Insights · How AI search works

Clarentino Aduk, Founder & Chief Executive Officer | October 2026 | 9 min read

When someone asks ChatGPT, Gemini or Perplexity to recommend a rainforest lodge, a university or a place to invest, the assistant does not show a list of websites in order. It writes an answer, usually naming one or two organisations, and moves on. Being on the second page of search results was a problem. Being absent from an AI answer is invisibility.

For leaders, this raises an obvious question: how does the assistant decide? No one outside the companies that build these systems knows exactly how each one chooses, and the details change frequently. But the broad mechanics are understood well enough to act on, and they are less mysterious than they first appear.

This article explains, in plain terms, how AI assistants find information, how they judge it, why familiar organisations are often left out and what organisations can practically do about it.

It helps to start with what has changed. A traditional search engine retrieves pages it considers relevant and ranks them. The person reading the results does the rest: opening several pages, comparing them and forming a judgement.

An AI assistant takes on much of that work itself. It reads, compares and summarises on the person’s behalf, then presents a conclusion. Some assistants show the sources they relied on, some do not, and many people never click through to check.

That shift has a consequence that is easy to miss. In traditional search, the judgement was made by the person. In AI search, a significant part of the judgement is made by the assistant before the person sees anything. An organisation that wants to be chosen now needs to be chosen twice: first by the assistant, then by the person.

AI assistants draw on two broad kinds of information.

What the model already knows

The first is what the underlying language model learned during training: a broad picture of the world built from a very large amount of published text. This knowledge is wide but dated. It reflects the world as it was when the model was trained, and it tends to favour information that was published widely and repeated often.

This kind of knowledge is slow to change and hard to influence directly. An organisation that was well documented several years ago may still be described accurately. One that was poorly documented, or has changed significantly since, may be described wrongly or not at all.

What the assistant retrieves now

The second source is what the assistant retrieves at the moment of the question. Most major assistants can now search the web, read the pages they find and use them to construct an answer. Google’s AI Overviews and AI Mode draw heavily on Google’s own search index. ChatGPT, Perplexity, Copilot and others use their own search tools and partners.

This is where the opportunity lies for most organisations. Retrieved information responds to what is published today. An accurate, clear and complete page published this month can influence answers this month.

In practice, assistants often combine both sources. The model’s background knowledge shapes how it interprets the question and which sources it expects to be relevant. The retrieved pages supply current, specific detail.

How an answer is assembled

Although each system differs, the process for a question that requires current information usually looks something like this.

1. Interpreting the question. The assistant works out what the person is really asking. A question such as where to see orangutans in Borneo becomes a set of narrower questions about locations, access, timing and operators.
2. Searching. The assistant runs one or more searches, often several in parallel, sometimes in different phrasings or languages.
3. Reading and selecting. It reads a selection of the results and decides which sources appear relevant, reliable and specific enough to use.
4. Composing. It writes an answer drawing on the selected sources and its background knowledge, often naming specific places or organisations.
5. Citing. Depending on the assistant, it may show links to some of the sources it relied on.

At every stage, an organisation can be included or excluded. It may not appear in the search results the assistant consults. Its pages may be found but judged too thin to be useful. Or its information may be used but attributed to an intermediary who published the same facts more clearly.

When an assistant retrieves information, it has to decide which sources to rely on. In practice, the sources that get cited tend to share several qualities.

They answer the question directly

A page that clearly explains how to reach a destination, what it costs, how long to allow and when to go is more useful to an assistant than one that only says the destination is beautiful. Assistants are trying to answer specific questions, and they favour sources that contain specific answers.

Much organisational content is written to impress rather than to inform. It describes values and ambitions rather than facts. Such content is rarely quoted, because there is nothing in it to quote.

They are complete

Depth matters. Thin pages give the assistant little to work with. A page that covers a subject thoroughly, anticipating the follow-up questions a person is likely to ask, is more likely to be used, and more likely to be used for several parts of an answer rather than just one.

They are consistent

When an organisation’s name, location, services or contact details differ across its website, directories, social media and news coverage, assistants have reason to be less confident in any of them. Inconsistency can result in an organisation being described vaguely, described wrongly or left out in favour of a source that appears more certain.

They are corroborated

Assistants weigh what others say about an organisation, not only what it says about itself. Independent coverage, reputable directories, reviews and references from other credible organisations all help an assistant judge whether a claim is trustworthy. An organisation that is described consistently by several independent sources is easier to recommend than one that only describes itself.

