Influencing What LLMs Recommend

It is easy to assume that when an AI assistant recommends something, it is doing what a person does. You ask a question, it thinks, it forms an opinion, and it gives you a neat little answer.

But most “recommendations” from tools like ChatGPT, Copilot, or Google’s AI summaries are better understood as something else. They are usually a summary of what the system can confidently assemble from the information it can access.

That difference sounds subtle, but it changes how you interpret almost everything that follows.

If the internet is the world’s biggest library, an AI assistant is not a new author. It is closer to a librarian who reads very quickly, stitches together a response, and tries to be helpful. And like any librarian, it depends on what is on the shelves, how clearly it is written, and which books look trustworthy.

This piece is about that dependency. Not how to game it, and not how to sell into it, but how to understand it in plain language.

1) What “AI search” is really doing

Traditional search has a rhythm most of us could describe without thinking.

You type a query. You get a page of links. You open a few tabs, compare, and decide.

AI search changes the middle of that process. Instead of handing you ten blue links and leaving the comparison to you, it tries to do the comparing for you and return one combined answer.

You can see this happening in a few places:

Search engines that show an AI summary above the results.
Chatbots that retrieve sources before answering.
Assistants embedded in operating systems and apps.

The interface changes, but the behaviour is broadly the same. Fewer clicks, more synthesis.

That is why people now talk about “being cited” or “being included.” If the answer is a summary, the most influential position is not the top link. It is being part of what the summary is built from.

2) Why assistants lean on some sources more than others

A useful way to think about it is that AI systems are always trying to do three jobs at once.

First, they have to find information that looks relevant.

Second, they have to extract claims that sound concrete.

Third, they have to explain those claims in a way that reads cleanly.

In practice, this pushes them toward information that is:

clear, because the point is easy to lift
consistent, because multiple sources seem to agree
specific, because it uses facts, numbers, and named things
repeatable, because it can be paraphrased without changing the meaning

It also tends to push them toward sources that look independent.

A company’s own website can be accurate, but it is obviously self-interested. Systems that are trying to sound neutral often lean more heavily on writing that appears to be about an organisation rather than by that organisation, because it is easier to treat as a reference point.

In everyday terms, that means mentions in established media, industry analysis sites, third-party reviews, and reputable databases can carry more weight than a polished “About” page.

When an assistant provides citations, those citations are often your best clue about what the system considered stable enough to reuse.

Industry analysis has started tracking these patterns across AI platforms, which shows that “authority” can look different depending on which assistant you are using.[1]

3) Reviews, mentions, and the role of “what other people say”

One of the simplest insights about AI recommendations is also one of the oldest insights about human decisions. People look for social proof. They want to know what others experienced, what experts noticed, and what problems show up in real use.

This is where PR and media mentions matter, but not in a mystical way.

A press article, a product roundup, or an independent review is usually written for readers who are trying to decide. That kind of writing tends to include the exact statements an AI assistant can reuse without guessing:

what the product is
what it competes with
what it roughly costs
what it is good and bad at
who it suits

Assistants often mirror that, because reviews and third-party mentions contain three things that are hard to fake and easy to reuse.

They contain comparisons.

They contain trade-offs.

They contain context, like who something is for and who it is not for.

And that context is exactly what people ask for when they use AI tools to research purchases. In 2025 reporting on consumer shopping behaviour, people described using AI chatbots to compare options, find deals, and summarise reviews.[2][3]

So when someone asks an assistant for “the best” option, it is common for the assistant to pull from material that already sounds like a comparison.

Not because it “likes” those sources, but because those sources are formatted in a way that makes the assistant’s job easier.

4) The quiet technical constraint: assistants need information they can safely reuse

Humans are comfortable with ambiguity. Someone can read a messy page and still come away with a clear understanding.

AI systems are less comfortable.

They need information to be extractable, because they are not really reading in the way people do. They are retrieving, selecting, and assembling. That is why conversations about AI search so often drift toward ideas like:

entities, like a product model or company name
attributes, like features, pricing, or compatibility
relationships, like what competes with what and what fits which use case

When that structure is present, the assistant can “ground” the response more confidently.

Interviews with AI search companies describe this as a retrieval-and-generation pipeline: retrieve sources, pull out relevant claims, and then generate an answer that tries to match the evidence.[4]

From the outside, it looks like a conversation. Under the hood, it behaves more like a high-speed research workflow.

5) Why regulators are paying attention (and why it matters)

Once a small number of systems become the main way people access information, the question shifts.

It stops being only “is it useful?”, it becomes “what power does this create?”

This is one reason competition regulators have been writing about AI foundation models and AI products through the lens of market structure and dependency.

In Australia, the ACCC’s 2025 paper reads less like a rulebook and more like a map of what could become sensitive: partnerships, acquisitions, data access, and the way AI changes how people browse and decide.[5]

Internationally, regulators have also issued joint statements about ensuring effective competition in generative AI foundation models and AI products.[6]

The core idea is simple: If AI becomes a dominant front door to the internet, the way it selects and summarises information is not just a technical choice. It becomes part of how markets and public knowledge are organised.

6) A calmer way to think about “influencing what LLMs recommend”

The phrase “influencing what LLMs recommend” can sound like a trick. A more honest framing is that assistants can only recommend what they can recognise, understand, and justify.

And a lot of what they can justify is shaped by third-party writing. If the only place a claim exists is a company’s own website, the assistant has very little to cross-check.

If the same claim shows up in independent coverage, even in small mentions, it becomes much easier for the assistant to treat it as a real-world fact instead of self-description.

When an assistant cannot find stable, consistent, clearly stated information about something, it will either avoid mentioning it or talk about it vaguely.

When it can find stable, consistent, clearly stated information, it usually becomes more confident and more specific. And as AI search becomes a mainstream way people research and decide, that constraint becomes part of how everyday buying decisions get shaped.