What AI content actually costs, and how often it fails
Live figures from every site the engine runs, read from its own job ledger at 2026-10-17 14:16 UTC: how often a first draft clears every check with no revision, how often it gets there within the three revisions allowed, how many drafts never publish at all, and what each finished run costs, failures included. Nothing here is sampled, and nothing is rounded up.
What the ledger shows
22.1%
of finished runs cleared every check on the first attempt, no revision at all
IT Master ledger, 2026-10-17 14:16 UTC
68.5%
of finished runs cleared every check within the three revisions allowed
IT Master ledger, 2026-10-17 14:16 UTC
127
drafts sent back or rejected, of 403 finished runs
IT Master ledger, 2026-10-17 14:16 UTC
263
articles live, 513,320 words across 7 active sites
IT Master ledger, 2026-10-17 14:16 UTC
12,580
on-page fixes shipped: 8,019 meta patches, 3,553 product and category descriptions, 1,008 rewrites
IT Master ledger, 2026-10-17 14:16 UTC
Aggregated across every site the engine runs. First-pass rate = drafts that cleared every check on the FIRST attempt, with no revision. Eventual pass rate = drafts that cleared every check within the three revisions allowed. Everything else never published. Median cost is per finished run, including the runs that failed. Updated 2026-10-17 14:16 UTC, cached for ten minutes. AI-assistant probes to date: 7,475, of which 974 cited a customer site.
Three numbers nobody selling AI content publishes, and none of them are rounded up.
How often a machine-written draft survives an independent fact-check on the first attempt. How many drafts are rejected outright. What an article actually costs to produce at each quality tier. The figures cover every site IT Master publishes for, aggregated, with no customer named, and they update as the engine runs.
In one sentence
Across all finished runs, 22.1% of AI-written drafts cleared every check on the first attempt with no revision, and 68.5% got there within the three revisions allowed. The remaining 127 of 403 never published at all.
By tier
| Tier | Finished runs | First attempt, no revision | Passed within 3 revisions | Median cost, per finished run |
|---|---|---|---|---|
| standard | 325 | 20.9% | 72.6% | $0.99 |
| lite | 57 | 28.1% | 54.4% | $0.41 |
| pro | 19 | 21.1% | 36.8% | $1.59 |
| premium | 2 | 50% | 100% | $7.30 |
Source: IT Master ledger, 2026-10-17 14:16 UTC.
Pro fails its own drafts more often. That is harsher judging, not worse writing.
Lite uses the smallest writer and the cheapest cross-vendor judges. Standard uses two mid-sized judges per check. Pro uses the strongest models for both writing and judging, and stronger judges reject more, which is why its first-pass rate is lower rather than higher.
The median cost is our spend on models and research tools per finished run, failed runs included because they cost money too, and it is not the price a customer pays: that is fixed per published article on the pricing page. The two rate columns are different questions: how often a first draft is good enough with no revision, and how often the engine gets there within its three attempts.
One run, five stages, and a rejection that costs nothing.
A run is one topic through the full pipeline, from research to a published article or an abandoned draft.
- A run is one topic through the full pipeline: research, outline, draft, four parallel checks, and up to three revision rounds.
- The four checks: a fact-check against the dossier and live sources, a novelty check against the pages already ranking, an E-E-A-T critique, and an AI-pattern scan.
- First-pass means every check cleared on the first draft, with no revision.
- Rejected means the final revision still failed; the draft is abandoned and never publishes.
- Costs are summed from every model and tool call in the run, in US dollars. Runs with no recorded cost are excluded.
Questions people ask
Where do these numbers come from?
They are computed on request from the engine's own job ledger across every site it runs, with no site named. A run counts as finished when the draft either passed every check or was rejected after its final revision. Nothing is sampled or adjusted.
Why is the first-attempt rate so low?
Because the checks are done by models from a different vendor than the writer, and they are told to refute. A first draft that clears a fact-check, a novelty comparison against the pages already ranking, an E-E-A-T critique and an AI-pattern scan on the first attempt is the exception. Most drafts are revised once or twice; some are rejected. We would rather publish this number than a flattering one.
Why does Pro pass less often than Standard?
Pro drafts are judged by stronger models, which are harsher. The tier buys better writing and stricter checking together, so more first drafts are sent back. The sample is also small, and the figure will move.
Is the cost the price a customer pays?
No. It is our median generation cost per article at each tier: the model and tool spend. Customers pay per published article; the price is on the pricing page.
Publish articles that survive this check
Every IT Master article goes through the gauntlet these numbers describe before it reaches your site.