Ask a large language model what’s trending, who’s buying what and why, and it answers fluently. The unsettling part is you have no idea where that answer comes from, or whether any of it is true. Brilliant models, no foundation: that was the provocation behind the EntityX × Ipsos session at SXSW London, where founder Tom Quick and Ipsos’s Dan Wong-Chi-Man worked through what it takes to make AI useful for one of the domains it reasons about least reliably, human culture.
01 / The problem
Brilliant reasoning, on a foundation of nothing
Tom’s image on stage was the philosopher’s thought experiment: the brain in a vat. A language model is an extraordinary cognition engine, but it isn’t plugged into the world except through the tools and context you hand it. Ask it about culture and it tells you something, but never where that something came from.
“An LLM can tell you about culture if you ask it questions. But you don’t really know where that data is coming from, what it’s grounded in. And grounding LLMs in data that’s well understood is really critical.”
Tom Quick, EntityX
02 / The grounding
Start with a real panel. Then make it addressable.
This is where the partnership does its work. Dan set out the Ipsos side: not a scraped or inferred signal, but a real, consented panel.
“We start with a panel — 10,000 nationally representative people in the UK. They’ve opted in, it’s UK-based, and it’s UKOM-endorsed. You get the obvious demographics, but what’s really rich is when you start thinking about the topics and interests that are trending — and where people are on the purchasing funnel.”
Dan Wong-Chi-Man, Ipsos
That panel is the ground truth. EntityX builds the bridge from knowing these people exist and what they care about to reaching them in live media: the Cultural Engine reads the open web at scale, billions of addressable media events a day, classified against a knowledge graph of named entities, matching the panel’s known interests to what people are consuming right now.
03 / What falls out
Audiences nobody briefed for
The most interesting things this surfaces are the ones nobody set out to find. Two from the work, in completely different worlds.
Example one · Money
Tom described work with Ipsos and a financial-services client targeting a lower-income audience building their credit. Around the Budget, that audience engaged hard with material it wouldn’t normally touch, because this time the Budget mattered to them. Reading what else they consumed alongside it revealed two clusters at opposite ends of a spectrum:
- Loud budgeters. People on a restricted income who treat managing it not as embarrassment but as pride: financial discipline as identity.
- YOLO spenders. The near-nihilistic other end: “I’m never going to afford a mortgage anyway, so I might as well max the card and go on holiday.”
Neither group was in the brief, yet a loud budgeter and a YOLO spender want to be spoken to in completely different ways, even though a conventional demographic model files them under the same age and income.
“It’s a completely different trader treatment for the loud budgeters versus the YOLO spenders. And we see that sort of thing all the time.”
Tom Quick, EntityX
Example two · Wellness
Point the same method at food and wellness and a different cultural map resolves out of the co-occurring interest, into clusters just as distinct:
- Longevity optimisers. The West Coast biohacking world of bodily metrics, pharmaceutical optimisation and performance-as-a-project.
- Pilates princesses. The cluster around a certain athleisure and a particular vocabulary of diet and wellness.
Those two are deliberately obvious. The non-obvious one is the payoff: GLP-1, the weight-loss-drug story, shows up reshaping food culture, visible in the engine’s live signal while it was still emerging. A shift nobody briefed for.
04 / The line that matters
Signal, not surveillance
There is an ethical shape to this the talk didn’t shy away from. The more you can derive signal about what people are interested in, and deliver it into activation, from context rather than personal tracking, the better. No PII, no IP addresses, no cookies following anyone around the web.
“In the case of using panel data, it is grounded in some individuals — but fully consented individuals. The mechanism of enriching that and delivering it doesn’t involve any violation of privacy, because drawing on the power of context decouples it from the individuals.”
Tom Quick, EntityX
05 / What comes next
Getting the signal to the machines
If the first half of the talk was about grounding AI in culture, the close was the inverse: making that grounded signal available to the machines. Brands and holding companies are getting good at AI; the open question Tom and Dan left on the table is how you put well-structured, panel-grounded cultural signal in front of those systems, through MCP servers and APIs an agent can query the way a planner would. EntityX’s cultural intelligence is already there, a live read-only planning layer AI tools can reach today: not AI you trust blindly, but cultural intelligence you can feed to the AI you already have.
Plan from what people do, not what a model guessed.
EntityX reads the open web at scale and connects it to real, consented audiences, so the loud budgeters, the YOLO spenders and the clusters nobody briefed for become something you can actually reach. Cultural intelligence, made addressable, with no personal tracking.
Find out more →Disclaimer
Of course, large language models don’t literally run on a foundation of nothing. What they know is encoded, inscrutably, across billions of parameters and frozen at a training cutoff, which isn’t a solid grounding in what is happening in culture right now. And yes, the major models can search the web, but web search gets you web-search results, not bespoke segments for a brand’s audience inside high-quality panel data, nor real-time analysis of content consumption across addressable media fed back into those same media spaces. EntityX closes that gap by giving these models, which are brilliant at using external tools, access to current, verifiable signal about culture as it moves, and to the deterministic audience data only a panel can provide.