Generative AI

Generative AI is software that produces new text, images, code or audio in response to a prompt, using a model trained to predict what plausibly comes next. The output is composed, not retrieved: nothing in it needs to have existed before as a whole.

Composed, not looked up

A search engine finds documents that already exist and orders them. A generative model writes. Given a question, it produces a sequence of words that fits the patterns in its training data, and it does this whether or not the true answer is among those patterns. That is why the same system can draft a letter well and state an opening time wrongly: both are fluent continuations, and only one of them was checked against anything.

The assistants a customer uses combine the two. They retrieve first, fetching pages that match the question, then generate from what they fetched, often with links back to the sources. The generated paragraph is still composed, but the material it composes from is a set of pages read a moment earlier.

Where a business enters the output

Through two doors. Training data, which is fixed when the model is built and which a business has little say over. And retrieval, which happens at answer time and reads whatever the site serves then. The second door is the one a business can act on this month, and it is what Aggarwal et al. studied in “GEO: Generative Engine Optimization” (KDD 2024): changing the source page changed how visible that source was in the generated answers. This site’s own audit work looks at the same door from the site’s side, which pages a crawler receives and what is on them.

A common question

Of the two doors into a model's output, which one can a business act on now?

Training data is fixed when the model is built, and a business has little say over what went into it. Retrieval happens while the answer is being written and reads whatever the site serves at that moment, which makes it the door that responds to work done this month. That second door is the one the GEO paper studied, finding that changing the source page changed how visible that source was in the generated answer. Audit work looks at the same door from the site side, asking which pages a crawler receives and what is on them.

Sources

Mechanism of source-side change: Aggarwal et al., “GEO: Generative Engine Optimization” (KDD 2024). No figures on this page are this site’s own.