How do AI overviews work?

Google is always looking for ways to develop its search platform. Whether it’s algorithmic changes underpinning its operation, or user-focused tweaks to improve functionality, it rarely stays the same for an extended period.

AI overviews are the newest feature to be added to Google search and have caused quite a stir in the SEO industry. Here, we explain AI overviews, how they work, and everything you need to know now.

What are AI overviews?

Google’s AI overviews are a new feature that sits within search engine results pages. They provide users with a summary of their searches, with information generated by large language models (LLMs) which are informed by subsequent search results.

AI overviews used to be called “Google’s Search Generative Experience”, which was a sandbox feature that Google tested in US markets. Like many other Google tools, it started with a beta program before being rolled out to the masses towards the end of 2023. That beta has now ended, and AI overviews have become a consistent presence in Google searches for all of its users.

The premise is simple: users continue to use Google in the way that they always have, except that at the top of some search results, they will receive a summarised version of the results that follow.

To be clear: AI overviews aren’t like ChatGPT, or Google’s Gemini for that matter – it’s not a chatbot, and users aren’t able to refine their queries further beyond what’s initially provided.

A quick note on LLMs

Before we go too heavily in-depth on AI overviews themselves, it helps to have a quick recap on how large language models operate.

At their core, LLMs are always trying to predict the next word in a phrase.

If you were to take the phrase “the quick brown fox jumps over the lazy…” and guess what word is missing, you’d probably guess “dog”, and you’d be correct in doing so. That’s the very core of an LLM – a predictive text tool.

LLMs use massive banks of data to estimate and guess what these subsequent words are. Unlike things like calculators, which use a deterministic approach to solving questions (facts, logic), LLMs take extremely educated guesses to predict what follows.

There are a few parameters that can be tweaked within an LLM to alter its predictions. They’re given a bunch of different names, but they’re commonly referred to as ‘temperature’ and ‘top P’. High temperatures mean more creativity with word choice, whereas a high top P value will result in more variation in overall sentiment and responses. Both these factors go hand in hand.

It’s like rolling dice down a cobbled road. You can never quite tell where it’ll end up, you can’t predict it, and there are elements of randomness. LLMs are designed to give different responses based on chance.

How AI overviews work

The presence of AI overviews in search is two-fold:

Organic searches

For organic, informational searches, AI overviews consider top-ranking sources of information to inform their responses. At the very top of a search result, a generative search summary will appear at the very top of the page (even above pages ranking in 1st place). What follows is a relatively short summary of the answer to said query, with links directing users towards more sources for further reading.

Shopping

AI overviews are also present in Google Shopping. When a user looks for a product, say a mountain bike, for example, AI overviews will suggest considerations and caveats for selecting the right mountain bike for their needs. In addition to these considerations, AI overviews will even directly suggest products that closely fit the “brief” (search criteria).

Limitations of AI overviews

While AI overviews do solve some browsing blunders for those using Google, it does come with their limitations and drawbacks:

YMYL restrictions

Google has, for a long time, categorised some search results as being associated with two key things in life: money, and health. This concept, which Google refers to as “your money or your life (YMYL)”, was introduced in 2014 and still has quite a large bearing on how Google works as a search engine. Essentially, any searches deemed as being YMYL-oriented are treated slightly differently from other searches, with search quality evaluators having specific guidance on how to gauge the quality of said results.

Google has openly admitted that it attempts to filter out AI overviews being delivered for YMYL queries. Rather, Google prefers to let verified experts comment on such subjects.

Hallucinations

Hallucination isn’t a concern that’s bound purely to AI overviews. All large language models (LLMs), chatbots, or whatever you wish to call them, will suffer from some degree of hallucination.

Hallucination is a term used to describe when the outputs given from an LLM begin to stray a little far from reality. This can include conundrums like inventing false sources, telling false information, or just giving completely incorrect responses in general. As you can tell, this can be extremely problematic on a platform like Google Search, a platform where users are previously accustomed to information being provided by valid, trustworthy sources. If AI overviews hallucinate, users could start distrusting Google Search in general – a major risk.

Size

Although less of a drastic concern, AI overviews have been known to take up significant portions of screen size. Some studies conclude that the viewports of AI overviews can be as high as 3,200 pixels – that’s more than most screen resolutions.

Placing AI overviews so prominently within a search result can mean more time is spent reading through the results rather than clicking through to individual pages. This, at least in this interim transitional period, can mean websites will see lower organic traffic.

More marketing wisdom, from the experts

So, you know how AI overviews work, and now you’re thinking of mastering them. Liberty has had its eye on AI for a while, and we’re constantly hunting out strategies to position your website higher. Contact us to chat about your goals – we’re happy to help you reach them.