How do you extract contextual ad signals from a URL?

One pass. The web_scraper extractor renders the page in a real browser and separates the article body from navigation and related-story rails. A structured-output call against a fixed JSON schema returns all five signals together: iab_tier1, iab_tier2 and iab_path; entities as an array of name and salience ordered by salience descending; sentiment with a score; brand_safety with brand_safety_categories; and keywords. Text embeddings index the same article text so the corpus is also searchable by meaning.

How do you score entity salience?

Salience is a number from 0 to 1 stating how central an entity is to the article, and it sits in the response schema next to the entity name, so the model has to commit to it rather than leave ranking to the reader. On an article about a Marvel film, the film scores 0.95, the studio 0.75, the streaming service 0.65, the franchise 0.55 and the lead actress 0.45. The main entity is the top score, which makes it a property of the structure rather than a downstream guess.

How do you classify brand safety for ad placement?

The flag is graduated rather than binary: safe, low_risk or high_risk, returned with the sensitive-theme categories that produced it. That distinction is what separates lookalikes from real risk. An article about a musician who died of an overdose returns high_risk with death/injury and drugs as its categories, while a warm retrospective about an actress who died at 99 returns safe. Both are death-related, and a blocklist keyed on the theme would treat them identically.

Why is a keyword blocklist not enough for contextual advertising?

A blocklist matches strings, so it cannot rank and it cannot grade. It cannot tell you which of nine entities the page is actually about, and it blocks every article sharing a word with something unsafe. The cost is inventory: pages that were never risky get excluded, and the advertiser pays for the false positives without ever seeing them.

Does the taxonomy cover categories outside entertainment?

Yes. The classification is IAB Content Taxonomy 3.x, so it spans the full tree rather than a hand-picked subset. A recipe article returns Food and Drink at tier 1 with Cooking and Recipes at tier 2, alongside the same entity, sentiment and safety fields as any other page.

Intermediate

Advertising

IP Safety & Copyright

6 min read read

Contextual Page Signals for Ad Placement

How to extract contextual advertising signals from a publisher URL: IAB Content Taxonomy category, main entities each scored 0 to 1 for salience, article sentiment, a graduated brand-safety flag, and keywords. One pass, no bespoke pipeline.

Who It's For

Contextual advertising and ad-targeting teams who decide, per article, which creative renders beside it, and who need that decision to be defensible to a brand.

Problem Solved

A flat list of entities found on a page cannot tell you which one the page is about, and a theme label cannot tell a routine obituary from a story that would embarrass an advertiser. Both are buying decisions, and both need structure a keyword list does not carry.

Ready to implement?

Schedule a Demo View Documentation

See It in Action

A narrated 52-second walkthrough: the salience ranking on a Black Widow article, a brand-safety pair that splits on severity rather than theme, an IAB category outside entertainment, and the raw JSON payload. Click to play, since it has sound.

Before & After Mixpeek

Before

✗

Entities

A flat list, unordered. The main subject is a guess.

✗

Brand safety

One unsafe bucket keyed on theme, so lookalikes are blocked with the real thing.

✗

After

✓

Entities

Each carries a salience score from 0 to 1 and the list is ranked by it, so the main entity is the top score.

✓

Brand safety

safe, low_risk or high_risk, with the sensitive-theme categories that produced the verdict.

✓

Why Mixpeek

The salience score sits in the schema next to the entity name, so the model has to commit to a number for how central each entity is. That turns an entity list into a ranking you can target or avoid by dominant entity, and it is a property of the response structure rather than a judgement someone makes downstream.

Overview

Contextual ad platforms have to understand what a publisher page is about before placing an ad, both for relevance and for brand safety. This use case covers a contextual video-advertising platform that decides, per article, which video renders beside it. Point Mixpeek at any article URL and five structured signals come back: an IAB Content Taxonomy 3.x category at tier 1 and tier 2 with the full path, the main entities each carrying a salience score from 0 to 1 and ranked by it, article-level sentiment with a confidence, a brand-safety flag of safe, low_risk or high_risk with the sensitive-theme categories behind it, and keywords.

Challenges This Solves

A flat entity list hides the main entity

Take a local news page about an actress returning to a sitcom. Nine entities come back, and three of them are politicians who appear only because the show once feuded with one. Read flat, the page looks political.

Impact: An ad-ops team either targets the wrong context or blocks a page that was never political. Both are decisions made on a list that carries no ranking.

Theme is not severity

Two articles can share a surface topic and deserve opposite verdicts. A story about a musician who died of an overdose and a warm retrospective about a beloved actress who died at 99 are both death-related.

Impact: A keyword blocklist keyed on the theme blocks both, so the advertiser loses inventory that was never unsafe, and a system with only one unsafe bucket cannot express the difference.

Every signal in a separate pass costs a pipeline

Category, entities, sentiment and safety are usually four services with four latencies, four failure modes and four schemas to reconcile per URL.

Impact: The integration work outgrows the signal itself, and per-page cost scales with the number of services rather than with the number of pages.

Recipe Composition

This use case is composed of the following recipes, connected as a pipeline.

Brand Safety & Ad Verification Pipeline

GARM-compliant brand safety for ad creatives

Taxonomy Enrichment Pipeline

Classify content into custom or IAB taxonomies

Metadata Enrichment Pipeline

Transform raw content into structured, queryable data

Feature Extractors Used

web_scraper

text_embedding

Retriever Stages Used

semantic search

filter aggregate

Expected Outcomes

Five, from a single pass

Signals per URL

Read from the top salience score, not inferred downstream

Main entity

Three graduated levels plus the categories behind the verdict

Brand safety

IAB tier 1, tier 2 and the full path

Category depth

Build this in the docs

The exact stages and extractors this use case runs on, with API reference and worked examples.

Feature extractorsHow an extractor turns a source into typed fields Data modelWhere the per-URL documents live RetrieversQuerying the scored corpus once it exists

Score your own pages

Point a collection at your URLs and get the same five fields on every page: IAB category, ranked entities with salience, sentiment, graduated brand safety, and keywords.

Estimated setup: 30 minutes

Run this on your own data, free Book a demo Documentation