Why human judgement is finally transforming brand safety systems - for publishers and advertisers

Published: 05 Oct 2026
Author: Peter Mason

It’s been more than a decade since any meaningful progress was made to brand safety practice in programmatic advertising, and the lack of innovation is currently failing both sides of the deal. A single word appearing anywhere on a page - "crash," "attack," "crisis" - can still pull a URL out of the bid stream entirely, whether it describes a stock market dip, a sports event, or an actual tragedy. Publishers are losing monetisable inventory to false matches and advertisers are losing reach into high-value inventory. No one is currently winning.

But transforming brand safety doesn’t mean getting better at building keyword lists - to us, it means shifting from static rules to systems that learn from real human judgment.

At illuma, we’ve long taken a contextual approach to brand safety, analysing the signals within a page rather than relying on individual keywords or static lists. Our technology also goes further than simply understanding each page in isolation; it builds a layer of understanding around what related pages have in common, while preserving the nuances that make each piece of content distinct.

In particular, to apply human judgement to brand safety, we have had a system in place that surfaces examples for review by a member of our team, or by publishers themselves. Every time a human reviewer corrects a content classification - adding a label that was missed, or removing one that was wrong - that correction becomes a valuable signal, which is used to continuously train our model. To extract even greater value from those signals, we’ve now started grouping similarly corrected pages into tight cohorts and associating a label that captures common features of that content, so that patterns in human correction can be identified and generalised, to avoid one-off judgment calls.

Within each cohort, labels are only kept as valid descriptors when they have strong, consistent majority support, and the same label can carry a different meaning depending on the cohort it comes from. For example, a cohort of football match reports and a cohort of World Cup coverage might both generate a “Sports” label, but each carries its own shade of meaning drawn from its own content. The result is a taxonomy that captures nuance and keeps improving as more corrections come in, rather than requiring someone to manually maintain a list.

What makes this possible is that the system isn't just matching words - it's reading content the way a person does. It picks up on tone, sentiment, and emotional register, not just topic, and can tell the difference between, say, an article reporting on an industry trend and a news report about a disaster. That depth of understanding is what makes the cohort-based approach work in the first place - it's not enough to group pages together, the system has to genuinely grasp what each page is saying, in context, before a human correction can be generalised into something reliable.

When it works, that shift changes the economics for both sides of the trade. For publishers, inventory that would have been auto-excluded under a blunt rule gets evaluated on what it actually contains, so quality journalism isn't penalised for covering difficult but legitimate topics.

For the buy side, it means media plans stop quietly shrinking around false positives, and campaigns can reach relevant, brand-safe content that a keyword filter would have hidden from them entirely. Overblocking has always been framed as a publisher revenue problem, but it's just as much a reach and efficiency problem for advertisers too - every wrongly excluded page is also a missed impression against an audience the campaign wanted.

There's a trust dimension that matters to both, too. Agencies and advertisers increasingly want to know why a piece of inventory was included or excluded, and "it matched a word on a list" no longer holds up under scrutiny. A classification system built on reviewed, traceable human judgment can be explained, audited, and improved upon, rather than being hidden inside a black box.

To us, this is what transformation actually looks like in brand safety - not a marginally better blocklist, but a move away from static, one-size-fits-all rules toward a system that understands content the way a human reviewer does, at scale, while continuously improving. As a result, we’re already starting to see publishers monetising content that was always safe to sell, and buyers finally reaching content that was always safe to buy. The same change in judgement, transforming value for both sides of the trade.