Week 1 · the mechanics

How AI Answers Are Actually Built

You can't change what AI says about you without knowing how it decides what to say. This is the machinery, in plain terms.

Makefield · Berlin

Two memories

What a model knows vs what it looks up

When you ask an AI assistant a question, its answer is drawn from two different places, and the difference matters more than almost anything else in this field.

The first is trained memory — what the model absorbed from its training data, months or years ago. This is why a model can describe a well-known company with no internet access. It's also why that description can be confidently out of date.

The second is live retrieval — what the model looks up at the moment you ask, when web search is on. The assistant runs a search, pulls in a handful of current pages, and writes its answer partly from those. This is the part you can influence quickly, because it reflects what's findable on the web today.

Most modern assistants blend the two. The takeaway: there are two ways to be described — from the model's memory, and from what it retrieves — and they call for different work.

The answer-building set

A shortlist, assembled before a word is written

When retrieval happens, the model doesn't read the whole internet. It pulls a small set of sources — often just a few pages — and builds the answer from those. Think of it as a shortlist the model assembles before it writes a word.

The entire game of AI visibility is about getting into that set, accurately. If your category's best pages don't include a clear, current account of you, you won't be in the shortlist — and you can't be named in an answer built from sources that don't mention you. This is the quiet mechanism behind being left out: not malice, just absence from the set.

Watch one shortlist form. A buyer evaluating CRMs types "best CRM for a small sales team" into an answer engine. It doesn't return ten blue links for her to weigh. It returns three named tools, one sentence on each, and a reason — a shortlist she didn't build and can't see the edges of. Two of those names were decided before she finished reading: the brands the engine could describe accurately and retrieve confidently. The fourth-best tool for her — maybe the best — isn't in the answer, so for this buyer it does not exist.

Now ask the same question again an hour later. The wording moves; the names mostly hold. That gap — stable shortlist, shifting prose — is the whole job of this course in one observation: measure the layer that holds (who gets named, and how they're described), not the layer that wanders (the exact sentence).

(One engine, one query, one moment — illustrative, not a measurement. Week 10 shows why a single look is a snapshot, not a verdict; Week 12 covers what to do when the answer moves.)

Why some brands make the cut

Recognition, repetition, structure

Whether a model knows and surfaces an entity comes down to a few repeating factors:

• Recognition as an entity. The model has to understand that "you" are a specific, distinct thing — a company, a product — not a vague phrase.
• Repetition across sources. Something described the same way in many places becomes the model's default account. If the web's consensus about you is thin or wrong, that's what gets repeated.
• Structure it can parse. Information a machine can read cleanly — plain text, clear headings, valid structured data — is easier to pull than information buried in an image.

None of this is a trick. It's what a careful human researcher would do: find the clearest, most corroborated account and repeat it.

Where citations come from

Citations — often, not you

When an assistant shows its sources, look closely: for many companies, the cited pages aren't the company's own site. They're third parties — review sites, directories, comparison articles, forums. The model is describing you through someone else's account of you.

That has a sharp consequence: the narrative about you can be owned by sources you don't control. Winning it back is one of the central jobs of this work, and a later lesson is devoted to it.

What this means for you

The whole course, in one idea

Everything ahead follows from the machinery above. Because there are two memories, you work on both what the web says now and the consensus that hardens over time. Because answers are built from a small source set, you work to be in it, accurately. Because models repeat the clearest, most parseable account, you make sure that account is yours.

But it all rests on this: an AI answer is assembled, not retrieved whole — and you can change what goes into it.

Common questions

Questions this lesson answers

Why does an AI assistant describe my company inaccurately when my website is current?

Because the answer may come from trained memory — what the model absorbed from its training data months or years ago — rather than live retrieval. A correct website today doesn't fix a stale account already settled into the model. The two problems call for different work: retrieval problems respond to changes on the live web now; memory problems only shift as the web's consensus about you hardens over time.

What is an answer engine?

Any tool that replies to a question with a synthesized answer instead of a list of links — ChatGPT, Claude, Perplexity, Google's AI features, and others. It may draw on trained memory, live web retrieval, or both; this lesson explains why that difference matters more than almost anything else in this field.

Why do AI assistants cite other websites instead of mine?

Assistants often build answers from third-party sources — review sites, directories, comparison articles, forums — because those read as corroborated, independent accounts. If the clearest account of you lives on someone else's page, that is the account the model repeats. Winning that narrative back is one of the central jobs of this work, and a later lesson is devoted to it.

Can you guarantee my product appears in AI answers?

No — and anyone who promises a specific placement is overselling. Answer engines are non-deterministic; the wording changes every time. What you can do is change what the engine is able to retrieve and how accurately it can describe you, then measure whether the stable layer improves. The honest claim is "this changes what the model can retrieve," never "this guarantees a placement."