Write only what you can source: AI content for a review site

A review site lives on one thing: readers believe what it says about the companies it lists. AI can now write the pages of such a site in days instead of months. It can also, in the same breath, give a company a feature it doesn’t have, a figure nobody published, or a warm “when I first tried them” that never happened. One of those, found by a reader or by the company itself, costs more than the speed saved.

We rebuilt a directory and review website in 18 days, every page checked one by one at phone and desktop width before launch. Most of the writing was done with AI. These are the rules that kept it honest, and they are the rules you should ask for if someone writes content with AI for you.

Rule one: never invent a fact, a figure or a memory

The first rule is the shortest. The AI writes only what can be worked out or sourced. Not what is likely, not what most companies in the category do, not what sounds right.

Facts and figures are the obvious part. The third word matters as much: memories. A model trained on millions of reviews has learned that reviews sound personal. Left alone, it writes “I signed up on a Tuesday and the support team answered within minutes”. Nobody signed up. Nobody waited. It reads well, and it is a lie told in the owner’s name.

We have written before about a pipeline where the AI writes and the data decides: figures computed first, then handed to the model. A review site needs the same idea applied to facts about other companies, which are not in any database you own.

Every fact about a company comes from its own pages

Each fact about a listed company is taken from that company’s own website, and stored with two things: the page it came from and the date it was checked. Not from another review site, not from a forum, not from what the model remembers about the brand.

This sounds slow. It is what makes the content maintainable. When a company changes what it offers, you know which pages said what, where it came from and how old it is. A fact without a source and a date can’t be checked again, so it can’t be trusted for long.

When it can’t be sourced, the page waits

Some pages needed facts we could not source, or could not keep current. Those pages were held, not guessed. A held page is marked as such, with the reason, and comes back when the fact can be found. A page with a gap is a small loss. A page with a confident wrong sentence is a liability that stays live until someone notices.

The same went for the content plan. Of the searches sorted one by one to decide which pages to write, some were held for exactly this reason: they asked for facts nobody could stand behind.

A site that refuses automated reading gets a person, later

Some company websites refused to be read by our tools. The easy move is to find a way around the block. We didn’t. Those pages were held and checked by hand later, by a person reading the site like any visitor would.

There are two reasons. A site that says no to automated reading has said something, and working around it is not how you want your vendor to treat other people’s rules. And a fact pulled through a workaround is one you can’t easily show your source for.

The owner’s voice, from the owner’s writing

The site is the owner’s, and it has to sound like him. So the voice was not described by us; it was written from his own writing samples into a voice profile. Every page written or edited then went through a voice check, a script that flags what drifts from it.

The voice check catches style. It doesn’t catch truth. That took a human review, and it found the problem from rule one: lines that read as memories nobody had, first-person moments slipped in to make a page sound lived. They were cut. A voice that sounds like the owner and says things he never did is worse than a plain voice.

Facts stored once, used everywhere

Pages are not the only place facts appear. A review site has lists, filters and comparisons, and those are where copied facts go out of date first.

So each listed company has structured fields: what it offers and the other facts readers filter by, each with its source. The filters and list pages fill themselves from those fields. A side-by-side comparison of any two listed companies is built from them too, every line with its source. Change a fact once and every list and comparison that uses it changes with it.

Speed is not the hard part

None of this slows the writing much. The AI still drafts in minutes. What changes is where the time goes: into sourcing, holding and checking, the parts that decide whether a reader can trust the page. As we told writers whose seats AI took, writing was never the job. Making something worth believing was.

We have done this at larger scale on thousands of lessons for a free school, and before the first page of another site we wrote the rulebook first. The pattern holds every time: the rules come before the volume.

What to demand from anyone writing content with AI for you

• Show me the source of this sentence. Pick any fact on any page. They should be able to show where it came from and when it was checked.
• What happens when a fact can’t be found? The right answer is “the page waits”, not “we write around it”.
• Who has used the product they describe? If the page says “I”, someone should have. Otherwise the line goes.
• What do you do when a site blocks your tools? A person checks it later. Never a workaround.
• Whose voice is this, and how do you check it? Built from your own writing, checked on every page.
• Where do the facts live? In one place, with sources, feeding every page that repeats them. Not copied by hand into fifty pages.

A vendor who can answer these has built a process. One who can’t is risking your reputation on the model never getting a fact wrong.