How AI assistants choose sources: what is known

The honest answer: nobody knows exactly

How an AI assistant decides which source to name in an answer is not disclosed by the providers. There is no published ranking of factors, and anyone who promises you a “guaranteed AI placement” is claiming more than can be known. The broad mechanisms are known. A sensible approach can be built on them without guessing.

What is established

Two ways in which assistants obtain knowledge are documented:

• Training: Language models are trained on large volumes of text. The knowledge from this is frozen at a cut-off date. Anything you publish afterwards is not in it.
• Retrieval from the web: Many assistants can search and read pages when needed. In doing so they draw on search indexes, some of their own and some belonging to search engines, and summarise the results. Some name the sources with a link, others do not.

The providers also run their own crawlers, which can be addressed in robots.txt. Which ones they are and how you control them is covered in AI crawlers and robots.txt. If you block these crawlers, you cannot expect to appear in answers that rely on these retrievals.

What is probable but not confirmed

Conjectures can be drawn from observation and from what Google says about its quality assessment. Marked as conjecture, not as fact:

• In our experience, pages that rank well in traditional search engines are used as sources more often, because assistants often read the top results. A fixed link between the two is not documented.
• Clearly worded answers to specific questions are easier to cite than long, vague texts.
• Information that agrees across several independent sites appears more reliable. This corresponds to the principle Google describes as “trust”; more on this in E-E-A-T explained.
• Structured data can make content easier to understand. Whether a particular assistant uses it is not publicly known, see Structured data for businesses.

What is not documented

Some widespread claims cannot be verified:

• That a particular file such as llms.txt reliably influences the selection. As things stand, it is not confirmed that some providers read such files at all.
• That a particular text length, choice of words or heading format is “AI-optimised”.
• That you can buy a mention in organic answers by paying.

On top of that, the systems change constantly. A rule that seems to hold today may be obsolete after the next model change. So rely on measures that make sense even without AI.

What you can do without speculating

Make sure your content can be reached

Check whether your pages are indexed and whether your robots.txt blocks any crawlers that you actually want to allow. Basics on this: SEO basics for businesses.

Answer questions directly

On important pages, state the question customers ask and answer it in the first paragraph in two to three sentences. Details follow afterwards. This helps readers just as much as machines. A collection of such answers is our Frequently asked questions page.

Make statements verifiable

Give sources, figures with their origin, the names of those responsible, an address and a way to get in touch. Avoid claims you cannot back up. An assistant cannot check whether something is true, but it can see whether the details agree across several sites.

Consistent key facts everywhere

Company name, services, location and contact details should match across your website, your Google Business Profile and directories. Inconsistent information unsettles people and systems alike.

Why restraint in making promises makes sense

Because the selection is not transparent, any promise of a particular mention is disreputable, including from us. As with traditional search engines, you can improve the conditions, but you cannot force the system’s decision. A sober attitude makes sense: check what demonstrably helps people, and treat visible effects in AI answers as an added benefit, not as a basis for planning. If you keep your content clear, correct and up to date anyway, you lose nothing, even when the systems change.

How to check whether it is working

Put together a list of ten questions your customers might ask, such as “Which provider for [your service] in [your region]?”. Ask them at regular intervals, for example once a month, in different assistants and note whether and how you are named. Bear in mind that the same question can produce different answers on different runs; what matters is the trend over several weeks, not the individual case. For an introduction, see Getting mentioned in ChatGPT and Perplexity, and for the overall picture GEO explained.

If you would like to know where your site stands, we explain our approach on the SEO and GEO page; we are happy to answer questions via contact.