ChatGPT disclosed things in this session that I did not expect it to disclose, and there is no guarantee it stays available. I would rather people had time to use it than had a tidier version of it later.
So: every word of every ChatGPT response here is verbatim, and that part is checked automatically on every build. What is not finished is the presentation. The color coding on the code blocks is incomplete and some of it is imprecise, and there are notes to myself still sitting in the page.
Parts 1 and 2 are close to empty and are getting a lot of detail over the next week or so, along with a cleanup pass on everything else. Worth checking back.
AI Search Optimization
How AI Search Works, From Prompt to Response
A guide for SEOs, AI search marketers, and marketing teams
If you want to get recommended, you have to understand how a prompt is handled from prompt to response.
Marketers are trying to work out how to get their site into the answer. That is hard to do without knowing how the answer gets built.
This resource follows one question all the way through. A person types a sentence. Some time later, a few company names come back. What happened in between, and which parts of it can you influence?
Most of what is written about this is guesswork. The middle section of this resource is not. It is a full transcript of a session where ChatGPT was asked to show its own tool calls, pushed on whether the answer was real, and then made to run the calls live and print what came back.
What is in here
Parts 1 and 2 deal with the prompt itself. A test that asks for hotel recommendations, where the only thing that changes between runs is one word about the car the traveler is picking up. The word should not matter. It does. Part 1 shows this with search turned off, so the answer comes from the model alone. Part 2 turns search on and shows the same effect getting stronger, because the word ends up inside the search queries.
Part 3 is what to do about it. If a single irrelevant word moves the answer that much, a one-line prompt is not a measurement. This part covers what a prompt has to contain before the result means anything.
Part 4 opens the machine. Eleven turns of a real conversation, verbatim. The actual arguments sent to the search tool. The actual text that came back. What the model can see about your page, and the much longer list of what it cannot.
Part 5 runs the whole pipeline on five real commercial questions: a hotel, a blender, a CRM, a local HVAC company, and a healthcare provider. Pick the one closest to your business.
Part 6 is the prompt that reproduces all of it on your own category.
Three things worth knowing before you start
The model cannot see any ranking score
No PageRank. No domain authority. No relevance score. No backlink count. It sees results in an order and nothing about why they are in that order. That does not mean the search engine behind it has no scores. It means those scores never reach the thing writing your answer.
Your structured data never arrives either
No JSON-LD, no schema.org objects, no Open Graph tags, no author markup. That markup may well shape the title, the date and the snippet upstream, before the result is assembled. It just never arrives as something the model can read.
What comes back is plain text, not data
Requests go out as JSON. Results come back as text, with a cap on how much of your page can be used, a rough age instead of a date, and no field connecting a result to the query that found it.
The color key explains which parts of the transcript came from a real call and which parts ChatGPT made up. That distinction is the difference between evidence and a story, and the colors are the fastest way to tell them apart.