Content
How do you publish AI content without inventing statistics?
AI writes a convincing statistic and a plausible citation with equal confidence. What an evidence classifier blocks, and what it lets through.
How MarketWar handles it
Every factual claim in a brief runs through an evidence classifier, and in a regulated context a claim that is AI-inferred or carries no source is BLOCKED from publishing and routed to source validation rather than published with a hedge. 26 output types are supported and not one of them will fabricate a statistic, a citation, a testimonial or a review — the engine is pure and deterministic, with no wall-clock and no randomness, so the same brief always produces the same plan.
Ask any AI tool to write about your industry and you will get a statistic. It will be formatted correctly, attributed to a plausible organisation, and frequently invented.
This is not a bug that will be fixed by a better model. A language model generates the most plausible continuation of your text, and a sentence of the shape "N% of small businesses report X" is enormously plausible — which is why it is the single commonest thing these tools invent. It arrives with identical confidence to a true sentence, which is precisely why it survives your read-through.
Why this is a business risk and not a writing problem
If you are a tradesperson, an invented statistic is embarrassing. If you are in finance, health, law, insurance or anything else with a regulator, it is a compliance incident — and the regulator does not care which tool typed it.
A fabricated citation is worse than a fabricated number, because it is a false attribution to a real organisation.
The evidence classifier
Every factual claim in a brief is classified by the evidence behind it: sourced, AI-inferred, or unverified.
What happens next depends on the risk of the context. In a regulated one, a claim that is inferred or unsourced is blocked from publishing and routed to source validation.
Blocked, not hedged. The tempting middle path is to soften it — "many businesses report" — and that is worse than either option, because it keeps the unsupported claim while removing the one word that would let a reader check it. You end up with prose that is unfalsifiable rather than true.
What it will not produce, ever
Not a statistic it cannot source. Not a citation. Not a testimonial. Not a review. Not a metric.
Those five are listed in the engine's own doctrine because they are the five fabrications that do real damage. An invented testimonial is a fake review, which is a consumer-protection matter in most markets this platform operates in — and it is not a feature anybody should be able to switch on.
26 output types, one control surface
The brief is where the useful work happens, and it takes the things that actually determine whether content lands: brand voice, audience, funnel stage, country and language, claims policy, legal disclaimers, approval gates.
26 output types — from SEO and answer-engine articles to landing pages, comparison pages, case studies and social posts — all planned from that one surface rather than from 26 separate prompts you have to remember to write consistently.
The planning is pure and deterministic: no wall-clock, no randomness. The same brief produces the same plan, which is what makes an approval meaningful. Approving something that would come out differently next time is not approval.
What this costs you
Honesty has a cost here and it is worth naming. The output is sometimes shorter, flatter and less quotable than what a careless tool would hand you, because the three best sentences were the ones with invented numbers in them.
What you get instead is content you can publish in a regulated industry, and a page that still says something specific — which is also the only kind of page that ranks or gets cited by an AI assistant now that the thin, statistic-stuffed version is exactly what the policies demote.
The question to ask
Give any AI content tool a prompt in your own field and ask it for three statistics with sources. Then check them.
The results will tell you more than any vendor comparison. We would rather you ran that test here and found the gap where a number should be.
What this does not do
It cannot verify a claim you have no source for. It can only refuse to publish it, which means the work it hands back is sometimes shorter and more boring than the version a less careful tool would have written for you.
Common questions
Why does AI invent statistics at all?
Because a language model produces the most plausible next words, and a sentence of the shape "N% of small businesses report X" is extremely plausible text. It is generated with exactly the same confidence as a true sentence, which is why it survives a read-through.
What does the classifier actually do?
It sorts each claim by what kind of evidence stands behind it — sourced, inferred, or unverified. What happens next depends on the risk of the context, and in a regulated one an unsourced claim does not publish.
Can I override the block?
You can supply the source. That is the override, and it is the only one, because an override button that bypasses evidence is just a slower way of publishing the same invented figure.
Does it write the whole article?
It plans briefs across 26 output types with the brand voice, audience, funnel stage, market, claims policy and approval gates as inputs. The writing runs through the same gateway as everything else, with its claims checked on the way out.
Related
Why AI assistants never mention your business
Something decides which names an assistant returns for "a plumber near me". What it reads, and why most small business sites are invisible.
What does a website audit actually check?
The twelve things worth measuring on a page, what each one costs you when it fails, and how to tell a real audit from a lead-capture form with a progress bar.
How do you research a competitor without writing knocking copy?
Research naming a rival's weakness is easy to turn into advertising that attacks them, which usually loses. What to do with it instead.