Ep 5.5: What Makes AI Cite Your Brand? Here’s What We Found

00:00: Genny: Why does it feel like every time we do an episode about AI search optimisation, or generative engine optimisation, or GEO, or whatever you want to call it, everything (well, not everything) we thought we knew about this new marketing channel has changed before our very eyes?

The short answer is that, in many ways, it has! Silly me!

Since I last made your eyes and ears bleed back in July, we ran an AI citation study in which we analysed over 11,000 citations from over a thousand questions, and what we observed about the sources AI systems cite was pretty insightful, and sometimes even surprising.

So on today’s episode, we’re going to go through our observations with a fine tooth-comb, or maybe more like a wide-toothed rake since we don’t want to be here all day. 

I will, of course, link to the case study in the show notes if you’re listening on your headphones or in the comments section below for anyone staring at my face right now.

But before I get started, a lesson in correlation vs causation. The study we conducted produced a few different types of findings: patterns, comparative associations, measured differences, or counter-patterns. For various reasons, such as the fact that the study captured just a snapshot of AI search behaviour and the fact that we didn’t experiment with each variable in isolation, we need to treat these findings as observations rather than laws, because we cannot say definitely that X causes higher citation rates, merely that X appeared more often in the citations we analysed.

So, without further ado, let’s see what our study dredged up from the sea and what we suggest you do as a result.

First observation: AI Cites Self-Contained Answers

The clearest pattern we noticed was that AI systems tended to cite passages that could stand on their own two feet. No burying the lead, no cliffhangers, just concise, self-contained explanations. 

For example, the median passage was 56 words. So the goal is to make sure that whatever question you’re answering is unmistakable when someone lands in that section of the page. Now I know that’s not super conducive to creative writing, but chances are you’re probably not writing about anything that exciting to begin with.

No offence!

Second observation: Cited answers are easy to locate on your page.

Citations often came from a quick paragraph or list placed near a clear heading, most commonly right under the heading. In the study, we noticed that ChatGPT did this more often than Gemini. 

This doesn’t mean you should turn all your content into a long FAQ section. Rather, it’s pointing to the idea that each section should have a clear job to do and do it.

I really do feel sorry for all the creative writers out there relegated to the ranks of content creation because this style of writing leaves little room for style and rhetorical devices. Where’s the metaphor? Where’s the stretched out rhythm of polysyndeton?

Well, good news for all Hemingwayites. Enjoy your days of glory, you minimalist bastards.

Third observation: Cited pages tend to be comprehensive. 

Get out of town. The cited pages in our study weren’t typically thin, with just over 2,100 words, 27 headings as the median. 65% of cited pages featured FAQ-style headings, and 41% had at least one table. But don’t confuse length and formatting for quality. Pages that were cited five or more times did not look structurally different from pages cited once. What distinguished recurring sources was topical coverage and brand recognition, not simply more formatting or word length.

By the way, that’s what I meant before about the difference between correlation and causation. That’s like saying that since shark attacks increase when ice cream sales go up, that ice cream causes shark attacks. 

In reality, it’s warm summer weather that creates the conditions that make both outcomes more likely. When people want to cool off, they buy ice cream and jump in the water, thus enticing the great whites with a veritable buffet of tasty human flesh.

Come on guys, we all know the real reason for shark attacks: You’re gonna need a bigger boat

Fourth observation: AI checks your reputation outside your website

AI search engines often look outside company websites when the question is related to the brand’s reputation. In fact, review platforms made up 36% of cited URLs for brand-reputation queries, compared with 6% for other query types. That creates a tricky problem for B2B brands, because the sources shaping an AI answer about your reputation often sit outside your own website. 

Now, I hate to state the obvious here, but it kinda points to this crazy idea that if you want to make sure AI search engines cite nothing but good reviews, maybe that means trying a little bit harder not to suck. Just spit-balling though. 

Fifth observation: Different AI platforms cite different sources.

One of the clearest lessons from the study is that there is no single, stable list of websites that “AI likes.” Different systems look in different places. Across the tests, ChatGPT and Gemini shared only 5% of their cited domains on average, and in 58% of searches, they didn’t even cite a single shared domain. So a page that performs well in Gemini may not be the page ChatGPT chooses, even when both are answering essentially the same question.

Sixth observation: AI citations are often highly localised

Geography also plays a factor. When we ran the same queries in both the UK and the US, the AI systems rebuilt their source lists much more substantially than conventional Google search results did. That makes sense, since pricing, regulations, providers, review platforms and language (obviously) can all vary by market.

The more of the story here is that building a generic page might only get you so far on your international website. Think more about how you can make your page appear as if it was written by the locals for the locals.

