You type a question into ChatGPT. Or Perplexity. Or Google AI Mode. You get an answer in seconds.
What happens in between those two moments is exactly what determines whether your brand exists in that answer, or doesn't.
We have been working on a tool that runs client content through exactly this process, and we are starting to roll it out this week. More on that at the end.
This new tool allows us to take you as a brand through all the 5 steps that happen inside an LLM, and see where your page might be excluded from the eligible mix. And it also helps us tune your content to stack the odds in your favor next time.
But first, here is what actually runs behind the scenes when you type in a prompt in major AI platforms.
Step 1: Your question gets broken apart
The AI didn't process what you typed. It decomposed it. One question became 5, 10, sometimes 20 sub-queries, each one probing a different angle of what you might have meant.
In traditional SEO, something similar happened in the background (you typed "pizza" and the engine might have expanded it to "pizza delivery new york"). Now the decomposition is far more complex and multi-directional.
Step 2: Each sub-query gets routed to a different tool
Some sub-queries go to a vector search. Some to a keyword index. Some to a live browse of the web — yes, how well you rank still matters, a lot. Some to structured data APIs.
The system picks the right retrieval method for each fragment of your original question. There is no single retrieval path anymore. There are several running in parallel.
Step 3: Chunks are fetched, not pages
Not your homepage. Not your article. Specific passages. Paragraphs. Sentences.
The AI pulls fragments that matched each sub-query, from wherever they live on the web. If your content is not structured in a way that makes individual passages retrievable and self-contained, it may never enter the mix at all.
Step 4: A critic reviews the draft
Before you see anything, an internal grader checks the assembled answer:
- Is it fresh enough?
- Are the sources diverse enough?
- Does anything contradict?
If the answer fails the check, the loop runs again. Content that introduces contradictions, lacks freshness signals, or comes from an over-represented source gets filtered out at this stage.
Step 5: A synthesizer decides what makes the final cut
Passage A versus passage B: which one better serves this query? That comparison happens repeatedly across every candidate chunk. The winners get stitched into the answer you read.
This is the stage where authority, specificity, and clarity determine survival. Vague content loses. Precise, credible, well-structured content wins.
Your brand either shows up in step 3 or it doesn't show up at all.
Not only because of your domain authority. Not only because of your keyword rankings. Whether your content exists as a retrievable, credible, specific chunk that survives steps 4 and 5 is what decides it.
What we built
The tool we have just built maps exactly where you stand across each of these five stages, for your brand, in your category, across the major AI platforms. It helps find the gaps responsible for falling out of AI citations and gives you a roadmap on what to fix.
We are rolling it out with clients on our consulting program this week.
If you have seen your traffic drop, or you would like guidance on optimizing for visibility in both traditional search and generative AI, book a diagnostic call. I will confirm if we can help and tell you the next steps to take.
Best, Gert
Hat tip to Mike King for the research work behind this five-stage model and for making it accessible to practitioners.