Advertisers care about where their ads appear. A food brand does not want its spot next to a story about a food recall. An airline does not want to sit beside a report on an air accident. At the same time, they want ads to land beside content that fits, such as kitchenware next to a cooking video.
Publishers and ad platforms use AI to make these placement decisions at scale. The model is only as good as the labels behind it. Brand safety labeling and contextual ad labeling are the work of describing content so that placement decisions can be made correctly. This post explains what that labeling involves and why it is harder than it looks.
Moderation asks whether content should be allowed on a platform at all. Brand safety asks a different question: would an advertiser want to appear next to it?
Plenty of content passes moderation and is still a poor place for some ads. A serious news report about a tragedy is legitimate and valuable, and many brands would still choose to stay away from it. For this reason, brand safety needs its own labels, designed around advertiser needs and not platform rules.
The most important design choice is to separate two things: what the content is, and whether a given brand finds it suitable.
Content labels describe the material: its topic, its tone, whether it touches sensitive themes, and how it treats them. Brand rules then use those labels. One brand may accept news about a sensitive event, while another may not. A children's brand and a financial services brand will draw different lines.
If annotators label directly for one brand's preferences, the data cannot serve anyone else, and it must be redone whenever those preferences change. Labeling the content once and applying brand rules afterward keeps the data reusable.
Many systems started with keyword lists. Block the page if it contains certain words. This is simple, and it fails in two ways.
It blocks too much. A photography blog that mentions "shooting," a finance article that says "crash," and a recipe that says to "kill the heat" all contain words that look risky and mean nothing of the sort.
It also misses things. A page can discuss a sensitive subject without using any flagged word.
For this reason, ad suitability data needs labels based on meaning, not on word matches. Annotators must read, watch, or listen to the content and decide what it is actually about.
Two kinds of labels carry most of the weight.
Topic labels place content in a category: sports, cooking, finance, health, politics, travel, and so on. A defined taxonomy keeps these consistent.
Tone labels capture how the content feels: positive, neutral, serious, tragic, humorous, or satirical. Tone changes suitability. A light story and a grim one on the same topic may be treated very differently by an advertiser.
Treatment matters as well. A report on a difficult event, an opinion piece about it, and a comedy sketch that references it are three separate cases, and the labels should show which one it is.
Ads do not always sit beside a whole article or a whole video. In video, they often appear at breaks within the content. A video may open calmly and turn tense near the end, so a label for the whole file would hide where the problem is.
Useful labeling therefore works at several levels: the page or video as a whole, and the segments inside it where ads may run. Annotators need to know which level a label belongs to, and the guidelines should say where segment boundaries fall.
Safety is only half the picture. Contextual advertising also aims to match ads to content that suits them. A cooking video fits kitchenware. A travel guide fits luggage and airlines.
Relevance labels describe the themes and interests in the content, such as recipes, hiking, home repair, or personal finance. These are separate from safety labels. A page can be perfectly safe and a poor match, or a good match and unsafe. Keeping the two label sets apart lets the system weigh them independently.
Relevance can also create risk. An airline ad is highly relevant beside a travel story and badly placed beside a story about a crash. The model needs both labels to tell the difference.
Meaning shifts across languages and cultures. A joke, a phrase, or a symbol can be harmless in one place and offensive in another. Labels for global content should be written, or at least reviewed, by people who know the language and the region.
Events move fast too. A page that was neutral last week can become sensitive after a news event. Labeling rules need a way to respond quickly, and the data needs a process for refreshing labels on content that has gained new context.
People with experience in advertising, media planning, or brand suitability read content differently. They know how advertisers think, can judge tone and treatment, and understand why one page is acceptable to one brand and not to another. They also help write the rules that keep a large team consistent: the topic taxonomy, the tone scale, and the handling of unclear cases.
This matches Globik's approach of using verified domain experts for work that depends on judgment. For contextual advertising AI training data, that expertise is what separates labels advertisers can trust from labels they must second-guess.
Before a brand safety or contextual labeling project starts, a team can settle five things:
Brand safety and contextual advertising AI works when labels describe content on its own terms, capture topic, tone, and treatment, and leave brand preferences to a separate layer. For another example of tone and mood labeling, see our guide on music and audio tagging: labeling genre, mood, and instruments consistently. In both cases, strong brand safety labeling depends on people who understand how the content will be used.
It is the labeling of content so AI can judge whether it is a suitable place for an advertiser's ad.
Moderation decides what is allowed on a platform, while brand safety decides what advertisers want to appear beside.
It blocks harmless pages that contain risky-looking words and misses sensitive pages that use none of them.
Brands differ and change their rules, so content labels stay reusable while rules are applied on top.
They understand tone, treatment, and how advertisers think, which keeps labels consistent and trustworthy.
