When AI Invents Facts About Your Business
The sentence you didn't write, in your name, on your site
A draft comes back and it reads well. Then, in the third paragraph, there is a callout fee you have never charged, or a warranty period nobody at your business has ever offered. It is written in your voice. It sits between two sentences that are perfectly true. For anyone who has read a machine-written draft closely, AI content inventing facts about your business is not a theoretical worry — and the invented sentence is often the one that reads most confidently.
The usual advice is to fact-check everything. As advice it is correct. As a process it fails, because "everything" is not a task with an end. Re-reading 1,400 words and interrogating each clause takes longer than writing the piece would have, so it gets done properly once, then skimmed, then skipped.
There is a better arrangement. The tool that wrote the draft knows where it was working from something confirmed and where it was filling a gap. It should say so, in a list, before you read a word of the prose.
Why AI content ends up inventing facts about your business
A language model predicts a plausible continuation of the text so far. When a sentence needs a number, a plausible number completes it as neatly as the correct one. Nothing in the mechanism separates "the figure this business publishes" from "a figure that would be unremarkable for a business like this".
That is why fluency is a poor signal. A sourced sentence and an invented one are written in the same register, with the same steadiness. There is no wobble in the prose to warn you. Model providers' own usage guidance makes the same point from the other direction: outputs are drafts that require human review before they are published.
The gap-filling gets worse the less the tool knows about you. Given only a topic and an industry, a model has no option but to generalise: typical response times, common warranty lengths, the sort of guarantee a competent firm in your trade might offer. Some of that will be right by coincidence. That is not the same as being sourced.
We are not going to attach a rate to this. Published research and provider documentation describe the behaviour clearly enough; the frequency in your drafts depends on your topics and on how much the tool actually knows about your business. Treat it as a category of risk to manage, not a percentage to plan around.
Which facts get invented most often
The errors are not random. They cluster, which is what makes a fast review possible.
- Prices, callout fees and minimum spends. Anywhere a sentence would be more useful with a number in it.
- Guarantees, warranty periods, response times and refund windows. Commitments phrased as reassurance.
- Statistics and percentages attributed to nobody. "Most homeowners…", "studies show…", a figure with no owner.
- Search volumes and ranking claims. Numbers that sound like data because they are shaped like data.
- Awards, accreditations, memberships and years in business. Credibility markers are easy to generate and hard to defend.
- Service areas and opening hours. Often not invented so much as drifted — correct on your site once, changed since, repeated from an old page.
When you review, sweep these categories rather than reading line by line. Numbers, dates, promises, superlatives.
Why a wrong price is the most expensive error to publish
A published price becomes an expectation. Someone reads it, decides you are affordable, and arrives already anchored. Your staff then spend the first minute of the conversation arguing against something your own website said. That cost lands every time the page is read, not once.
Advertising standards and consumer protection guidance generally treat misleading price and guarantee claims as the responsibility of the business doing the advertising. Who or what composed the sentence is not much of a defence. The name on the page carries it.
There is also a difference in kind between error types. A wrong statistic embarrasses you: you correct it, and it looks like carelessness. A wrong guarantee can commit you. If your site says you will be there within a certain window, or refund within a certain period, a customer is entitled to take you at your word — and reasonably annoyed when you cannot.
Corrections also travel slower than the original page. Search caches, screenshots, quotes on forums and in review threads: the wrong figure keeps circulating after you have fixed it. This is why catching it in review is worth so much more than catching it later. Pre-publication, it costs you a deleted sentence.
Guessing from general knowledge versus working from a business profile you confirmed
The fix is not a better writing style. It is a source of truth.
Kwerivo crawls a customer's public pages and builds a business profile — what you sell, who you sell it to, the words you use for it — and you correct and confirm that profile before anything else is built on it. The crawl respects robots.txt. It does not sign in, submit forms, or reach non-public pages. What it can read is what any visitor can read.
