This morning I watched an AI research the next article for my daughter's wedding band. It ran a proper grounded scan of what already ranks, came back with a tidy gap analysis, and recommended an angle. The research was good. The recommended angle was wrong — in a way that would have quietly damaged the brand. The interesting part of the day was the bit where it knew that.
That gap — between research that's correct and an angle that's right for you — is the whole game. It's also the part nearly every AI-content workflow on the market skips, because the way most of them are built makes skipping it inevitable.
The normal way to make AI content is broken by design.
Here's the model most businesses are actually buying, whether they know it or not. Someone hands the cheapest available person — an offshore contractor, a junior, a content mill — a keyword and a login. The instruction is, more or less, "go write an AI article about this." Sometimes there's a research step: run a tool, pull the top-ranking pages, summarise what they cover, fill the gaps. It looks diligent. It produces something competent and grammatical and on-topic.
And it is almost always generic — not because the writer is lazy and not because the AI is weak, but because the writer is starting from zero context every single time. They've never read your site. They don't know what you've already published, what you deliberately don't do, or the one sentence your whole positioning hangs on. So they write to the only thing they have: the average of what everyone else already ranks for. The output regresses to the category mean, which is the precise opposite of what content is supposed to do for you.
The research phase isn't the problem. The missing context is.
I want to be fair to the tools, because the research really was good. For the wedding-band brief, the scan read the live results for "live band vs DJ for a Belfast wedding" and handed back a genuinely useful map: which questions are already answered to death, which are under-served, where the ranked pages go thin. It even spotted that not one of them covers what happens if a performer falls ill on the day. That's real signal, and it would take a person an afternoon to assemble by hand.
Then it did the thing these tools always do: it recommended an angle. Write a comprehensive local guide — drill into Northern Ireland cost breakdowns, package inclusions by band size, three-piece versus five-piece versus a ceilidh option, venue logistics. Reasonable. SEO-sound. And for this client, brand-corrosive.
Followed literally, the recommended angle would have written the band up as exactly the interchangeable function band their entire identity is built against.
Because the band's whole proposition — argued at length on their own site — is that they are not a 200-songs-on-a-list, four-hours-for-a-price function band. A page that slots them into a cost table next to every covers act in the province doesn't just miss the point. It actively undoes the positioning the rest of the site is working to build. It would rank, and every ranking visit would arrive with the wrong idea of who they are.
Here's the part that only works if you've read the whole site.
The reason the angle got caught and corrected is that the tool had actually read the site first — the weddings page, the long essay on why they started playing weddings at all, the existing piece on what a grunge wedding really sounds like. So instead of writing to the scan's default, it reframed the brief: the real article isn't band vs DJ on a spreadsheet of cost and space and song count. It's playlist vs event — and "live band" isn't even one thing, because a safe function band and a DJ have more in common with each other than either has with an act that plays your evening like a headline show.
That reframe is the difference between a page that ranks and dilutes you and a page that ranks and sharpens you. It is not a cleverer prompt. It is not a better model. It is the direct, unglamorous output of having read everything the business has already said about itself before writing the next sentence.
Context is the cheap part to skip and the expensive part to fake.
There's a second layer, and it's the one I think is genuinely hard for the commodity model to copy: memory. The system didn't just read the site once and forget. It carried forward what was already published, what's still in draft, which questions the FAQ already answers — so it didn't propose a topic the band had half-covered three times, and it knew which gaps were already owned and which were actually open.
A contractor handed a keyword starts from nothing on Monday and starts from nothing again on Thursday. Every brief is the first brief. There is no accumulation, no thread, no sense of we already made this argument over here, so this piece should pick it up rather than repeat it. You can buy a great writer and still get disconnected articles, because the asset that ties them together — the running, specific memory of the business — was never theirs to hold.
This is the same idea I keep coming back to on the visibility side of the work. AI summaries fabricate when your articulated content is thin; the fix is to give the model more of you to stand on. Writing is the mirror image of the same principle. The output is only as specific as the context you feed in. Thin context in, generic content out — every time, regardless of how good the underlying model is.
What this means if you're paying someone to write for you.
Stop evaluating AI content by the model and start evaluating it by the context behind it. The model is a commodity — your competitor has the same one open in the next tab. Ask the questions that actually predict whether the writing will be any good: Has whoever is doing this read my whole site? Do they know what I deliberately don't do? Will the piece they write this month know what the piece they wrote last month already said?
If the answer is "they were given a keyword and told to go," you're not buying content. You're buying the category average with your name printed on top — and on the surfaces that matter now, average reads as invisible.
The article we ended up briefing this morning is sharp, true to the band, and says something the first-page results don't. It came out that way for one reason: the thing doing the writing had read the whole site and remembered what mattered. That's not a prompt you can paste. It's an asset you build — and it happens to be exactly the asset we build for the people we work with.