It depends entirely on what you mean by works, and that is not a dodge. On one lane we ran, generated pages produced a sustained five times rise in impressions, from about 28 a day to about 156 a day, and delivered zero additional visitors. Both of those statements are true at the same time. Google indexed the pages, showed them to people, and nobody clicked, because the pages surfaced at an average position in the forties where nothing gets clicked.
If you had measured that lane on impressions you would have called it a success and built a thousand more pages. If you had measured it on clicks in the first month you would have killed it before it had been indexed long enough to judge. Both readings are wrong, and which one you land on is mostly an accident of which number you happened to look at.
The useful question is not whether AI content works. It is whether you have written down, before you spend anything, the number that will tell you to stop. Operators who skip that step do not run an experiment. They run a purchase.
The three pages in this section are the parts we got wrong and had to measure properly: what generated pages do once they are indexed, what image generation actually costs once you include the renders you throw away, and why the keyword volumes you are planning against are probably contaminated.