2026 UK edition
MediaCityUK, Salford · 10-11 September 2026
Twenty-two data and analytics leaders from British and European media organisations spent two days at MediaCityUK in Salford. Twelve talks ran to a strict format: fifteen minutes of presentation, then fifteen of discussion, under Chatham House Rules. Two speakers dropped out late and both Friday gaps were filled from the floor. Where the previous edition asked whether language models would change this work, this one took that as settled and argued instead about what comes after: who owns generated software, where corporate memory goes when everyone can produce analysis, and how anyone proves a probabilistic system still works.
Talks
- “Building a Demand-Tracking Product from Public Web Data” – An analytics consultancy reconstructed the demand curve for a seasonal live-entertainment operator from seven public web sources, built in roughly four weeks of spare hours and sold only once it worked. Almost nobody opens the dashboard that came with it: what gets used is a recurring email and a messaging thread.
- “Adventures with Conversational Analytics” – A news publisher built a routing agent over four data sources and found that ordinary data governance did most of the work, while the fashionable context layer was worth 10 to 20 per cent at the margin. Adoption reached 70 per cent of users because the same agent sits on four surfaces rather than one.
- “Democratising Software Development Across a Media Company” – A public service broadcaster opened a platform where any employee publishes a working tool by dragging a folder into a browser, and 17 apps landed in the first weeks with no promotion. Regional journalists built audio editors better than anything they could buy, and nobody could agree who maintains them.
- “Measuring Incremental Reach from Off-Platform Distribution” – A public broadcaster used copulas and tetrachoric correlations to turn daily, global social figures into weekly domestic reach. A festival week that an independence assumption put at 32 per cent came out at 13 per cent, and one correlation matrix taken from the top of the funnel still held two levels further down.
- “Designed to Advise: Building a Centralised Insight Function” – A former insight director at a commercial broadcaster described merging six scattered teams into business partners sitting over three specialist pools, after which the function became the only team besides the chief executive that thought about the business as a whole. The regret was building on data infrastructure that was not ready, having flagged it and pressed on anyway.
- “Informing Content Exploitation Decisions with Scientific Models” – A studio and streaming group spent four years modelling every licensing and scheduling decision, carrying more than 50,000 interpretable parameters for a single metric. Against a consultancy prototype it doubled accuracy on less data and cut running costs sharply, by connecting to data where it already sat rather than copying everything into one lake.
- “Making Ad Effectiveness Work” – Three advertising measurement veterans argued that the industry's favoured triangulation of models, experiments and attribution fails in practice, and that what separates teams who get it right is organisational rather than technical. One client's own credible evidence supported anything between spending nothing on television and spending seven figures a year.
- “Go Big, or Stay Home: The Economics of Stadium Concerts” – An economics consultancy built the dataset that did not exist: headlining artists up 16.5 per cent in a year, concerts up 12.4 per cent, and the number of stadiums flat. Italian artists play 93 per cent of their stadium shows at home, Korean artists 92 per cent of theirs abroad.
- “Our New Viewing Segmentation” – A public broadcaster cut a year of single-source panel data roughly 25 ways and kept getting the same four groups, defined by what people watch on rather than who they are or when they watch. Day of the week and time of day made almost no difference, and the finding the business used was that moving a show from Wednesday to Friday changes nothing.
- “Analysing a Century of Parliamentary Language on Immigration” – A newspaper's data science team classified the stance of five million parliamentary speeches across a hundred years, using a two-stage pipeline that judged relevance first and sentiment second. General sentiment models cannot do this, because they score the vocabulary rather than the position, and a passage quoting hostility in order to attack it reads as hostile.
- “Evaluations: Getting AI into Production” – A consumer data group shipping customer-facing generative features described the harness around the model, and the two slow qualitative stages that consume the effort: defining what success means, then gathering the evidence. Nobody in the room had a source for a golden question set.
- “Drivers of Frequency: Causal Inference Modelling” – A subscription publisher used a three-month propensity-weighted design to find which content format drives frequency, because withholding the thing people paid for is not a testable option. Every format came out the same size, because subscribers use several at once, so the ranked list the business wanted does not exist.
What we learned
Corporate Memory Went Backwards: Once anyone could produce analysis, almost nobody filed it. At one publisher more than 90 per cent of the knowledge produced used to reach the central repository; within three months of production federating, that fell to about 3 per cent. Output exploded and institutional knowledge shrank. The experiment library was the clearest casualty: far more experiments were being run, and not one result had been added since April.
One Question, Five Answers: Cost per acquisition on search advertising came back as five credible figures roughly ten times apart, depending on which method produced it. The recommended fix, multiplying models, experiments and attribution into a single picture, fails on plumbing rather than theory: channels get lumped together, time periods do not line up, and the calibration ends up as one number hard-coded forever. Asked to describe how they recalibrate a model from an experiment, six practitioners, several of whom had presented on doing exactly that, gave no credible answer.
Only Thirty Per Cent Is Contestable: Five services take 75 per cent of all television time and 11 take more than 90. The same concentration repeats inside each person's own viewing: a preferred service holds about half of it, and the top three take nearly 70 per cent. If you are not already in someone's top three, the remaining 30 per cent is the entire market you are competing for, and scheduling will not get you in.
Models Agreed With Humans More Than Humans Did: Twelve annotators labelling political speech for stance agreed at an alpha of 0.56, which is what a genuinely subjective task looks like. Three foundation models from different providers, prompted to describe a reader's reaction before assigning a label, agreed with the humans at 0.63 and with each other at 0.69. That turned 1,330 human labels into 22,000 training examples. The team would not cut the human share any further, because doing the labelling is how you find out whether the task is feasible at all.
Everybody Wants It, Nobody Will Own It: A broadcaster's self-service build platform worked, and staff shipped tools the market does not sell, including one that replaced an external creative service costing a five-figure sum per programme. Maintenance is the unsolved half. Developers who reviewed the generated code said it was fine but that no human could work on it, and the team that built the newsroom video editor declined to look after it. Nobody could say how you stop a prompted tool becoming load-bearing after its builder leaves.
Twenty-Two Years to the Top: The median gap between an artist's first release and their first stadium headline is 22 years. Britain has the oldest headliners, at a median of 26 years, against about ten for South Korea, whose production system starts artists at 14 and has them headlining within a decade. Stadium rent for a single night has tripled in two years, which is why multi-night runs are now standard, and why the middle of the live market is disappearing.
The argument has moved on from whether these tools work to who is accountable for what they produce, and how anyone would notice when they quietly stop working.