Cheap models do the logistics, the best model writes, rival models inspect.
No single model runs an article end to end. Each piece is routed, researched against your own data, written, and then judged by models from other companies before anything reaches your site. This is the whole path, in order, including what happens when a draft fails.
Five stages, and nothing skips a line.
Discovery decides what is worth writing, research decides what the piece will know that its competitors do not, generation writes it, the gauntlet tries to break it, and publishing puts it on your site with the schema and links already in place.
1Discovery
Real demand decides what gets written: keyword and result-page data from DataForSEO, the questions attached to results the site already appears for, and its own Search Console queries once connected. Cheap models do the routing and classification.2Research
The pages currently ranking are read in full, alongside forum threads and the customer's own first-party data: product records, supplier datasheets, the answers support has already written. A large-context model reads the lot and finds what the top ten left out.3Generation
The tier's writer model works from that dossier, outline first and then the article, under the site's brand voice, its terminology and the facts it must never state. Imagery is generated in parallel.4Gauntlet
Four checks run in parallel, none of them by the model that wrote the draft and each including a model from a different company. Any miss returns a specific objection and generation tries again, up to three times.5Publish
What clears every check publishes: internal links into its topic cluster, JSON-LD schema, a push to the CMS, a CDN purge and an IndexNow ping. Takedown, pause and the kill switch stay with you afterwards.
A nightly loop runs underneath all five: it reads Search Console, Bing and GA4, finds pages whose traffic is decaying, queues refreshes, and probes ChatGPT and Perplexity with your buyers' questions to record whether the answer cited you.
Stage four is the gate. Four checks, none of them run by the writer.
They run in parallel against the finished draft. A miss comes back as a specific objection rather than a score, and generation tries again with that objection attached, up to three times. A draft that still fails is abandoned, and it is not charged for.
Fact-check
Every claim in the draft is checked against the research dossier and live sources by a model from a different company than the one that wrote it. Fabrications and contradictions come back as specific objections.Novelty
The draft is scored against the pages currently ranking for the same question. A piece that restates what the top ten already say does not clear the threshold, however well written it is.E-E-A-T
Experience, expertise, authority and trust signals are audited the way a reviewer would: who is speaking, what they have actually handled, and whether the page earns being believed.AI-tell
A read for the patterns that mark text as machine-written: stacked hedges, empty intensifiers, a sentence rhythm that never varies, a conclusion that restates the introduction.
20.8%
of drafts clear every check on the first attempt, with no revision
IT Master ledger, 283 finished runs
55.1%
get there within the three revisions allowed
IT Master ledger, same runs
127
drafts were abandoned rather than published
IT Master ledger, all sites, all time
$0.96
median model spend on a finished Standard run, failures included
IT Master ledger. This is our cost to produce, not your price
The right model for each job, and a vendor spread.
Routing and schema go to a cheap fast model. Research reading and novelty scoring go to a large-context one. Only the writing and the critique go to the frontier models. The spread is not only an economy: it is what lets one company's model audit another's.
Route, classify, write the schema
Anthropic
Claude Haiku 4.5
Read the research, score novelty
Google
Gemini 3.1 Pro
Write the outline and the draft
Anthropic
Claude Haiku, Sonnet, Opus or Fable, by tier
Fact-check the draft, adversarially
Google and OpenAI
Gemini 3.1 Pro with a GPT panelist
Embed and retrieve your first-party data
Google and Voyage
Gemini Embedding, Voyage rerank
What you actually do: connect a site once.
After the connection, the engine chooses the topics, writes, checks and publishes on the cadence your budget sets. The controls you keep are the ones that matter after the fact.
Connect
A WordPress plugin, the @itmaster/sdk package on a Next.js site, a read-only pull API with a per-site key, or a signed webhook into any endpoint you like. One key, once.Describe the business once
Your brand voice, the terminology you insist on, and the facts the engine must never state. Those rules are enforced deterministically before publication, not left to a prompt to remember.Keep the controls
Take an article down and connected sites remove it on their next pull, including anywhere it was reposted. Pause one site, require review before publishing, or switch publishing off entirely.Watch what it did
Search Console and Bing traffic, indexing status per URL, GA4, and the citation probes that ask ChatGPT and Perplexity your buyers' questions and record who the answer named.
Questions people ask before they buy
How long does one article take?
A few minutes end to end. Discovery and research run first, the writer produces an outline and then the draft, and the four validators run in parallel against the finished piece. A draft that has to go back for fixes takes another round, up to three.
What happens when a draft fails a check?
It goes back to the writer with the specific objection: the unsupported claim, the passage that repeats what competitors already say, the sentence that reads as machine-written. It is then rewritten and re-checked. After three failed rounds the draft is abandoned. Failing is the normal case for a share of drafts, and those never reach your site.
Why does a different model do the checking?
Because a model is a poor judge of its own output: it shares the habits and blind spots that produced the draft. Fact-checking and novelty scoring run on models from a different vendor entirely, so a weakness in one company's model does not silently pass through both the writing and the checking.
Where does the first-party data come from?
From what your business already holds and nobody else can publish: product specifications, supplier datasheets, real buyer questions, returns, stock. It is read only, never written to, and it is what makes an article impossible for a competitor to reproduce without the same source.
Does anything publish without me seeing it?
By default an article that clears every check publishes to your connected site, because passing the gauntlet is the gate. You can require review before publishing, pause a site, or switch publishing off entirely, and you can take down anything already live at any time.
See what the first stage would find on your site.
The free check runs the same discovery reading against your address: what Google and the assistants can see today, and the first articles worth writing. No account, no card.