Seventh observation: A slight shift in wording makes a big difference. 

We grouped related searches around the same topic, such as things a user might plausibly ask before, after, or instead of the main search. 

Just one seed query surfaced a median of 13 cited domains. Looking across the related-query group surfaced 60 cited domains. A fan shape, if you will. That’s nearly five times as many. 

What difference does that make to you? Well, nothing except for the fact that if one prompt around a single topic tells you very little about the full citation landscape. To fully understand citation behaviour around a topic, you’ll have to broaden your prompt tracking to cover related questions.

Now, from a content writing perspective, you could hypothesise that writing about a specific topic more broadly will likely help you show up in more answers. So for example, if your topic cluster is around ‘ice cream and weight loss’ (hey, why not?), then writing pages revolving around this topic could increase your chances of a citation. 

So, if your topic cluster is ‘ice cream and weight loss ’, then your related pages might be something like ‘5 Guilt-Free Ice Cream Flavours’ or ‘Ice Cream vs. Frozen Yoghurt: Can the Lactic Acid Effect Help With Weight Management?’

I think I talked more about topic clusters in a previous audio-only episode, and I’m fairly certain it hasn’t become obsolete just yet, so I will be sure to include the link in the comments.

Now let’s bust some myths 

While our study is really only a snapshot in time, it did not corroborate some of the accepted beliefs about AI search behaviour, so without further ado, let’s clear the air.

Myth: AI trusts neutral editorial sources? Who the hell thinks that? According to our study, AI citations showed 75 to 82% commercial bias, which means the cited page recommended something that it itself was selling. That doesn’t mean AI only cites company sites, but it does challenge the assumption that it only trusts neutral editorial sources.

Myth busted!

Myth: Platforms like Reddit are taking up all the citation space. Actually, in our study, Reddit and other user-generated content platforms only made up 4% of all citations in this B2B dataset. Now, I have to insert B2B because that’s what we were looking at. I can’t say that if we ran a test of 1200 queries based on A Song of Ice and Fire fan theories, it wouldn’t bump up that 4%. But I will say, in the context of this study, that Reddit myth is busted!

Myth: The stronger your backlink game, the higher the likelihood that you will be cited by AI search engines. So if you’re coming from a traditional organic search background, you know that getting high-quality backlinks from reputable sources is a strong authority signal that Google search algorithms lean on heavily when ranking pages.

But in the study, the median AI-cited page had zero page-level backlinks, and 53% of them didn’t even have referring domains. I know! It’s like…excuse me.

Don’t throw out your backlink strategy completely because, A) that would be a tad hasty; and B) even though AI search is growing rapidly, traditional organic search still holds the vast majority of market share, so you still need those backlinks.

But for the purposes of this list, consider that myth busted!

Myth: If you want to be cited in ChatGPT, you should focus more on your Bing search strategy rather than Google. This myth makes tons of sense to me. Microsoft is an OpenAI investor and partner, not to mention a fellow player in the fun little game called global ecosocial annihilation, in which being a billionaire and a total douche are the only barriers to entry, so it stands to reason that ChatGPT’s algorithm would align more closely with Bing Search. 

But actually, ChatGPT citations overlapped with Google’s top 10 search results three times more than Bing’s. Now, of course, the winner for the strongest overlap between Google SERPs and AI citations obviously goes to Google AI Overviews, but we don’t need to explain that clear case of nepotism.

But even after all of that, the study showed that there actually wasn’t that strong of an overlap between pages cited and top 10 search results. That’s interesting. That tells us that while Google rankings are related to AI visibility, they’re not the whole picture. Ranking in the top three improved the likelihood of a citation, yet many cited sources sat outside Google’s top 10.

Mind blown!

Myth busted! Busted! Busted!

I think I’m actually going to leave it here because I think I would like this last point to resonate a little bit. It seems like while there is obviously a very strong overlap between SEO and AI search optimisation techniques, studies like this show that if you’re keen to get your AI search visibility off the ground, your SEO should probably be treated more like a jumping-off point rather than a one-size-fits-both strategy. 

With that said, be sure to tune in next time, where I’m confident the rules about AI search will have shifted yet again, but before you do that, don’t forget to read the full AI Citation Study linked in the comments, where you will find many more gems to marvel at. 

And not to sound like a broken record, remember this citation study is a snapshot in time, which is important to consider when working on your strategy because it will take time to see the trends in your visibility. That’s even more of a reason to utilise prompt-tracking tools like GeekyIQ (hey, I haven’t plugged that once yet), which do the heavy lifting for you.

Our next episode and last episode of the year will be released some time in November, so until then…be sure to…something…I never know how to sign off these—