An autonomous organic-growth engine for SEO, AEO and GEO that learns a site, decides the next action, writes it, checks it and publishes it — our own AI product, in its final stretch.

AI Implementation

Wovio

CASE STUDY CASE STUDY

(01). PROJECT

How do you let a language model publish on a client's website every week without letting it decide what is true?

Wovio is Le Website Tech's own product: an autonomous organic-growth engine for SEO, AEO and GEO. It learns a business's site, Google Analytics and Search Console, decides the single highest-impact action of the week, writes it, runs it through a quality gate and publishes it as a draft in WordPress — then measures the result and plans again. It works in Spanish and English and is in closed beta with the first Le Website clients.

We designed and engineered the whole system: data model, planner, LLM layer, quality gates, evaluation and the WordPress plugin.

Walkthrough – 2026

(01) ABOUT THE PROJECT

What does it take to trust an LLM with a client's site, week after week?

The first decision was to separate retrieval, decision and generation. Everything Wovio knows lives in a fact store in PostgreSQL — each fact with its source, measurement date, time-to-live and confidence, so an expired fact is never used. Planning is arithmetic over those facts, and a recurrence filter written in SQL removes the actions already taken before any candidate is scored by impact, confidence and effort. The model never chooses what to do: it only executes step three of six.

The LLM layer is LLM engineering, not prompt guessing. A provider abstraction covers Anthropic, DeepSeek, Kimi and Qwen; a routing table picks the model tier by action type, on one criterion: what a mistake would cost. A draft nobody has seen goes to a mid-tier model; a rewrite of a page that already ranks goes to the top one. Briefs are built in code under four rules — one brief, one job, one target, zero open decisions — with numeric limits written out, because a cheaper model does not write worse Spanish, it decides worse. Every completion carries its tokens and its real cost, outputs are capped per task, and the ceiling is fifteen dollars per client per month. Retrieval augments each brief with the facts that matter and the three closest neighbour pages, by lexical overlap on purpose: at this scale a vector database would be a crane to lift a chair, and the code marks the one function to change the day a client has thousands of pages.

(02) QUALITY AND GOVERNANCE

Who checks the model's work?

A mechanical gate, not another model: minimum word counts, character ranges per field and a filler-phrase banlist in Spanish and English, because an LLM judging an LLM fails in correlated ways. Governance assigns a risk tier to every action, protects pages already in the top ten, throttles bulk changes and lets each site run in manual mode — every piece waits for approval — or automatic once the client trusts it.

Nothing is recorded without evidence attached to the action, every publication is a WordPress draft first, and everything can be rolled back. The plugin is first-party and uses application passwords, so no client password is ever asked for.

(03) EVALUATION AND INFRASTRUCTURE

How do we know it works before a client pays for it?

Two evaluation harnesses ship with the code: a frozen corpus of twenty real pages scored by the same gate used in production, and a thirty-day simulation of six weekly cycles that confirmed the recurrence filter on all six criteria. A Claude CLI transport lets the engine run on a subscription with no API key, isolated from the operator's tools so results are reproducible.

Underneath: Python 3.12 and FastAPI, SQLAlchemy async on PostgreSQL 16, Alembic, a hard 500-line cap per file enforced by a test, tenant isolation on every query, Argon2id passwords and an encrypted per-tenant credential vault. Deployed with Docker Compose behind Caddy on DigitalOcean, with two timers: Monday measures and re-plans, every day executes.