Drafts are then written from that confirmed profile rather than from an impression of your industry. The rule that matters most for the fear described above: Kwerivo will not state a price, guarantee or specific figure unless that figure is confirmed in the business profile. Where the profile is silent, the sentence gets written without the number, or not written at all.
It also never invents search volume or ranking data, for a structural reason rather than a virtuous one — no search data provider is connected. There is no source for those numbers, so they do not appear.
Topic proposals work the same way. Each one states why the topic is worth writing and is tied to something the business actually sells; generic topics that would suit any company in the industry are flagged as filler. Competitor analysis fetches approved competitors' public pages to understand what a topic needs to cover.
One more constraint, because crawled text is an input like any other: pages are treated as untrusted and passed to the model in marked blocks. A page that tries to issue instructions to the model is reported, not obeyed.
What the list of claims it could not support does for your review
Every draft arrives with a score for structure and voice, plus a list of the claims it could not support from the business profile.
That changes the shape of the job. Instead of reading defensively from the first word, you start with the list and make a decision about each item: confirm it, rewrite it, or cut it. Then you read the piece as an editor rather than an auditor.
Be clear about the limit. No list catches every error. Models get things wrong, including in ways nothing flags, which is why human approval stays mandatory rather than recommended. The list narrows the search; it does not end it.
This is only a sensible arrangement because nothing goes out on its own. Articles are created as drafts, and publishing requires an explicit action from you. A tool that can publish by itself cannot offer you a review list in good faith — the review would be optional, and optional review is no review.
What the list contains for your business depends on what your profile confirms. A thinly confirmed profile produces more flagged claims. That is the system working, not failing.
A short review routine for any AI draft
This routine transfers to any tool, with or without a list to start from.
- Start with whatever the tool flagged. Clear the known unknowns first.
- Sweep the categories. Numbers, dates, guarantees, superlatives. Ignore the prose between them on this pass.
- Ask for a source you could point to. Editorial practice is unforgiving here: a figure needs a source you can name, not one you assume exists somewhere.
- Delete rather than soften. "Around", "typically", "up to" do not rescue a claim you cannot support. Cut the sentence; the paragraph rarely misses it.
- Fix it at the source. If the figure is real, confirm it in the business profile so the next draft has it and the same review does not recur.
- Read the opening and closing paragraphs last. Confident summary claims collect there, and by then you know what the piece can honestly say.
Who is accountable when the wrong figure goes out
The business whose name is on the page. That is the whole answer, and no vendor arrangement changes it.
So the approval step has to be a real decision made by a person who knows the business, not a button pressed to clear a queue. Design the process so that decision is visible. In Kwerivo, workspaces support members and roles, with an audit log of which actions were taken and by whom — so "who approved this" is a question with an answer.
Ownership sits the same way round. Customers own their inputs and outputs, including website content, business profile and drafted articles; Kwerivo claims no ownership. The work is yours, and so is the sign-off.
What to insist on before you let a writing tool near your site
Six questions. They apply to anything that will write in your name.
- Where does it get facts about your business — and did you confirm them? An impression of your industry is not a source.
- Will it refuse to state a figure it cannot source, or fill the gap? Ask which behaviour is the default.
- Does it hand you a list of what it could not support, or leave you to find it? Unstructured checking is checking that stops happening.
- Can it publish without you? It should not. Kwerivo requires an explicit action to publish.
- What does it do with your content and credentials? Kwerivo never holds site credentials and never makes requests to your server; the WordPress plugin pulls approved articles using a key Kwerivo issued. Prompts are not stored — only the template, version, parameters, an input hash and cost are recorded.
- Is your content used to train models or write for other customers? It should not be. Kwerivo's is not, and deleting a workspace or account is self-service in Settings, scheduled with a 30-day window you can cancel, after which data is permanently destroyed. The privacy policy sets that out in full.
See the list on a draft of your own
The argument here is easier to check than to describe. Request an invitation, correct and confirm your business profile, then look at the list of claims it could not support on a real draft about something you sell.
No payment details are collected; billing is not connected and nothing is charged at present. Questions to support@kwerivo.com.