Seven years of BrightonSEO agendas: what the industry stopped programming

BrightonSEO has published its agenda 15 times since 2019. Across 1,828 talks that is a record of what the SEO industry thought was worth an hour of a practitioner’s time, written down twice a year by people with no reason to flatter anyone’s theory of the field. Counted by subject, the agenda rotated: talks naming AI or an LLM went from 3.0% to 31.5% of the programme, while technical SEO gave up 4.6 points.

That is the useful shape of it. The room did not get bigger; talks about AI took slots that used to go to content, technical SEO and link building. Nothing was retired as solved. Something else simply started winning the pitch.

Of 20 themes tracked, 5 gained at least 3 points and 4 lost at least that much. The rest held roughly steady, which is its own finding: most of the agenda is stable, and the churn is concentrated in a handful of subjects.

The turn has a date, and it is one programming cycle

AI search visibility, talks about being found inside an answer engine rather than a results page, is the cleanest signal on the agenda, because it starts at nothing. It is absent from the first 6 reportable events, through Apr 2022. Then it appears, and the largest single-event move is 9.0 points, between Apr 2025 and Oct 2025. By Oct 2026 it is 14.7% of the programme.

A conference agenda lags reality with a known delay: talks are pitched, selected and scheduled months before anyone stands up. A step that size means a programme committee’s read of the field changed between two cycles rather than gradually. If you want a date for when optimising for answer engines stopped being a fringe topic, Oct 2025 is the one this dataset supports.

How a talk gets counted, so you can disagree

Each title is matched against 20 themes by word-boundary keyword patterns. The patterns are the method, so they are published in the repository rather than described here in prose: a reader who disagrees with a boundary should be able to read the exact pattern that drew it.

Three consequences worth stating plainly. Themes are not mutually exclusive, a talk called “Technical SEO for AI search” counts in both, so the columns do not sum to the agenda. A title matching no theme is counted as unmatched and reported, so the classifier cannot quietly hide its blind spots. And the unit is the title, never the talk: see the limits below, because that is the constraint that decides what this page can honestly claim.

A second corpus, pointing the same way

An agenda is somebody’s forecast of what an audience wants. That is a real limitation, so it is worth checking against a record of what the industry actually talked about, and this show has one: every episode since 2009, transcribed and counted year by year in The 18-Year Shift.

The two corpora share no text and no classifier. One is conference titles chosen by a programme committee; the other is spoken transcript from interviews and monologues. They agree on the inflection: AI topics climbed from 3.0% of this agenda across Sep 2019, Apr 2020 to 31.5% across Oct 2025, Oct 2026, on the same timeline over which the podcast corpus goes from no AI mentions to AI in most episodes. Neither measurement can corroborate itself. Together they are harder to dismiss than either alone.

What a rotating agenda is telling you

The subjects that lost programme share, technical SEO, content, links and digital PR, did not lose it because they were solved. No conference stops programming a topic that still decides whether a site works. They lost it because a finite number of slots got allocated to something newer, and a programme committee has to choose.

So the read is not “stop doing technical SEO.” It is that the attention market moved faster than the work did, and the gap between those two is where an advantage sits. If fewer people are being taught the craft that still decides outcomes, knowing it is worth more than it was in 2019, not less.

What genuinely arrived, on a datable step, is a second surface to be found on. That is an addition to the job rather than a replacement for it, and this agenda has had 15 chances to show the older work being retired. It has not.

A finding this page used to make, and no longer does

An earlier version of this page split AI talks into two senses, AI as a tool for doing SEO work, and AI as an interface to be optimised for, and reported that the first was flat for seven years while the second stepped up sharply. That split has been withdrawn.

Two reasons, both worth more than the finding was. The classifier that produced those numbers could not be found: not in the repository, not in its history, not in the offline archive of the project. Nothing could refresh or re-check them. A replacement was then written from scratch and hand-validated against three random samples of talk titles, 420 in total, labelled before the code was run on them. On the third sample, the one it had never been tuned against, it scored an F1 of 0.64 on the tool sense and 0.53 on the interface sense. The near-perfect scores on the first two samples were the classifier agreeing with the data that shaped it.

The obstacle is not the code, which is why more of it would not have helped. Archive-era agendas publish no abstracts, so intent has to be read from a title written to fill a room. “Automate to elevate: making space for human brilliance” is either an AI talk or a time-management talk, and the title does not say. Labelling the same samples three times, the same reader had to correct nine of their own earlier judgements. A distinction that cannot be applied consistently by a human cannot be reported to one decimal place by a machine.

The themes on this page ask a question titles can actually answer: which subject does this title name. That is a weaker claim than the one that was withdrawn, and it is one the data supports.

What this data cannot tell you

Published in full rather than footnoted, because a number that turns out to be overstated and undisclosed costs more than the citation was worth.

• The unmatched rate falls over time, so every rising share is marginally overstated (36.6% of talks matched no theme across the first 3 events, 18.2% across the last 3). Titles have got more explicit about their subject. The direction of each change survives that; the exact percentage is soft.
• Titles only, on both eras. 8 events were read from archived pages that carry abstracts; 7 from live listings that do not. Abstracts cut the unmatched rate sharply, so using them on one era and not the other would compare two different instruments. Everything here is title-only on both sides, which also means a talk whose title hides its subject is invisible to this count.
• Titles are marketing. Speakers name talks to fill rooms. “AI” in a 2026 title may be doing promotional work the content does not back, which would inflate the AI lines specifically.
• Events under 20 talks are excluded as non-reportable. A share computed on a handful of talks reads like a data point and is noise.
• One conference. This is BrightonSEO, not the industry. It is the largest SEO conference in Europe and it programmes twice a year, which is what makes a seven-year series possible at all, but a single programme committee’s taste is inside every number.
• Share is not volume. Every figure is a percentage of that event’s agenda. A theme can lose share while holding its talk count, simply because the agenda grew.

The data

1,828 talks · 15 reportable events · Sep 2019 to Oct 2026. Every talk counted here comes from BrightonSEO’s own published agendas, which anyone can go and read.

Citing a number from this page? Cite it as: The Best SEO Podcast, “The Conference Agenda Topic Shift,” October 2026, https://www.bestseopodcast.com/conference-agenda-topic-shift.