Content
Content that answers the question without inventing anything about you.
Most AI writing tools will happily state that you have been in business for twenty years, serve four hundred clients, and won an award. If none of that is true, you have just published a liability. We only write what your knowledge base can support.
Ungrounded
Draft paragraph
With over 20 years of experience and more than 400 satisfied clients, our award-winning team delivers industry-leading results.
Four claims. None of them checkable.
Grounded
Draft paragraph
Cross-border employment contracts must satisfy the posting rules of the destination country. We advise on Italian and German postings; for other jurisdictions we work with local counsel.
Every claim traced to a knowledge-base entry.
How a draft is built
Four steps, and a human at the end of them.
01
Build the knowledge base
We crawl your site, take what you tell us directly, and turn it into discrete claims. Each one carries a confidence level and a source, so the system knows the difference between something you stated and something it inferred.
Artifact: claims with confidence and source
02
Find the questions worth answering
We work out what your buyers are actually asking answer engines, in their language, and drop the ones that only mention competitors or that you have no standing to answer.
Artifact: a ranked question map
03
Write it, grounded
The model commits to which knowledge-base claims it will use before it writes, then writes only from those. Anything it cannot support is left out rather than smoothed over. The finished HTML is sanitised before it goes anywhere near a browser.
Artifact: an article with traceable claims
04
Score it, then let a person decide
Each draft is scored against the rubric below and checked for grounding and valid schema. The score is advice. The decision to publish is always yours.
Artifact: a scored draft awaiting approval
The rubric
We publish the weights we actually score against.
10 factors, 100 points. This table is generated from the same configuration the scorer reads, so it cannot quietly drift from what we do.
We demoted schema markup from 9 points to 5 after a matched study found no meaningful citation lift from adding it. It is still generated and validated, because it is hygiene — it is just not a lever. We also dropped multilingual consistency to zero for want of evidence.
What the research supports
Real effects, smaller than the industry claims.
Visibility lift, best methods
+40%
Citing sources, adding quotations and adding statistics improved visibility by 30–40% on a position-adjusted word-count metric.
Depends entirely on your rank
+115%
Citing sources lifted a page ranked fifth by 115%, and lowered a page already ranked first by 30%. The same tactic helps and harms.
Tactics that survived testing
3
Of 54 method-domain combinations benchmarked independently, only three produced a statistically significant ranking improvement. Most of this field is overclaimed.
Puerto et al., C-SEO Bench, NeurIPS 2025
Questions people actually ask
That is the failure mode the pipeline is built to prevent. Every factual statement has to trace back to a specific claim in your knowledge base, and each claim carries a confidence level. Where the knowledge base has nothing, the draft says so rather than filling the gap. Medium-confidence claims are hedged; low-confidence ones are flagged or dropped.
You do. Nothing publishes automatically. Drafts pass a pre-publication gate that checks grounding and schema validity, then wait for a human decision. A gate result is advice, not permission.
No, it is advisory. We used to treat it as an approval gate and that was the wrong call: a score is a proxy, and optimising a proxy is how you end up with content that scores well and reads badly. Grounding and drift checks are the hard gates.
Never. In the controlled study that established most of this field, keyword stuffing produced no benefit at all, and performed 10% worse than an unmodified baseline against live Perplexity. It is a tactic with a downside and no upside.
Whatever your plan covers, and it is written in that language rather than translated into it. Translated content reads like translated content, and answer engines are asked questions in the buyer's own words.
See what we could honestly say about you.
We build a knowledge base from your site first, so you can judge the grounding before anything gets written.
✓ Free✓ No card✓ Two minutes✓ GDPR
