One person, three languages, 300 URLs: y8y's editorial system with AI agents

A competitive niche publication needs, at minimum: documented writing, claim verification, translation for each market, technical SEO (markup, hreflang, sitemaps), distribution to search engines and a content map that prevents cannibalization. With a traditional team, that is four or five people before the first reader. The y8y experiment was to ask: how much of that can become a system, leaving the one available person the work that genuinely requires judgment?

The answer was not "use a chatbot to write". It was to design an editorial chain where every link (research, draft, localize, validate, interlink, distribute) has a checkable definition of done, and the AI executes inside that definition.

The central piece is not an AI model: it is the data structure. Every entity (robot, story, comparison, term, answer) is a JSON file with trilingual fields, a search query it owns and an extractable answer. On that graph, AI agents do the competitive research, draft to a written editorial playbook, generate the three language versions and propose the internal links. A battery of automatic validations rejects any invalid entry: data schema, broken references, duplicate queries (the anti-cannibalization guard) and even style rules.

Distribution is also a system, not a chore: the content index is published as a machine-readable file, search engines are notified automatically after every deploy, and the full answers are served to AI engines in a single file. The person stays exactly where they should be: choosing topics, verifying facts against primary sources, signing verdicts and deciding what does not get published. The full editorial method is public in the methodology.

Direct costs: monthly infrastructure is zero (a static site on its host's free tier) and the real spend is AI tool subscriptions plus the person's hours. That is the pattern that matters for a marketing team: the dominant cost of an agent system is not the software, it is the supervision, which is why the entire design aims at making human review fast (automatic validations before, clear diffs after).

Governance in three rules. One: the AI proposes, the person publishes; no content reaches the site without human review. Two: every factual claim carries a cited source, and figures only a manufacturer asserts are labeled as such. Three: there are hard rules the tests make impossible to skip, from the data schema to style. What this case does not claim yet, out of honesty: revenue and consolidated traffic; the publication is young and those numbers will be added here, dated, once they exist. The system that will produce them is what stands documented today.

The order matters more than the tools: each step creates the condition the next one needs. With one person's partial dedication, the full journey takes weeks, not quarters. The full prescriptive version (budget-leak diagnosis, rebuilt team, tools with pros and cons, token economics and risks) is in the content operation blueprint.

A manager's real decision is not which tool to buy but which team model to operate. All three work; they differ in where they place the cost and the risk.

A case that doesn't show its alternatives teaches nothing: any architecture looks inevitable told backwards. The principle that ordered every choice was this: boring, replaceable technology wherever there is no competitive advantage, and custom design only where there is one. In this system the advantage lives in exactly two places, the data structure (the entity graph with its query map) and the editorial playbook with its tests. Everything else was chosen for cost, robustness or easy exit, and every piece has an honest substitute: the static generator could be Hugo or Eleventy instead of Astro; the JSON file store could migrate to a CMS with a UI (Keystatic, Contentful) the day a team writes; the general-purpose agents could be any competent LLM operating the same playbook.

The table summarizes the six structural decisions with the reasonable alternative that was discarded and the real reason. None is dogma: in another context (a large team, high-risk regulated content, a brand without proprietary data) several flip, and the last column says when.

Three real failures, documented because each one rewrote a rule. One: a story shipped with an unverified embedded video (the ID came from a search result and the player had a styling defect); a reader found it as a black box on mobile. Since then no external content gets embedded without prior technical verification, and the player is tested on mobile before publishing. Two: the first draft of the first data study came out thin and one of its key figures, one hundred thousand vacancies attributed to a public agency, did not survive verification: the real source was an employers' association quoted in a press release. It was caught before publication, the study was rewritten around the discrepancy (which turned out to be the most interesting finding) and the primary-source verification pass stopped being optional.

Three: automated distribution to search engines failed silently for days. The cause was not in the code but in a hosting domain configuration, and it only surfaced by examining real behavior from a neutral environment. The resulting rule: after every deploy, the system checks the site's live behavior, not the assumption that the deploy worked. None of the three failures was caught by the AI alone; all three were turned into rules by a person. That is the honest account of operating with agents.

This section is the author's analysis, dated July 2026, and will be revised in public like everything else. The underlying bet: search is splitting into two economies, the click economy and the citation economy. Engines that answer (AI summaries, assistants, agents) don't send visits, they cite sources, and they cite best what they can verify and extract cleanly. That is why every piece of this system has a double output: the page for people and the surface for machines (extractable answers, structured data, the content index, the full answers file following the llms.txt convention). I expect those surfaces to standardize and become as mandatory as the sitemap; whoever has them first accumulates citations while everyone else debates.

Second expectation: readers will increasingly be agents acting on someone's behalf (comparing vendors, preparing a report, verifying a claim). An entity graph with dated, sourced data is exactly what an agent can consume with attribution; a blog of running prose is not. The opportunity I see: small sites structured like this one function de facto as trust APIs for those agents, and that depends not on domain size but on the quality and verifiability of the data. It is the first era of SEO where the small, rigorous operator holds a structural advantage over the big content farm.

Third: the cost of producing content will keep falling toward zero, which means production stops being the business. All the value shifts to what this case calls the account and the governance: proprietary data, verification, named accountability and the published record of hits and failures. What I will add to this system next, in this order: scheduled re-verification of dated claims (the system flags when a cited figure ages), per-surface answers (the same entity answering differently to a search engine, a voice assistant and an agent), and this same graph operating client content. When it happens, it will be documented here with its numbers.