Case study · 03
Almost every post on this blog was written by software. None of it published itself.
GEO Engine is an AI-powered web app we built to write the posts on this site. It works to a structure that search engines rank and answer engines quote, researches real sources, and draws its own images. Then it stops. A person reads the post, edits it, approves it, and the tool opens a pull request. Merging that pull request is the publish, and it is the one thing the tool cannot do.
Search changed shape, and writing for it by hand got slower.
Ranking a page is no longer the whole job. Answer engines summarise instead of linking, so a post now has to be built to be pulled apart and quoted, not just crawled. That is a different way of writing, and it is more work, not less.
Doing it properly takes hours. The structure, the sources, the links to your own pages, the markup, an image. Every time. It is the kind of job that gets done well twice and then quietly stops getting done.
There is a third problem underneath both of those and we have not solved it yet. Nobody knows what to write next. Without data about your own site and what people actually search for, topic choice is guesswork. That is what we are building now.
It writes to a spec, then it stops and waits.
It learns the site
Reads your sitemap, or you type the pages in. If a site refuses to be crawled it records the refusal and the reason. It does not report a success it did not have.
It writes to the spec
Body, structure, real sources, and links to the right pages of your own site. Written to be quotable, because that is what an answer engine picks up.
It draws the image
The model describes the picture as data rather than as markup, so nothing untrusted ever reaches your site.
You read it
In the app. Edit it, approve it, or throw it away.
It opens a pull request
Against your site. It does not merge it.
You merge
Then it notices, and marks the post published. It never says live unless the post really is live.
It crawls and records
What every page is, and every link between them, so you can ask which of your pages point at this one. That is how you find the ones nothing points at.
Three jobs, and one it has no way to do.
AI
It researches and writes
A whole post against the spec, with real sources it names.
AI
It works out the internal links
Reads the site and decides which of your pages a post should point at. This is the part nobody finishes doing by hand.
AI
It draws the hero image
As a description in data, never as markup.
You
You merge the pull request
That is the publish. There is no button in the tool that does it, and that is the design, not a gap.
The part that matters
Most of this was built to stop it, not to make it clever.
Anyone can make software that writes. The decisions worth showing you are the ones where it could have done something and we built it not to.
It will not get round bot protection
Volta Compliance sits behind Cloudflare. Its robots file answers normally while its pages refuse, so a crawl there is recorded as refused. It does not try again wearing a different name. There is no setting that turns that on.
It will not merge
It opens the pull request and stops. The publish is a person, every time.
It will not spend without you
Every run has a ceiling it cannot go past, and every call records what it cost. Nothing runs on its own and nothing runs twice by accident.
It will not call a page verified because a crawl worked
That proves the address answered. It does not prove it is the page you meant, and we are not going to pretend otherwise.
The posts on this blog, and you are already on it.
Almost every post on this blog was written by it. You are on the site, so you can go and read them and decide for yourself, which is not something most software gets to offer.
A person still reads every post before it goes anywhere. That was never the expensive part and it is not the part we would take out.
The crawl costs nothing to run and it stays that way, so a site can be looked at as often as you like without anyone thinking about the bill.
Next step
Twenty minutes, no obligation.
If you want something like this on your own site, start by telling us what you would actually want to publish. Nothing gets built until you have approved the scope.
Would rather write than talk? Send us a message instead and it reaches the same two people.