Ask an assistant which running shoe suits flat feet and you will not get a list of ten links. You will get four sentences naming two or three brands. That output is a different commercial object from a search results page, and the strategies that won the first do not automatically win the second.
The good news is that the selection is not mystical. Retrieval-augmented systems follow a describable process, and each stage of that process has inputs you can change.
Stage one: does your brand resolve as an entity?
Before a model can recommend you, it has to be confident about what you are. Entity resolution is the step where a system reconciles every mention of a name into one coherent thing with attributes: category, audience, price band, location, notable properties.
This is where most brands quietly fail. Their homepage says one thing, their About page says another, their Amazon listing says a third, and their schema markup says nothing at all. Faced with contradictory inputs, a model does not pick a winner. It hedges, and a hedge reads as a competitor getting named instead of you.
Stage two: can your content be retrieved in pieces?
Retrieval systems do not read your page the way a person does. They chunk it, embed the chunks, and pull back the handful that best match a query. A chunk that only makes sense in the context of the three paragraphs above it is a chunk that will never be quoted.
Content that retrieves well shares a shape: a self-contained answer in the first forty words, stated without pronouns or dependence on the preceding section, followed by evidence, followed by nuance. It is the inverted pyramid that newspaper editors have insisted on for a century, which should tell you something about how durable the format is.
What this looks like in practice
- Every heading is a question a buyer would actually type or say aloud.
- The paragraph directly beneath it answers that question outright, in one or two sentences, before any setup.
- Facts are numbers, not adjectives. 'Ships in two business days' retrieves; 'lightning-fast shipping' does not.
- Comparison and definitional pages exist, because a large share of buyer prompts are comparative or definitional.
- Critical data lives in HTML, as tables, lists and structured markup, not inside a PDF or an image.
Stage three: does anyone else vouch for you?
This is the stage that separates brands that move from brands that plateau. Retrieval systems weight independent corroboration far above self-description, for the obvious reason that every brand describes itself favourably. When an assistant assembles an answer, it is disproportionately drawing on sources that are not you.
Which reframes digital PR from a vanity exercise into the highest-leverage activity in the discipline. Not any coverage, though. It has to be coverage on the specific sources these systems demonstrably retrieve from in your category. You can find out which ones those are by reading the citation lists in the answers themselves, at scale, across a few hundred prompts.
The fastest way to change what a model says about your brand is to change what other people's pages say about your brand.
What does not work
A short list, because the category has attracted the usual opportunists. Keyword stuffing for models does not work; embeddings are semantic and repetition adds nothing. Prompt injection in page copy does not work and will get you filtered. Paid 'guaranteed AI placements' do not exist, because nobody can sell you an entry in a system that has no such inventory. And publishing three hundred thin AI-written pages actively harms you, because undifferentiated content gives a retrieval system no reason to select it over anything else.
How to know whether it is working
Build a prompt set of 150 to 400 real buyer questions. Run it weekly across the assistants your customers use. Record mention rate, citation rate, position within the answer, sentiment and competitive share of voice. Because generated answers vary between runs, sample each prompt several times and report rates rather than screenshots.
That scorecard will be uncomfortable at first. It is also the only version of this work that can be defended in a budget meeting, which is why we insist on it before anything gets built.