It’s no secret that AI tools like Claude, Gemini and ChatGPT have disrupted the search ecosystem. According to research published by Search Engine Land, approximately 71.5% of people use - or have used - these large language models (or LLMs) to retrieve information, ask for a recommendation or solve a pressing problem. 

Traditionally, this activity would have been routed through a search engine like Google or Bing. These search engines generate results pages (or SERPS) with a ranked list of websites that purport to answer the user's query — giving savvy businesses who invest in search engine optimisation (SEO) a chance to compete for eyeballs and clicks. 

But LLMs don’t have listicle-style results pages. Instead, they generate 2-3 paragraphs of summarised content that may (or may not) directly reference brands that are relevant to the user's question. In many cases AI results will also contain a list of cited sources that it's supposed to have drawn on while compiling a summary, but we'll get on to that later. 

For now, let's home in on the way AI is reshaping search behaviour. Anecdotal evidence published by HubSpot suggests that it’s top-of-funnel, informational searches that are taking the biggest hit. 

For clarity, these are searches like "what is digital marketing?" or "How do I clip my dog's nails". Experts estimate that these searches account for about 88.81% of all searches made via AI, presumably because people trying to solve a specific problem prefer the simple, summarised answers provided by a large language model. 
 

AI search is reshaping the marketing funnel

There's also some evidence to suggest that mid-funnel, research or comparison type searches are being hit. Think search behaviour like "who are the best events agency?" or "is Salesforce better than HubSpot?" which is prime real estate for anyone that's serious about snagging high-intent traffic. 

In short then, people are using generative AI to query the web, answer questions and compare brands. 

If we want to grow awareness and maintain our share of voice, we need to make sure we're showing up in AI search summaries, and that means developing a reliable strategy for generative engine optimisation (increasingly shortened to GEO). 

To help you do this, we've pulled together the latest thinking on the subject; drawing on research papers and expert commentary to explain concepts like vectorisation and understand the relevance of authority, content structure, natural language and trust signals.

This is a nascent and therefore noisy space. Any number of people claim to be able to help you improve brand awareness - or brand placement - in an LLM. 

They often lean on arcane or esoteric references to ‘knowledge graphs’, claim that latent semantic indexing is key to cropping up in AI summaries, or make vague statements about the undeniable correlation between well-structured content and inclusion in AI search results. 

None of this is (strictly speaking) true. In fact, the idea that you’ll be able to reliably ‘place yourself’ in a generative search result is inherently flawed because these models are probabilistic and therefore inconsistent. 

Ask one to recommend a brand of lawnmower and it might tell you to buy Bosch. Wait five minutes and ask it to recommend a lawnmower manufacturer and it’ll tell you that you that Toro’s the only choice and the seemingly-random nature of these responses renders any effort to ‘dominate’ generative search an essay in futility. 
 

So, what can we do?

We can increase the likelihood that our informational content will be picked up and referenced by AI tools like ChatGPT or Google’s search AI. 

Positioning yourself as a trusted and reliable authority will increase your chances of being referenced in AI summaries, and it's important to remember that the content used to generate an AI summary is also cited as a source. 

Ultimately, the combination of brand mentions and direct citations will push people into your arms: increasing referral traffic from LLMs and growing mid-funnel, branded search activity that you can convert on your website. 

This isn’t a vain hope either. People’s inherent distrust of AI means that they do like to verify sources and clients who get regular AI mentions do see a steady stream of AI referral traffic in their Google Analytics dashboards. This traffic may not be ready to buy straight away, but we can retarget it, nurture it, and work on turning it into a source of value further down the line. 

To be clear, nobody outside Google Labs, Anthropic or OpenAI actually knows how to increase your chances of showing up in generative search. Anyone claiming otherwise is over-exaggerating their own competence. AI tools like Claude or Gemini are black boxes, and search marketers simply aren’t privy to their inner workings.

That said, we can test some of the ideas floating around on the web, draw on other people’s research/data and mine technical documents for hints at the underlying workings of these incredibly complex data models. 

In time, this science-based approach will yield ironclad best practices that mirror the tried-and-tested tactics we have for traditional SEO, but while the industry is working its way towards that lofty goal, we can still take an evidence-based approach to optimising written content for generative AI -- using the knowledge we have gleaned to maximise our chances of appearing in AI summaries. 

It’s this approach that we’ve outlined below, starting with the foundational blocks and working our way towards the fringe ‘GEO’ ranking factors that may or may not actually impact on our efforts. 

Fair warning: this is a fairly exhaustive breakdown that draws on six months of research and delves into the inner workings of modern large language models , but if you can stick with us through to the end, you will have a much clearer understanding of how and why content gets picked up by LLMs, and a solid plan for improving your own contents’ chance of appearing in AI summaries. 

We’ll dive straight into the deep end and look at the way large language models process information. Contrary to popular opinion, AI doesn’t understand words or concepts in a way we’d recognise. 

