How AI Chooses Its Sources: From Question to Answer

This mechanism is central to brand visibility. A page must do more than simply exist or rank well: it needs to be found for the right search angle, then contain a passage clear, relevant and self-contained enough to be selected. The LLM Monitor webinar illustrates this with a simple question: “Which everyday health tracker would you recommend?”

AI does not always answer from memory alone

A language model first acquires knowledge during training. It then learns how to formulate answers, before its behaviour is refined through different alignment and reinforcement stages. Once a version is deployed, part of its knowledge is tied to the data on which it was trained.

That foundation is not always enough. When a question requires recent, precise or factual information, an assistant with web access can rely on a search index. This is the principle of search grounding: the answer is no longer generated solely from the model’s internal knowledge, but also from content retrieved at the time of the query.

Step 1: understand the intent and identify the entity

Before looking for sources, the AI must understand what the user actually expects. Two elements structure this interpretation:

• Intent is the underlying need: getting a recommendation, comparing products, checking a price, assessing the reliability of a feature or choosing according to a specific profile.
• Entity is the object targeted by the question: a brand, product, place, organisation or, in this example, the health tracker category.

The question “Which everyday health tracker would you recommend?” sounds simple, but it is incomplete. The best choice depends on age, activity level, health goals, budget, expected battery life and the mobile ecosystem being used. The AI must surface these dimensions before it can produce a credible recommendation.

This gap between a short query and the needs hidden behind it explains why traditional SEO and AI search do not follow the same logic. A search engine ranks pages for a query. An AI system tries to build an answer that covers several facets of the user’s intent.

Step 2: break the question down with query fan-out

Once the question has been understood, the AI can break it down into several subqueries. This stage is called query fan-out. Instead of running one very broad search, the system explores each of the angles required to answer properly.

In the webinar example, the initial question can generate subqueries such as:

Query fan-out completely changes the playing field for brands. You do not only need to be visible for the broad question. You also need to appear in the sources used for the subquestions that make up the final answer. This is why query selection is critical when you want to measure your visibility in AI systems.

Step 3: search for pages for each subquery

Each subquery can then be sent to a search engine. For the sleep angle, the assistant might look for pages answering the question: “Which fitness trackers offer the most reliable sleep tracking?”

The webinar uses an example in which this search returns twelve sources. That number is not a universal rule: it simply helps visualise the process. The important point is that the AI builds a pool of candidate pages, then analyses them to identify information capable of supporting the answer.

At this stage, being indexed and appearing in search results is the first filter. But it does not guarantee that the content will be retained. A page may be found and then discarded because it does not answer the intent well, lacks precision, contains too much noise or fails to connect a characteristic clearly to an entity.

This selection process explains why certain types of content are used more frequently by AI systems: structured comparisons, detailed reviews, factual data, well-supported opinions and passages that directly answer a precise question.

Step 4: split pages into usable chunks

The AI does not necessarily send an entire page to the model responsible for writing the answer. Retrieved content is divided into smaller fragments called chunks. This process, known as chunking, isolates passages that contain useful information.

In the example shown, one page may contain a passage about the sleep-tracking accuracy of the Fitbit Charge 6 and another about the overnight comfort of the Xiaomi Smart Band 9. A different page may explain the sleep stages being measured without mentioning a specific model.

Each fragment can therefore contain:

• a clearly identifiable entity, such as a brand or product;
• information connected to the intent, such as sleep accuracy, comfort, battery life or price;
• enough context for the passage to remain understandable once separated from the rest of the page.

This final point is decisive. Content can be excellent for a human reader yet difficult to reuse if its paragraphs constantly depend on information located earlier on the page. By contrast, a self-contained paragraph that names the product, the criterion being assessed and the conclusion can be extracted and understood immediately.

Step 5: score, filter and deduplicate the fragments

After chunking, the system may have hundreds of fragments. In the webinar’s educational example, 315 chunks are identified before a much stricter selection takes place.

The fragments then go through several processing steps:

• Cleaning — removing menus, ads, tags and other irrelevant elements.
• Semantic clarity — checking that the passage stands on its own and remains understandable outside its original page.
• Relevance ranking — assessing how closely the chunk matches the subquery.
• Organisation by source and entity — grouping information that refers to the same product, brand or criterion.
• Deduplication — removing passages that repeat the same information.
• Re-ranking — carrying out a final ranking to keep the excerpts most useful for the answer.

The webinar illustrates this reduction by moving from 315 candidate fragments to 12 selected chunks. Again, these figures are examples. The essential point is the competitive logic: several pages may cover the same subject, but only the most usable passages enter the final context.

Selection therefore does not depend only on domain reputation. A source must first be found, then judged relevant, and finally provide a stronger excerpt than competing sources for the intent being explored. This mechanism helps explain why some brands are cited by AI while others remain invisible.

Step 6: build the context sent to the model

Once the best chunks have been selected across all subqueries, the system assembles them into a context. The webinar presents this assembly as a “super prompt” made up of three components:

The model does not simply copy one source. It synthesises several fragments, connects the information, follows the instructions and formulates an answer tailored to the initial question. This is also why two assistants, or two slightly different phrasings, can produce different recommendations on the same topic.

The complete workflow, from question to answer

The process presented in the webinar can be summarised in eight stages:

What this workflow changes for your brand visibility

The consequence is clear: publishing a page entitled “best health tracker” is not enough. To enter an answer, your brand needs to be present in content that responds to the subqueries the AI actually explores.

Three conditions must be met:

• Be found — your content, or a source mentioning you, must appear for the relevant search angles.
• Be understood — the passage must clearly connect your brand, product and the benefit being sought.
• Be selected — your fragment must be clear, complete and relevant enough to score better than the alternatives.

This logic means working both on your own website and on third-party sources that talk about you. Specialist media, comparisons, reviews, tests and reference pages can each cover a different part of the fan-out. That is the core of a strategy designed to influence the sources used by AI without trying to manipulate the model directly.

Common mistakes to avoid

• Working only on the main question. The answer is built from several subqueries. Ignoring fan-out leaves entire parts of the topic open to competitors.
• Confusing ranking with selection. A strong search position can help a page get discovered, but it does not guarantee that its passage will be retained in the final context.
• Writing paragraphs that depend on the rest of the page. Vague wording such as “this product”, “this solution” or “as explained above” becomes difficult to use once the content is split into chunks.
• Adding keywords without answering the question. The selected chunk must resolve an intent with precise information, not simply repeat the query vocabulary.
• Speaking only from your own website. Third-party sources add credibility, comparison and diversity to the information available about your brand. To appear in ChatGPT responses, you need to build a documented presence across this wider ecosystem.

Guillaume