HONEST

What v1 doesn't do yet

The output is fully trilingual, but the form labels and the orchestration prompts the buyer edits are still in French. A buyer who doesn't read French can run it, but configuring it deeply assumes they do.

NEXT An English configuration pack is the obvious first upgrade, and it is scoped rather than hand-waved.

Results

A self-contained engine that produces three-language articles, each with a full on-page SEO package, for cents per run. Validated end to end on a real topic, then the test instance torn down and the export checked for any trace of the original brand.

What it demonstrates. Product judgement, more than code: choosing a free data source over a billable one, and turning down the dishonest sales claim everyone else in the category makes.

Stack

  • n8n, self-hosted by the buyer: ships as one importable workflow
  • Gemini: the master article and its two persona adaptations
  • Google Autocomplete: keyword expansion, free and unauthenticated
  • Google Drive & Docs: the buyer's own, article plus ready-to-paste HTML
  • A single CONFIG node: brand voice, personas, domain, link rules

What I took away

The hardest product decisions were the ones about what not to build and what not to claim. Picking the free keyword source and refusing the “undetectable” promise both cost me a flashier pitch, and both were right.

Honesty is a feature. It is the one a saturated market can't copy.