What they are genuinely good at
Being fair to the machines first, because the parts they do well are the parts that used to eat my afternoons.
Killing the blank page
A sensible hierarchy and rough copy you can react to. Reacting is far easier than inventing.
The repetitive middle
Alt text for forty images. Meta descriptions. The eleventh rewrite of a paragraph you can no longer read objectively.
Questions you forgot
Ask for the twenty things a customer might worry about and you will get four you had not considered.
Work that drains without distinguishing
The real category. If a task exhausts you and does not differentiate you, hand it over.
Why the output all looks the same
These systems are trained on what already exists, so they are excellent at producing the average of everything they have seen. The average is safe. The average is also why so many generated sites share the same soft gradient behind the same confident headline above the same three-column feature grid that fades in on scroll.
For a business trying to be chosen, being interchangeable is worse than being slightly bad. A slightly bad site gets remembered. Most of these tools are also not designing anything — they are picking which pre-built components to stack, which is why the result feels familiar. Because it is.
Four judgements it cannot make for you
- When to break the pattern. Knowing that this particular brand needs an uncomfortable amount of white space, or that the sticky navigation everyone recommends is ruining this specific portfolio. A model optimises toward the average; it has no gut, it has a probability distribution.
- What actually happened. The true story is the most persuasive thing on most small-business websites, and it is the one thing that cannot be manufactured without lying.
- Who your one real customer is. Not the imaginary everyone. The person comparing you against a cheaper quote and quietly worried they will regret going cheap. You know she does not care about your credentials, she cares whether you will text her back. Writing to that worry is precision no model can infer.
- Whether it is true. These systems produce confident sentences regardless of accuracy. One invented fact on your site damages credibility more than a slow page ever will.
An honest admission
I found a line in a draft of my own site claiming that thousands of engineering teams trusted a studio that had, at that point, no clients whatsoever. It was fluent, it was well-placed, and it was a lie. It got caught because I run a script that specifically hunts for claims like that before anything ships — not because I noticed while reading.
That is the real lesson. The danger is not that the machine writes badly. It is that it writes plausibly, and plausible is much harder to catch than bad. Whatever you use, something has to check it. Mine is a checklist and a script; yours could be a friend who will tell you the truth.
Where that leaves us
The contest was never human against machine. It is people who use these tools well against people who refuse to touch them, and the winners will be the ones letting the machine take the busywork so they can spend their attention on the parts that carry a fingerprint. The tools got faster and made the human judgement more valuable, not less.