The  Institute of Analytics puts it best:

“Contemporary AI systems, including state-of-the-art language models do not understand language in any cognitive or semantic sense. These systems operate by identifying statistical patterns across massive datasets. Models such as GPT-4, Claude and Gemini do not contain mental representations or grounded concepts. They do not know what a tree is or what it means to be cold. Instead, they generate likely continuations of text based on distributions learned during training.”

To expand on this, Large Language Models are using complex maths (natural language processes or NLP) to generate the next most likely words in a sequence. They do this via two distinct processes: Tokenisation and vectorisation.
 

Tokenisation and vectorisation explained

To begin with, LLMs break phrases, or parts of phrases into chunks and assign them a numerical value. Duck might become ‘3’, tree ‘4’, and ‘the’ 16. 

LLMs then trawl an incredibly large database of so-called ‘training data’ to model the statistical relationships between its tokens and use this information to build or embed these correlations as vector embeddings that it can use to predict likely sequences of words.

This approach may sound rudimentary, but when you do it on a vast scale -- using thousands of terabytes of data and hundreds of coordinates to work out how, for example, the presence of the word ‘duck’ in a phrase might modify the verb used six or seven steps down the chain, you get to a point where you can reliably generate text that looks and feels ‘correct’ in the sense that it appears to make sense and offer a coherent response.

As Rodney Brooks, Professor of Robotics (emeritus) at MIT, points out, this ability to produce human-like text is easy to mistake for a broader competence. 

People often make the mistake of thinking that ChatGPT or Gemini understands the questions they’re asking it, and works to produce a real answer when it’s just looking at a string of tokens and saying “if this is the input, what is the most likely output? What tokens should it contain, and in what order?”

Fundamentally, we need to remember than an LLM isn’t trying to offer a good, accurate or correct answer: it’s trying to accurately imitate a passably ‘human’ output based on the conditions proscribed by the model’s programmers. 

Incidentally, this is basically the premise of the Chinese Room; an old thought experiment that goes like this: 

An English speaker with no Chinese language skills is locked in a room. The room is empty, except for pens, blank paper cards, erasers and a large book of instructions that walks you – step by step – through the process needed to manipulate Chinese symbols in response to specific inputs. The book is in English, but it allows the person trapped in the room to act as though they understand Chinese. 

When someone outside the room slips a question under the door, they get an accurate response written in Chinese. From their perspective, the person in the room clearly understands Chinese. But we’d probably say otherwise. 

Now, this is a practical guide for ranking in generative search, so we don’t really want to embark on a philosophical journey for the true meaning of intelligence (or comprehension). But it is important to understand how these models operate because it is that understanding that allows us to manipulate their output. 

Broadly speaking, if we want to make sure that AI uses our content as a resource, we need to start by ensuring that:

• We are using words and phrases that are semantically linked to the prompts our audience will use when querying an AI search tool
• LLMs can understand what we are writing about
• Tools like ChatGPT and Gemini see our content as a comprehensive resource that speaks in detail about the subject it’s trying to write a summary for. 

To a degree, this is just about writing clearly and using words and phrases that closely align to a given model’s expectations vis-à-vis your chosen topic. 

Drawing on its vector maps, AI would expect an article about keeping ducks to talk about ponds, water quality, food and reeds. If we want our content to be referenced when someone asks Google how to keep ducks, we best be sure we tick all those boxes instead of, say, ramming our article full of various synonyms for duck.  

Strategic keyword research gets the ball rolling, and after that, it’s about thinking broadly: We want to produce a structured article that addresses intent and uses language naturally, but we also want to be far-ranging in scope and focus -- so that we naturally include a lot of the tokens associated with our chosen subject. 

There’s also some (anecdotal) evidence to suggest that it helps to write in a conversational or clear tone, structure content around specific questions or themes, and use an active voice to encourage tools like Chat GPT or Claude to reach the conclusion that your content is both authoritative and easy to read.  

According to Crescat and Sulu, structured data (where you mark up specific entities using schema.org vocabulary) may also help LLMs quickly scan, understand and categorise your content.  

Again, we’ll stress that this is all foundational: Writing clearly and ticking the right linguistic boxes is important because it gets your foot in the door, but it’s not enough to push LLMs into using your content. 

A lot of the so-called ‘GEO wisdom’ currently circling on the web is eerily similar to conventional best practices for SEO, but that doesn’t mean we should discard it out of hand. Talk about authority is a prime example. 

On the surface, the idea that a probabilistic model would care about – or even comprehend – concepts like authority seems extremely far-fetched, but if we introduce the concept of grounding, everything starts to make a little more sense.
 

Grounding explained

If you think back to the heady days of ChatGPT 2.0, you may remember that early AI models were hilariously inaccurate and frequently ‘hallucinated’ in nonsensical ways. To kerb this behaviour, programmers sought to ‘ground’ LLMs like Chat GPT 4.0 by forcing them to search the internet for relevant content before they formulate a response. 

This is done using a technique (or family of techniques) called retrieval augmented generation (RAG) to scrape relevant information from the web, vectorise it and then use that vectorised data to enrich or alter a response. 

This process is governed by retrieval models, built to try and ensure that LLMs use accurate and trustworthy sources. Now, retrieval models are unique, and each individual model will have its own set of criteria, but most do look at something called MMR or mean reciprocal rank, which is a measure of how quickly sources appear in search results for an undisclosed number of relevant queries.

