When a Content Network Starts Publishing to Itself

TL;DR

Thorsten Meyer AI reported that a 28-day audit of its 474-site WordPress publishing network found 80% of output concentrated on 38 sites, while 249 sites received no posts. The company attributed the imbalance to both placement logic in DojoClawAI and supply mismatch in Stenvrik, then applied caps, recency ordering, feed changes and scheduler updates.

Thorsten Meyer AI said a 28-day audit of its 474-site WordPress publishing network found that 80% of posts were landing on just 38 sites, while 249 sites received no posts, exposing a distribution failure hidden by healthy overall publishing totals.

The audit covered a network fed by two systems: Stenvrik, described by the company as a news-intelligence layer that ingests feeds, scores and geo-tags stories, and DojoClawAI, a content engine that rewrites stories for individual sites and distributes them across the catalog. According to the report, the system was still publishing at volume, and individual placements appeared valid, but the site-level distribution showed heavy concentration.

Thorsten Meyer AI said the top 38 sites, about 8% of the catalog, carried 80% of total output during the audit window. The company also said the top four sites, all technology titles, were receiving more than 200 articles per week each. At the other end of the catalog, 249 sites, or 53% of the network, received zero posts.

The company attributed the issue to two separate causes. In DojoClawAI, the matcher kept surfacing the same broad technology sites for technology-related stories, while rotation only worked inside the already selected candidate pool. In Stenvrik, the supply mix was skewed: the report said 53% of incoming supplied content was technology or AI-related, while only about 13% of the site catalog belonged to that category.

Why It Matters

The report matters because it describes a failure mode that can occur in automated publishing systems even when output volume, error rates and per-item relevance checks appear normal. If a network only measures how much it publishes, it may miss where the publishing is landing.

For operators of large content catalogs, the case shows how aggregate performance metrics can conceal audience, indexing and editorial problems. Sites receiving hundreds of posts a week may face quality, repetition or crawl-management risks, while dormant sites may lose freshness, search visibility and internal value. The report also shows that distribution problems may sit across separate systems rather than inside a single ranking rule.

WordPress Explained: Your Step-by-Step Guide to WordPress (2020 Edition)

WordPress Explained: Your Step-by-Step Guide to WordPress (2020 Edition)

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Background

Thorsten Meyer AI framed the finding as an engineering note about scale, not as a public product launch. The network includes 474 WordPress sites and uses separate systems for supply selection and placement. That separation shaped the diagnosis: one system influenced what stories entered the pipeline, while the other determined where eligible stories landed.

The company said the fix had three parts. On placement, DojoClawAI added a per-site weekly cap, a network-wide least-recently-used ordering rule and a starvation floor designed to keep idle eligible sites inside the selection path. On supply, Stenvrik audited feeds for liveness, removed feeds that returned HTTP 200 but no items, added verified feeds across non-technology categories and flagged throttled feeds for replacement. On throughput, the scheduler raised fan-out width from five to seven sites and increased quota depth from two to three.

The company said the daily ceiling moved from about 188 posts per day to about 280 posts per day after the changes, a stated increase of 49%. It also said the fix changes future placement behavior and does not retroactively alter the 28-day period in which many sites received no posts.

“The throughput graph looked fine.”

— Thorsten Meyer AI engineering note

“80% of output on 8% of sites”

— Thorsten Meyer AI engineering note

“Not one bug — two independent causes”

— Thorsten Meyer AI engineering note

Mastering GitHub Actions: Advance your automation skills with the latest techniques for software integration and deployment

Mastering GitHub Actions: Advance your automation skills with the latest techniques for software integration and deployment

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

What Remains Unclear

It is not yet clear how quickly the dormant portion of the catalog will receive steady posting after the changes. Thorsten Meyer AI said the proof will come from the next weeks of data, meaning the reported fixes are behavioral changes whose long-term effect has not yet been fully measured. The report also says a previously documented intent of about 950 posts per day was not delivered by the code because of a units issue and remains gated behind a separate sign-off.

Creating a Website: The Missing Manual

Creating a Website: The Missing Manual

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

What’s Next

The next test is whether future audits show lower concentration, fewer dormant sites and healthier posting across non-technology categories. The company identified instrumentation as the real deliverable, indicating that follow-up measurement will determine whether the cap, recency ordering, feed changes and scheduler updates have corrected the imbalance.

Amazon

feed monitoring and auditing tools

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Key Questions

What happened in the publishing network?

A 28-day audit found that most posts in a 474-site WordPress network were concentrated on a small set of sites. Thorsten Meyer AI said 38 sites received 80% of output, while 249 sites received none.

Was this described as a single software bug?

No. The company said the issue came from two separate causes: placement logic that repeatedly favored broad technology sites and a content supply mix weighted toward technology and AI while most sites were in other categories.

What changes were made?

The company said it added per-site caps, network-wide recency ordering and a starvation floor in DojoClawAI; refreshed and rebalanced feeds in Stenvrik; and raised scheduler fan-out and quota depth after adding those controls.

Does the fix prove the problem is solved?

Not yet. The report says the changes affect future placement and that the next weeks of data will show whether dormant sites shrink and distribution becomes more balanced.

Why should readers outside this network care?

The case shows how automated systems can look healthy in aggregate while failing across distribution. It is relevant to publishers, platform operators and anyone relying on automation to route content across a large catalog.

Source: Thorsten Meyer AI