They are current

For many questions, recent information is preferred. Opening hours, prices, regulations and availability all change. Pages that are clearly maintained, with accurate dates and current details, are more useful than pages that appear abandoned.

They are readable by machines

Clear structure, plain language and properly described facts make information easier to retrieve and quote accurately. Headings that reflect real questions, short paragraphs that answer them, and structured data that identifies the organisation, its location and its services all help. So does making sure the content is visible in the page itself, rather than hidden in images, downloadable files or elements that only load after interaction.

Assistants reward information that is correct, complete and consistent, not slogans.

It is tempting to speak of AI search as if it were one thing. In practice, the main assistants behave differently, and an organisation that appears in one answer may be absent from another.

Some assistants search the web for almost every question; others rely more heavily on what the model already knows unless the question clearly needs current information. Some show their sources prominently; others summarise without obvious links. They use different search indexes, weigh sources differently and are updated on different schedules. Even the same assistant can give different answers to the same question asked in slightly different words, or in a different language.

For this reason, checking a single assistant once tells you relatively little. A useful assessment asks a consistent set of questions, across the assistants your audiences actually use, in the languages they actually speak, and repeats the exercise over time. The pattern across many answers is far more informative than any single response.

Being well known is not enough. We regularly see established organisations described inaccurately, or passed over for an intermediary such as a booking platform or directory, because the intermediary publishes clearer and more complete information than the organisation itself.

There are several common causes.

• The website says little that is specific. It describes the organisation in general terms but does not answer the questions people actually ask.
• Important information is hard to reach. It sits in downloadable brochures, image-based pages or sections that require several clicks.
• Old information persists elsewhere. An outdated directory entry or an old article continues to be found and repeated, and nothing more recent contradicts it clearly.
• Intermediaries are more thorough. Platforms that list the organisation have published more detail about it than the organisation has published about itself.
• No one is looking. The organisation has never checked what assistants say, so errors go uncorrected for years.

The reverse is also true. A smaller organisation that answers its audience’s questions thoroughly can be recommended ahead of much larger competitors. On our own websites, which are relatively new, AI assistants now bring one in four visitors. That did not happen because the sites are large or well known. It happened, we believe, because they answer travellers’ questions clearly and completely.

Because AI search is new, it attracts confident claims. A few are worth addressing directly.

That it is simply SEO under a new name

There is real overlap. Many of the practices that help search engines understand a website also help AI assistants. But the goal is different. Search engine optimisation aims to rank a page among ten links. AI search is about how the organisation is described and recommended across many sources, including sources it does not control. An organisation can rank well and still be described poorly by an assistant.

That recommendations can be bought or guaranteed

No one outside the companies that build these systems can honestly guarantee that an assistant will recommend a particular organisation. Answers vary between assistants, between phrasings of the same question and over time. What can be done is to improve the quality, consistency and visibility of the organisation’s information, and to measure the effect.

That more content is always better

Large volumes of thin or repetitive content, especially content written for machines rather than people, rarely help and can do harm. Assistants are looking for the best answer, not the most answers. A smaller number of thorough, accurate pages usually outperforms a large number of shallow ones.

That it only matters for consumer businesses

Investors, students, patients, journalists and officials all use AI assistants to form early impressions. For institutions whose reputation abroad matters, what an assistant says may be the first thing an important audience learns about them.

The practical response is straightforward, even if it requires sustained effort.

1. Test what assistants say today. Ask the questions your audiences ask, in the languages they use, across the main assistants. Record the answers so you have a baseline.
2. Fix inaccuracies at their source. Outdated information usually comes from somewhere specific: an old directory entry, a stale page, an out-of-date article. Correcting the source is more effective than publishing a contradiction elsewhere.
3. Publish the answers people need. Write for the questions your audience actually asks, in depth, on your own website. Make the facts easy to find and easy to quote.
4. Keep facts consistent everywhere. Names, locations, services and contact details should match across every channel you control and, as far as possible, the ones you do not.
5. Earn independent references. Coverage, listings and endorsements from credible third parties help assistants trust what you say about yourself.
6. Measure over time. AI answers change. Track them monthly, the way you would track any other part of your reputation, and track visitors and enquiries arriving from assistants alongside other channels.

Most of these steps are not technical. They depend on clear ownership, accurate information and cooperation between the teams that publish it. That is why AI search is, in the end, a leadership question rather than only a marketing one.

<- All Insights

hello@mudahtemplates.com.my

+60 16-840 7481

Lot 25-3(R), 3rd Floor, Beverly Hills Plaza, Jalan Bundusan, 88300 Kota Kinabalu, Sabah, Malaysia