Every figure sourced, or it never reaches the page
BenchWrite reads the live results for your keyword, pulls the source pages behind them, and writes the article in your own voice. Each number is stored with the source it came from. When a figure contradicts a primary source, it is held back rather than smoothed over.
No card required. Free plan covers two full articles end to end.
A 140-grain ELD Match leaves a 24-inch barrel at , carries a ballistic coefficient, and drops at 500 yards.
Taken from the manufacturer's own published specification, so it is treated as a primary source and used as the baseline every other figure is checked against.
Click any figure to see where it came from. Every number in a finished article has a row like this behind it.
Slop has a measurable shape
Generated text fails in patterns, not at random. A 2026 University of Maryland and Google DeepMind study of machine-generated text (arXiv:2604.03136) reported five structural differences between model and human writing. These are the five, as the paper describes them.
- 01
Over-explanation
The text states a theme it has just demonstrated instead of trusting the reader to infer it: "what this shows is", "in other words", "the key takeaway".
- 02
Low narrative diversity
Generated writing moves in a straight line — define, list, expand, conclude. Human writing doubles back, starts mid-subject, and leaves some threads unresolved.
- 03
Bodily metaphor
"At its core", "the heart of", "the backbone of": models reach for the body to explain abstractions at more than twice the human rate.
- 04
Low specificity
Humans name real products, people, places, prices and dates at roughly twice the model rate. It is countable, which is why our pipeline enforces a specificity floor per thousand words.
- 05
Structural uniformity
Sections of the same length, the same internal shape, the same sentence rhythm, and no position anyone could disagree with.
And we check for more besides
These are our own checks, not findings from the study.
- 01
Thematic repetition
The same three ideas recycled under different headings. We score cross-section overlap and rewrite the sections that repeat.
- 02
Hedged, unfalsifiable claims
"Can vary depending on several factors" survives any fact check because it says nothing. We cap hedges per article and demand a number instead.
- 03
Decorative or blank data
Tables with no populated cells, or figures that never appear in the argument. A cell we cannot source is shown as blank, not filled with a plausible guess.
See what the results already cover
- 01 What every ranking page answers, and in what order
- 02 Entities the top results mention and yours does not
- 03 Coverage gaps worth writing about — the questions nobody answered
- 04 The intent behind the query, which decides the structure
Mirror is a reading tool, not a generator. It tells you what the page has to contain before Authority writes a word.
Nothing publishes with an open question on it
The Pre-Publish tab lists every figure the run could not settle by itself. On our own blog it is a hard block: an open item means the publish button refuses.
- 1
Conflicting figure, unresolved
Two sources give different values for the same claim and neither is a primary source. The run states both and asks you to pick.
- 2
Blank cell left blank
The table guard removed a cell it could not source. You confirm the gap is honest before the piece goes out.
- 3
Baseline the table states about itself
Barrel length, zero distance, altitude — the conditions the numbers assume. Confirm they match the setup you meant.
- 4
Claim without a ledger row
A sentence carrying a figure that has no evidence row behind it. Either it gets a source or it comes out.
Built for subjects where a wrong number matters
Pricing
One credit is the unit of work; a balanced article costs ten. The number on the plan and the number the app meters are the same number.
Free
$0/month
Kick the tyres on one brand.
- 20 credits (2 Balanced articles)
- 1 brand
- Brand DNA + Growth Agent
- Full anti-slop and fact-audit pipeline
Starter
$29/month
For one site you're serious about.
- 250 credits ≈ 25 Balanced articles
- 3 brands
- SERP Mirror + live SERP grounding
- Search Console integration
- Image studio
Most popular · lowest cost per article
Growth
$79/month
For agencies and multi-brand operators.
- 1,000 credits ≈ 100 Balanced articles
- 10 brands
- Bulk writer and bulk images
- Deep quality tier
- Priority queue
Scale
$199/month
High-volume publishing programmes.
- 3,000 credits ≈ 300 Balanced articles
- 50 brands
- Everything in Growth
- API access
- Priority support
Every plan resets to its credit allowance each billing period, with no rollover.
Questions
- Will this rank my page?
- No tool can promise that. Ranking depends on your domain's authority, your link profile, competition for the query and how well the page answers it. What BenchWrite controls is the page itself: accurate figures with sources behind them, a structure that matches the query's intent, and clean technical markup. That is the part you can actually influence.
- Where do the figures come from?
- Live search results and the source pages behind them, plus any documents you add to the brand's knowledge base. Manufacturer and primary sources outrank aggregators, and every figure that survives is stored as an evidence row with its source URL, date and tier.
- What happens when two sources disagree?
- The run keeps the higher-tier figure and records the conflict rather than averaging them. If neither source is primary, the figure is held back and appears on the Pre-Publish checklist for you to resolve before publishing.
- Does it write in my voice?
- It builds a voice profile from writing samples you provide — sentence length, vocabulary, how much you hedge, how you open a section — and drafts against that profile rather than a generic house style. You can edit the profile directly.
- Can it publish to my site?
- Yes. Connect a WordPress site with an application password and BenchWrite uploads the images, rewrites the body against the uploaded URLs, and creates or updates the post with the title, slug, meta description and structured data. Re-publishing updates the same post rather than creating a duplicate.
- Is the output disclosed as AI-assisted?
- That is your call and your legal responsibility in your market. BenchWrite does not strip provenance signals or attempt to defeat AI detection; it improves the writing and verifies the facts. If your jurisdiction requires disclosure, disclose it.