Key points
- An assistant does not look at your store, it reads it. Information that cannot be read unambiguously (a price only on an image, terms buried in a PDF, a "standard" return window) effectively does not exist for the model.
- Knowledge about your store comes from four groups of sources: structured product data, machine-readable content, external signals, and the model memory topped up by fetching information live.
- Advertising does not buy a place in the recommendation. Paid formats inside assistants do exist, but according to platform statements and industry sources they are labelled and kept separate from the answer itself.
- An answer from model memory and an answer from a live fetch are two different problems: a stale brand picture, or invisibility in what the model finds right now. The audit in chapter 2 tells them apart.
Before you change anything in your store, it pays to know what an AI assistant builds its answer from, the answer where your brand name either appears or does not. This is not a technical curiosity, it is a question of sequencing: if you do not know where the model gets its knowledge, you will fix things it never reads and skip the ones it bases decisions on. This chapter explains the mechanics, meaning how AI decides what to recommend, and which four groups of sources shape its picture of your store. Implementation details belong to the chapters that follow. Here we are drawing the map you will lay everything else on.
The model does not look at your store, it reads it
A customer lands on a product page and within a fraction of a second reads things nobody wrote down: that the photography looks solid, that the store feels substantial, that the price is fair against what they saw earlier. An AI assistant does none of that. Even when it fetches your page while answering, it takes text and code from it, not an impression. It has no emotions, does not remember your banners, and does not assume that a tidy layout means a trustworthy company.
Two consequences follow, and they hold for the rest of this guide. First: information that cannot be read unambiguously effectively does not exist for the model. A price written only on an image, delivery terms hidden in a PDF, a return window described as "standard" are blank spaces to a machine. Second: contradictory information is often worse than missing information. When free delivery starts at 199 zloty in one place, at 250 zloty on another page, and the terms and conditions say something else again, the model has no way to settle which fact is true. It will either state something false or, more safely, reach for a store that speaks with one voice.
A recommendation needs one more thing: confidence about who a fact belongs to. That this product is exactly the model reviewed on another site. That this price is your store's, not the comparison site quoting it. That the company behind this domain is the same one the trade press writes about. Where that attribution cannot be made confidently, the model usually does not gamble and reaches for an offer it understands better.
The four sources behind what a model knows about you
The knowledge behind an assistant's answer comes from four groups of sources. One thing separates them, and it is the thing that matters most to you: how much control you have. You own the first two outright, you shape the third indirectly, and the fourth you do not control at all, you can only feed it.
| Source group | What the model takes from it | How much control you have | Where in the guide |
|---|---|---|---|
| Structured product data | Name and brand, GTIN identifier, price, currency, availability, specs: facts stated outright in page code and in the feed | Full | Chapter 3 |
| Machine-readable content on site | Answers to buying questions: what the product suits, how it differs from others, who it is for, how delivery and returns work | Full | Chapter 4 |
| External signals | Customer reviews, rankings, "top 10" round-ups, mentions in media and directories: second-hand confirmation | Indirect, you influence but do not decide | Chapter 5 |
| Model memory and live fetching | The general brand picture fixed during training, plus fresh information pulled while answering | None directly, only through the three above | Chapters 2 and 6 |
The order is deliberate: from sources you control entirely to ones you influence only indirectly. The work starts at the top of the table.
Structured product data
These are facts handed to the machine outright instead of left to be inferred: product name and brand, GTIN identifier, price, currency, availability, specifications. They live in two places: in structured data inside your page code (Product and Offer schema) and in the product feed you send to comparison engines, marketplaces and buying systems. According to industry sources, shopping recommendations from assistants lean heavily on exactly this layer, a complete feed and correct schema, rather than on how the product page looks. GTIN acts as the key that ties your offer to the same product described elsewhere. How to build this and what to watch for is the subject of chapter 3.
Content a machine can actually read
The second layer is ordinary text, only written so that an answer can be lifted out of it. The model looks for specifics: what this product suits, how it differs from the cheaper variant, who it will be too big for, how long delivery really takes, how returns work. According to industry sources, buying-guide style content tends to be cited more often than product pages alone, because it answers the customer's question instead of merely describing the goods. That is good news for stores without the widest catalogue: you can be a useful source without being the biggest seller. How to write so the model has something to quote is covered in chapter 4.
External signals
A model rarely stops at what you say about yourself. It looks for confirmation: customer reviews, "top 10" round-ups in your category, rankings, mentions in media and industry directories. The principle is simple: a fact confirmed in several independent places is safer to cite than a claim on the seller's own page. You do not control this area, because you do not decide what somebody else writes about you, but you influence it more than most teams assume. Chapter 5 deals with it in full.
What the model remembers and what it fetches
The last group is the least obvious, because it concerns how the answer is produced at all. Some answers come from what the model "knows" from training. That is a picture from some time ago, with no sources, occasionally a year or more out of date. Other answers come from fetching fresh information while responding, and then you usually see links and citations. This has a very practical consequence for you, because these are two different problems. If the model answers from memory and gets facts wrong, you have a stale brand picture fixed somewhere outside your own site. If it fetches live and still skips you, the problem sits in what it finds. Telling those two cases apart is one of the jobs of the audit in chapter 2.
Advertising does not buy a recommendation. Paid formats inside assistants already exist, but according to platform statements and industry sources they are labelled and separated from the answer itself: media budget does not move you onto the list of recommended stores. That changes the economics of visibility, because there is no bid to raise here, only groundwork on data, content and reputation. One caveat, stated plainly: these are platform statements, and the advertising market inside assistants is still forming, so they are worth re-checking at the source from time to time.
What this means for the order of work
The source map gives you a natural running order and protects you from the most common mistake, which is starting with content production. First you check what models say about you today, because without that you cannot tell whether your problem is absence or distorted facts, and those are two entirely different jobs. Then you tidy the layer you control outright: product data first, content second. External signals come last, because they move slower and depend on other people's willingness. Measurement closes the loop and immediately opens it again.
It is also worth seeing where this leads. An assistant that recommends a store today is starting, in some scenarios, to close the purchase itself, and then only a machine reads your store, with no human at the other end. We go deeper into that thread in our piece on agentic commerce. The next chapter comes back down to earth: you will check what assistants say about your store today and record it in a form you can return to a quarter from now.
Questions
How does AI decide what to recommend in answer to a shopping question?
The model assembles an answer from facts it can read unambiguously and confidently attribute to someone: structured product data (Product and Offer schema, GTIN, the feed), on-site content that answers the customer question, external signals such as reviews and round-ups, plus what it remembers from training and what it fetches while answering. Where data is missing or contradictory, the model usually picks the offer it understands better.
Does the model visit my site when it answers a customer?
Sometimes it does, and sometimes it answers from memory. When it fetches information live, you usually see citations and links. When it answers from training, there are no sources and the facts may be stale. Even when fetching, though, the model takes the text and code of the page, not its appearance, so anything visible only on an image or in a layer rendered later in the browser may not exist for it.
Can I pay to have an AI assistant recommend my store?
According to platform statements and industry sources, no: paid formats inside assistants are labelled and separated from the answer, so ad spend does not buy a place in the recommendation itself. Visibility in answers is earned through the quality and completeness of your data, content and external signals. The advertising market inside assistants is still forming, so these rules are worth verifying at the source periodically.
