When calculating ROI, a fundamental question is ‘when do we do it?’ In other words, at what point do we decide that we’ve got everything we’re going to get out of the investment we’ve made, and that now it’s time to grade the work.
While there’s always an asterisk for the residual benefit to long term brand building, the vast majority of an activity’s impact is usually pretty apparent. Mixed media modelling boffins have developed, sweated, tested and debated a range of formulae for determining the rate at which advertising impact decays.
The IPA, Binet and Field and the Ehrenberg-Bass Institute have all scrutinised thousands of assets of varying quality in every media environment imaginable. The TLDR? Most ads lose 95% of their impact within a couple of weeks of running. The best performers (emotional, visually compelling TV commercials) can hold on for a month or two once the seven-figure media budget runs dry.
However, the widespread adoption of large language models (LLMs) has fundamentally changed this dynamic, extending the ROI window of earned media significantly.
In the olden days (pre mobile and social), the return window on a PR activation or comms spike was essentially a day. Your brand would drop the story or take over Martin Place, pray for good weather and hope that enough people read the papers or walked past your activation to make it count.
By the mid 2010s, the ROI window on the same initiatives extended to a few days, thanks to the prevalence of mobile phones and social media. Suddenly, if what you did was interesting enough, people would take selfies, share videos, post links, tag friends and add comments, keeping the brand in circulation for a little bit longer. Some of the best examples from this era include ‘A dramatic surprise on a quiet square’, Delite-O-Matic and of course, Volkswagen’s The Fun Theory.
But what about ROI today?
Welcome to the LLM era
While LLMs are no longer sci-fi, their integration into daily search and decision-making behaviour is where the real shift lies. Rather than relying solely on static training snapshots, modern models rely heavily on real-time web grounding and retrieval. They dynamically query live data to evaluate, synthesise, and surface authority signals for millions of user prompts every hour.
This is where the concept of “share of model” comes in. Pioneered by Jack Smyth and Tom Roach at Jellyfish as the AI-era successor to “share of voice”, share of model measures how often, how prominently and how favourably LLMs cite and recommend a brand relative to its competitors across thousands of category queries.
Analysing share of model data shines an interesting light on how these engines weigh external sources. As a general rule, the harder something is to fake, the more weight it carries in the model’s response engine.
Content on your own website that says your gluten-free cookies are awesome? Cool story.
Hundreds of reviews or conversations between individuals who have tried and enjoyed your gluten-free cookies? Pretty good.
An independent, widely respected media publication, definitively proclaiming your gluten-free cookies as the best in the world? Now we’re talking.
And here’s the catch.
Once ChatGPT or Gemini decide your cookies, washing machines or cruise experiences are undefeated based on trusted authority inputs, that recommendation persists until new, high-authority consensus proves otherwise.
That could be months, years or an eternity away.
The extended shelf life of earned media
The land grab for default status
So what’s the ROI on creating premium, earned media attention today, that drives visibility, recommendations and choices for years to come? It’s still tough to say, but also tough to imagine the conditions being any more favourable.
As we’ve witnessed with every major tech revolution that’s come before this one, fortune favours the first. The brands that move early cement their position as the default option, causing every late comer to play catch up.
So, how do you actually influence the machines? If you can’t buy ads inside ChatGPT’s reasoning engine, how do you manufacture an infinite half-life?
It requires a two-pronged approach to earned media: Culture spikes and press desk.
Culture spikes: Moments that command coverage
The first way is to develop the creative, high-impact activations we’ve always loved – think modern equivalents of the fun theory. But today, the objective of the stunt isn’t just immediate social impressions; it’s finding ways to strategically seed ideas in the quiet corners of the internet – think sub-reddits, niche forums, specialised Discord servers, or creator networks.
Ideas that resonate deeply in these communities often find ways of naturally bubbling up to the surface. In a few short hours, they can make their way onto mainstream social feeds, gain velocity and create a cultural gravity that mainstream journalists simply cannot ignore.
When those journalists pick it up and publish it on high-authority, trusted news sites, the mission is accomplished. You’ve successfully engineered a high-value data point that the LLMs will index and trust.
The share of model data is illuminating here too. Even a small dive into local analysis reveals around a dozen core Australian news properties, spanning broadcast, broadsheet, tabloid and trade. Of course, online articles from the usual suspects like ABC, 7news, and the Financial Review are in there, but so are many stories in the form of TV clips, social posts and even the odd podcast.
Press desk: Always-on validation and reviews
The second way is the structural, everyday work that traditional PR agencies have always done well. For a while, the standard ‘press desk’ function was starting to feel old-fashioned, buried under the weight of algorithmic social feeds and real-time ad bidding. But LLMs have completely reversed that decline. Because these models constantly scan authoritative news sources for consensus, the everyday press desk has suddenly become hot again.
If the AI is looking for signals to prove your brand is the right recommendation, you need to feed it a constant stream of them. This means keeping the brand in regular circulation via media-ready voices, original research and active community engagement.
Crucially, it also includes a rigorous strategy to drive customer reviews, creating a layer of decentralised, peer-to-peer verification that acts as hard-to-fake earned proof. When the LLM reads that collective, continuous vibe check across the web, an old-school PR capability transforms into definitive, modern category proof.
The new scorecard
The traditional marketing hamster wheel is exhausting. We spend millions on paid channels only to watch our impact decay rapidly the moment the budget stops.
But the LLM era offers an alternative. By adding a focus on premium, earned media and systematic review acquisition, we aren’t just buying temporary attention. We’re building a permanent corporate asset.
So the next time you drop a campaign, don’t just count the clicks or check Tuesday’s social feeds. Ask yourself what the machines will say about it next year. Because if your brand doesn’t exist in the model, to the modern consumer, it may not exist at all.