A Korean eyewear maker’s ROAS rose 39% when Jellyfish introduced insights from its Share of Model tool into the brand’s Google advertising.
Gentle Monster, which sells upmarket (and often weird-looking) glasses, used Jellyfish’s LLM insights platform to optimise its pre-Christmas campaign on Google Performance Max (PMax) in the US.
Specifically, Jellyfish used information from its Share of Model platform to augment the search themes — the terms that potential customers might use to find a certain product — that Gentle Monster put into PMax at the beginning of the campaign.
In an article published in the Journal of Brand Strategy, John Dawson, VP of strategy at Jellyfish USA, and Jack Smyth, the Australia country lead at the Brandtech Group (which owns Jellyfish), wrote:
‘This produced a 17% improvement in click-through rates and a more than 14% uplift in conversion rate, indicating that the AI-optimised search themes were more relevant and attracted higher quality traffic. Overall, this resulted in a more than 39% improvement in return on ad spend.’
A simplistic way to explain the Share of Model tool, says Dawson, is that Jellyfish quizzes the big LLMs as if they were a focus group, asking what they think about different categories or brands.
The agency then uses a process it calls ‘embedding’ to group the responses into a complex word cloud, from which it can derive search themes. When these themes frame brands or product-use occasions in novel ways, it’s often because the LLMs have picked up on customer needs that marketers haven’t yet identified, and they can create new targeting opportunities in Google ads.
In the Journal of Brand Strategy article, Dawson and Smyth wrote that Jellyfish used the same method with MSC Industrial, a supplier of industrial equipment, and improved its incremental ROAS with PMax by 758%.
The ambient era of marketing
These case studies offer a glimpse at best practice in what Dawson and Smyth call marketing’s ‘ambient era’.
While the internet of the 2000s (the direct era) created an imperative for brands to guide people to their websites, and the platforms of the 2010s (the distributed era) forced them to create content for the places that people were congregating online, the LLMs of the 2020s (the ambient era) have introduced a new audience into the marketing equation, says Dawson, one that is constantly evaluating and contextualising information about brands.
It was this ambient era that Share of Model was built for, says Dawson, adding that the tool is also able to ‘identify the preferences and behaviours that these models exhibit’ and deliver campaigns that appeal to the LLMs.
When dealing with LLM chatbots, says Dawson, the main metrics of salience are mention rate (how often models will cite a particular brand) and average position (how far down in a list of competitors a brand appears).
Brands will perform differently against these measures within different models, and even within the same model, a high mention rate will not always equate to a superior position in a list of brands. On its own, neither metric is sufficient to tell a brand where it sits in the LLM’s pecking order. But when the mention rate and average position are considered together — and combined with some semantic analysis — brands can understand how they are perceived by the LLMs, and then ‘attack the limiting beliefs that these models have about a brand’, says Dawson.
In this ambient era of marketing, organisations must think about their brand in much broader terms than they are used to, says Dawson, because LLMs take into account information that most people never consider. When, for example, Jellyfish probed LLMs’ perceptions of financial institutions, it found that weight was given to how bureaucratic they were, as well as more obvious factors, like how expensive they were.
‘Historically, investor relations, ecommerce, HR and all these other teams developed content and tactics that sat in different areas of the internet and were visible to different people by virtue of their interests,’ says Dawson. ‘That’s all changing because the LLMs understand the full context of the brand, and what you do in one area can affect another.’
Understanding and manipulating LLMs to influence humans is only one aspect of the ambient era of marketing, as Dawson and Smyth see it. The bigger question, says Dawson, is ‘how do we think about the role of communications when these models provide not just recommendations but independent actions for the benefit of an end user that may not even be human? That’s really the paradigm we’re in.’
Fully agentic shopping, in which agents transact on behalf of a person or organisation, is not a foregone conclusion. There is no evidence yet that people want to shop that way, and analyst Eric Seufert has pointed out that there is little incentive for either brands or retailers to give up their relationship with consumers for the benefit of AI companies.
Dawson cautions against using analogies with the past to predict the future, however. ‘I think Eric is one of the smartest people and I don’t disagree with him. But I think a neater way to look at what might be happening is the agents that Ramp, the B2B finance company, launched last week.
‘Ramp agents use existing data to basically identify procurement [opportunities for clients]. It’s not agentic shopping in the classical sense, but it still frames and pre-decides the set of decisions that businesses will make.
‘And if I think about my own behaviour, I use [Google] Gemini a lot for recipes. And now, when I’m in the supermarket, if I say, “what can I do with this vegetable?” it will recommend a whole lot of options for me. Is that agentic commerce in the way that Eric rightly critiques it? No, but it’s agents involved in commerce, and I think that is going to be a really big part of the future. Not just for small decisions, like what to cook for dinner, but very big ones, like what B2B provider will get my multi-million dollar enterprise contracts.’
