# How answer engines build an answer

> The stages an AI engine goes through to answer a question, what a brand can affect at each one, and how each stage shows up in AI visibility measurement.

## Reading brief

- Decision: Which stage of answer-building explains what a brand sees in its AI visibility numbers.
- Evidence: Each stage, what a brand can affect there, and the pattern it leaves in the four signals.
- Action: Read a gap in the numbers back to its stage, then work on the lever that stage responds to.

An answer engine turns a buyer's prompt into a written reply in a few stages, and a brand can be lost at any of them. Knowing the stages matters less for their mechanics than for one practical reason: each stage leaves a different pattern in the measurement, and each responds to a different kind of work.

The mechanics are well covered elsewhere. This chapter keeps them to a line each and spends its length on what each stage means for a brand.

## What happens at each stage?

| Stage | What a brand can affect | Where it shows |
|---|---|---|
| Training | How widely and consistently the brand is described across the web | Mentions with no source cited |
| Search | Pages that answer the narrower questions | Citations on specific questions, gaps on sub-topics |
| Retrieval | Whether its pages can be reached and indexed | Cited on one engine, never on another |
| Reranking | Passages that answer directly and stand alone | Citation share against competitors |
| Writing | How clearly the brand is described as a fit | Average position and share of voice |

The rest of the chapter takes the stages in order. Each section ends with the chapter that covers its lever.

## Which route did an answer take?

A large language model can answer in two ways. It can answer from its training data, the text it learned from before release, which stops at its knowledge cutoff. Or it can search first and write from what it finds, a method called retrieval-augmented generation.

The two routes leave different traces. An AI mention is the brand appearing in the text of an answer. An AI citation is a source the answer links or lists, and a cited source is the page or domain behind it.

An answer that mentions a brand and cites nothing most likely came from the model's memory. One that cites sources came at least partly from a search.

The training route changes slowly. A brand moves in it only when a later model learns from a web that describes the brand more widely and more consistently, which is the subject of chapter 6 on earned authority.

## How do search and fan-out decide what gets found?

When an engine searches, it rarely runs the buyer's question as typed. Google [describes AI Mode's query fan-out](https://blog.google/products/search/google-search-ai-mode-update/) as breaking a question into subtopics and issuing many searches at once, and OpenAI's [ChatGPT search help page](https://help.openai.com/en/articles/9237897-chatgpt-search) says ChatGPT rewrites a question into one or more targeted queries. Query fan-out is the general name for this.

The searches match meaning, not only words. Semantic search turns text into an embedding, a list of numbers that places similar meanings close together, and vector search finds the passages nearest to the query. Retrieval can also show recency bias, a tendency to favor pages that are new or recently updated.

For a brand, the consequence is that the buyer's question is not the unit of competition. The narrower questions behind it are, and a brand with nothing that answers them has nothing to retrieve. Chapter 5 covers content built for those questions.

## How are passages chosen?

Engines retrieve passages, not whole pages. Chunking splits a page into passages that can be retrieved on their own, and reranking reorders the candidates so that only the most relevant few reach the model.

The limit is fixed. A model reads text in tokens, and its context window caps how many it can consider at once, prompt, sources, and reply together. Many retrieved passages never make the cut.

A passage survives reranking when it answers the narrower question directly, without depending on the paragraphs around it. That is a writing discipline more than a technical one, and chapter 5 covers it. Retrieval failures that come from access rather than writing belong to chapter 4.

## How is the answer written?

The model writes from the passages it kept. Grounding ties its claims to those sources, which is what makes citations possible and checkable.

The order of brands in an answer is not a ranking of merit. Research on language models has found that they use information at the start and end of their input more reliably than information in the middle, a pattern known as position bias and documented in the [Lost in the Middle study](https://arxiv.org/abs/2307.03172). Answers also change from run to run, a property called non-determinism.

Both are reasons to read average position across many answers and never from one. Chapter 3 covers how.

## What happens when an engine gets the brand wrong?

A hallucination is a fluent statement that is false or unsupported. A brand hallucination is one about a specific brand: a wrong price, a feature the product lacks, a competitor's founder.

The route tells the team where to look. If the wrong answer cites a source, the error is usually in that source, and correcting or outweighing that page is the fix. If it cites nothing, the error sits in the model's memory and changes only as later models learn from a web that states the fact correctly and consistently.

Either way, the correction is made on the web the engines read, not inside the engine. Chapters 5 and 6 cover the owned and earned sides of that work.

## What does this chapter not cover?

The stages describe how any answer is built. They do not describe which answers matter commercially, or where buyers meet them. The next chapter covers the surfaces where buyers read AI answers and how a shortlist forms inside them.