BrightonSEO is on 8 and 9 October at the Brighton Centre. It’s the tenth year the conference has been in that building. There’s a lot of AI on the agenda, and most of what gets said about it won’t be checkable afterwards.
That isn’t a dig at the speakers. It’s how conference talks work. You get twenty minutes, a deck, a chart with no axis labels and a confident line about what AI now does. People leave with a takeaway and nothing they could test.
I’ve been doing SEO for about 12 years and I’ve sat through enough of these decks. So this year I’m only writing down claims that come with a sample and a date.
The sessions I’ll be watching
Jean Sarunporn, “Getting Your Brand Recommended by AI Agents: A Practical Guide to GEO” (Thursday). GEO here means getting your brand into AI-generated answers. It becomes checkable if it names the agents, the prompts and the date they were run. Without those, it’s a list of tips.
Paris Childress, “Information Gain: The Missing Metric in Every GEO Content Strategy?” (Thursday). A metric is only useful if someone else can calculate it. I want the formula, and a before and after on pages that were actually changed.
Chris Donnelly, “What a billion sources teach us about getting mentioned & cited in AI” (Thursday). A big number is a decent start. The questions are which billion, collected when, and what counts as cited. A link and a passing mention are different results.
Gintarė Rimolaitytė, “Who AI Cites Now: How Source Patterns Shifted Since December 2025” (Thursday). The title already has a date in it. If the same queries were run in December and again now, this is the talk most likely to stand up.
Tom Rudnai, “Your AI Visibility Metrics are Lying to You: How Conversational Context Impacts AI Responses” (Friday). Fair point. The test is whether it shows the same prompt giving different answers under controlled conditions, and publishes the prompts so anyone can rerun them.
Malte Landwehr, “ChatGPT Shopping: Lessons from analysing 1M+ products in 250M+ chats” (Friday). There’s a sample size in the title, which already puts it ahead of most of the agenda. I’ll want to know where the chats came from and what dates they cover.
Ross Simmonds, “How to Engineer Your Brand Into the Memory Layer of Every LLM” (Friday, closing keynote). This is the hardest one on the list to check. Nobody outside the AI labs can see what a model has stored. The nearest thing to evidence would be a fixed set of prompts run against named model versions, before and after. If the talk has that, I’ll be impressed. If it doesn’t, it’s a keynote.
MeasureFest
MeasureFest, the measurement track, runs alongside the main conference, and it’s where I’d expect to find claims that can actually fail. Measurement talks have to show their working or they come apart in the questions afterwards. Tom Capper’s “How not to measure AI” (Thursday) is the obvious one for this subject. Kumar Amrendra’s “Money In. Money Out. The Rest Is Measurement Theatre.” (Thursday) has the most honest title on the agenda. I don’t need to agree with either talk. I need them to show the numbers.
What I’m listening for
The first thing is a sample: which queries or sites, and how many of them.
The second is a date. I mean the date the data was collected, which is often months before the deck was finished.
The third is a definition of cited. An answer can link to a source, name it without a link, or do neither. A lot of AI visibility charts count all three as the same thing, and they aren’t.
The fourth is whether anyone separates crawlers that collect training data from crawlers that fetch pages for answers. Blocking one doesn’t block the other. I looked at how the UK’s biggest news sites handle this in my study of their robots.txt files, and I’d like to see the same split in other people’s data.
After the event
I’m not writing a recap. If a speaker publishes a query list, a crawler setup or a raw data file, I’ll check it against what was said on stage and write that up. If nobody does, there won’t be anything to write.