Ask an assistant for the best merino wool dress socks for someone who stands all day and you get three names back. Not ten. Three, with a sentence of reasoning each. If your product is not one of them, you were not in the consideration set at the moment the customer decided.
Most explanations of this stop at “the model was trained on the internet”, which is both true and useless. What actually happens is more specific, and the specifics are where the work is.
The answer is assembled, not recalled
For a commercial question like this one, the engine does not answer from memory. It retrieves. It runs its own searches, pulls a set of pages, reads them, and assembles a response from what it found. The set of pages it pulls is smaller and more stable than people expect. Run the same question ten times and you will see the same handful of sources doing most of the work.
That stability is the whole opportunity. If the retrieved set were random there would be nothing to optimise. It is not random, so there is.
Three things decide whether you are named
Presence in the retrieved set
If none of the pages the engine pulled mention your product, you cannot be recommended. This is the blunt one, and it is where most absent brands actually fail. Your own product page being excellent does not help if the engine did not retrieve it, and for broad category questions it usually does not.
The quality of the evidence on those pages
Being mentioned is not the same as being recommended. When two products appear on the same buying guide, the one with specific, checkable detail tends to get named: the needle count, the fibre percentage, the fit range, the return window, the number and recency of reviews. Vague positioning language loses to specifics every time, because specifics are what the model can safely repeat.
How the sources frame you
Sentiment and framing carry through. If the pages that mention you describe your product as a budget option, the answer will describe you as a budget option, regardless of what your own site says.
Which sources actually matter
Not all citations are equal, and the difference is measurable. Some pages are retrieved for almost every question in a category. Others appear once. The ones retrieved repeatedly are worth many times more than the ones retrieved occasionally, and they are usually not your own site.
- Buying guides and listicles, the single most-retrieved format in practice
- Independent reviews and comparison pages published by third parties
- Community threads where real users argue about the category
- Review platforms, weighted heavily by recency
- Your own product and collection pages, which matter most for questions that already name you
The practical consequence is uncomfortable for most marketing teams: the highest-leverage work is often not on your own website. It is getting onto the pages other people control.
What to actually do about it
Work one product at a time. Take a product that matters commercially, generate the ten or so buying questions a real customer would ask, and run them. Record every source the engine used. Find the sources where competitors appear and you do not, weight them by how often they were retrieved, and start at the top of that list.
Then, and almost everyone skips this, freeze a couple of comparable products as controls and leave them untouched. Rerun the same prompts at two weeks and four weeks. If the treated product moved and the controls did not, you learned something. If both moved, you learned that the engine changed and you learned nothing about your work.
The short version. Find the pages that get retrieved. Get onto the ones your competitors own and you do not. Make the evidence on them specific enough to quote. Measure against controls, or you are guessing.
What this does not do
None of this guarantees a position. The engines change, competitors act, and retrieval sets shift. What it does is improve the evidence available at the moment the answer is assembled, which raises the probability of inclusion. Anyone promising a guaranteed ranking in an AI answer is selling something they cannot deliver.
The honest version of the claim is narrower and more useful: close the specific gaps that matter for one product, measure against controls, and follow the result through to sessions and orders. Do that repeatedly and you build something better than a tactic. You build a record of what works in your category.