The idea being that a source with a high MMR is more likely to be correct or reliable. 
 

Wait, good SEO = good GEO? 

This is an oversimplification but yes; making sure content ranks well in traditional search probably improves your chances of appearing in an AI summary. 

Most modern search engines used a link-based algorithm. There’s obviously a lot more to traditional SEO (which is what we are discussing now) but the easiest way to make sure your content ranks well is to ensure that it has plenty of inbound links from respected and topically relevant domains.

Why? Well, imagine that you’re a search engine, and you’re trying to work out who to rank for an informational search like “what type of engagement ring is best for women who like classic jewellery?”

You’re looking at two candidates: One very thorough brand with a great page full of detailed content, but no trust signals, and another brand with an equally great page and links from the National Association of Jewellers, Jewellery Focus and the Luxury report.

Chances are, you’ll rank the second page and this is fundamentally true of real-life search engines too. 

Yes, there are other factors, but all else being equal, the page with the most citations will always win out and if we are trying to maximise our chances of ranking in generative search results, it makes sense to focus our energy on improving the most meaningful measures of authority. 
 

Heavy emphasis on the ‘could’ here. We see a lot of chatter about properly using h2 and h3 tags, keeping paragraphs to a set length, making heavy use to tables, or using bolded summaries to ‘force’ AI to grapple with your main points.

Unfortunately, a lot of this is advice is too specific -- and quite misguided.

To get to the crux of the problem with structure, we need to understand LLMs. Early models like ChatGPT-2 couldn’t parse (or understand) formatting cues at all, so they’d completely ignore heading tags, bullet points, bolding and other, structural elements.

Optimising for these models was all about semantic clarity; making sure that ideas were clearly addressed; that content flowed in a natural way, and that you reinforced (or repeated) important points to make sure AI tools understood what you were talking about.

Newer LLMs like ChatGPT-4 can parse HTML, although recent papers published by he Gaoling School of Artificial intelligence suggest that there’s still some debate about whether or not they do this when retrieving web content via the RAG processes we talked about before. 

Assuming that all AI tools will parse HTML in future, I makes sense to pay attention to formatting cues like <h> tags, <strong> tags and other HTML elements, while still applying all the advice about semantic clarity, breaking text down into simple chunks, and repeating ourselves ad nauseum that we mentioned before. 

Care must be taken here though: Some formatting cues are mor helpful than others. Headings are great, but the HTML formatting for tables actually creates artificial distance between the various entities (or words) contained therein, and studies published in 2023 suggest that AI tools really struggle to understand tables in general.  

In brief, focus on making sure that your content is chunked up nicely, and that you use summaries and other tools to reinforce ideas, use headings and bullet points, but don’t get too bogged down in prescriptive rules about paragraph length or content structure.

And above all, remember that parsing text is the one thing LLMs are good at. If you’re clearly answering a question and providing plenty of relevant detail, you should be fine. 
 

Again, heavy emphasis on the ‘might’. For those not in the know, E-E-A-T (expertise, experience, authority and trustworthiness) is a framework Google uses to evaluate how useful and/or user-friendly content is for traditional search, but its importance has always been slightly overstated.

While it’s obviously important to showcase all these things from a user perspective, E-E-A-Ts were never an algorithmic ranking factor, and it seems unlikely that LLMs would be conditioned or coded to seek out specific trust or experience signals in your content. Instead, we’d think of the whole E-E-A-T philosophy as generally good advice for writing content that engages users and improves their impression of your brand. No need to link that to efforts to appear in AI summaries. 
 

Most of the guides we read in the run up to this article claimed that quoting authoritative sources would help you rank in generative search results, but we couldn’t find any data to back up these assertions.

Again, this is generally good advice for writing authoritative content that inspires trust but doesn’t seem to have any bearing on your chances of showing up in an AI search result. 
 

There’s a marked uptick in the number of people using LLMs to query the web and navigate purchase journeys. While there’s no suggestion that these AI tools will completely replace traditional search in the foreseeable future, it’s still important to start thinking about how you can leverage them to ensure you’re reaching customers at key moments. 

As outlined above, this is a long game that demands careful thought. It’s also an esoteric science, which is to say that there’s a lot we still don’t know about the specifics. 

As a rule of thumb, it helps to write and structure content in a way that helps AI to navigate and understand what you’re writing, and it helps to send clear signals about your brand’s authority. Quoting trusted resources and writing in a conversational tone may also be helpful, but again, there’s less evidence for this. 

Above all, we think that it’s important to note that this is a long game. Don’t sweat it if it’s taking several weeks or months for your content to start appearing in generative results. Traditional search is still very powerful and while we think generative AI results will become more important as time goes on, it’s still possible to run highly effective marketing campaigns while you’re waiting to accrue the credibility needed to break into AI summaries.  

If you're looking for a competitive edge and want to maintain your share of voice, the digital team at Mearns & Gill are ideally placed to help you come up with a GEO strategy. 

Complete the form below or contact us to get started.