How AI Picks Its Sources

What a searching assistant does

In Lesson 1 we saw that some assistants search before they answer. This lesson looks inside that search. How does the assistant decide which pages to rely on?

We can't see every detail from the outside. Providers don't share the pages they considered but didn't show. Still, the process usually runs in three steps:

1. Search. The assistant turns the question into one or more search queries and finds candidate pages.
2. Select. It reads a handful of the candidates that look most relevant.
3. Write. It combines what it read with what it already knows and writes one answer. Some pages are shown as sources.

Pages drop out at every step. Ranking in search results isn't enough to be selected, and being selected isn't enough to be mentioned in the answer.

Which pages get picked?

Selection rules vary by engine and aren't published. In most cases, though, these qualities work in a page's favour:

• It answers the question directly. The answer sits near the top, in plain sentences.
• It's accessible. AI crawlers can fetch it. It isn't behind a login and isn't blocked by robots rules.
• It's quotable. Concrete facts such as dates, prices, specs and locations are stated clearly.
• It's consistent. The brand's name, category and key facts match across different places.
• It's current. For time-sensitive questions, recent pages usually have an edge.

None of this guarantees selection. These are conditions that can make it more likely.

Types of sources

The pages behind AI answers come in many types. For an English-language question, common ones include:

• The brand's own website.
• Wikipedia and other reference pages.
• Software and product review sites such as G2.
• Community discussions on Reddit and similar forums.
• News sites and industry publications.
• Comparison articles and "best of" lists.

The variety tells you something: AI doesn't learn about your brand only from your website. What others write about you shapes the answer too. Module 3 covers source types in depth.

Mentioned, cited, recommended

These three are often confused, but they are different things:

• Mentioned: the brand's name appears in the answer.
• Cited: the answer links to one of the brand's pages as a source.
• Recommended: the answer explicitly suggests the brand.

One doesn't imply another. A brand can be mentioned without being cited, cited without being mentioned, or mentioned without being recommended.

Even a cited link only tells you one thing: that page was part of the evidence the answer showed. It doesn't automatically mean the brand was recommended. More detail: Citation evidence.

Example: the sources behind one answer

The example below is representative. "Northwind Boards" and "Contoso Tasks" are fictional products.

Question: Which project management tool has the best client permissions for agencies?

Answer: Contoso Tasks lets agencies give clients view-only access per project, and reviewers often praise how easy it is to set up. Northwind Boards also supports client access, though some users in a Reddit thread describe the permission settings as confusing.

Sources: contosotasks.example/features/permissions, a G2 review page, a Reddit thread

In this answer:

• Contoso Tasks is mentioned, its own site is cited and it is described favourably, but with no explicit recommendation it doesn't count as recommended.
• Northwind Boards is mentioned, but the source is a Reddit thread, not its own site. It isn't recommended.
• If Northwind Boards had a clear, crawlable page explaining its permission settings, the assistant might have been more likely to find and use it. That isn't a guarantee.

The part you can't see

You can't know which pages an assistant considered but didn't show, and you can't say for certain from the outside why a page wasn't picked. If an answer changes after you edit a page, that alone doesn't prove your edit caused the change. See What we do not claim to measure.

So the honest approach is: measure the visible evidence, date it, and compare again under the same conditions.

SOV Tracker is a measurement tool that can record brand mentions and cited sources in AI answers over time.

Summary

• A searching assistant usually searches first, then reads a few pages, then writes its own answer.
• Pages that answer directly and are accessible, quotable, consistent and current are usually more likely to be picked. Nothing is guaranteed.
• Sources go beyond the brand's site: reference pages, review sites, forums and news shape the answer too.
• Being mentioned, being cited and being recommended are three different things.
• You can't see the pages an assistant didn't show or prove why an answer changed. Measure the visible evidence over time.