A context hint is a short clause describing a user's intent or conversational moment — not a noun phrase about the topic. Operators who write good context hints describe what someone is trying to do right now; operators who write them like keywords are still in the keyword era and will get keyword-era results.
A context hint is one short clause naming a USER INTENT or CONVERSATIONAL CONTEXT in which an ad would be appropriate. It is read by the relevance function as a semantic description of the moment — not as a match term.
The format rule:
Two ways to test whether your hint follows the rule:
| Wrong (keyword shape) | Right (context-hint shape) |
|---|---|
| "contextual ads" | "user is comparing ad-tech vendors for privacy posture" |
| "agent advertising" | "developer is evaluating how to monetize an agent surface without behavioral tracking" |
| "brand safety" | "brand team is sourcing placement rules for AI-generated answer surfaces" |
| "ergonomic chair" | "remote worker is comparing chair options for a small home office under a stated budget" |
| "FCS protocol" | "publisher is evaluating which provenance headers to emit on AI-generated responses" |
The right column is longer because it carries information the left column compressed out. That information is what the relevance function uses; without it, the protocol is doing keyword matching with a different label.
Search-era ad targeting could survive on keywords because the user typed the keyword. The system saw the compressed signal directly and bid against it.
Agent surfaces do not produce keywords. They produce conversations — multi-turn, contextual, with constraints and goals named in full sentences. A keyword-shaped context hint forces the relevance function to throw away the surface's actual signal and re-derive a keyword from it. Reduced to keyword matching, the system loses precisely the information that made the agent surface valuable.
A clause-shaped context hint preserves the signal. The relevance function reads the moment, not a re-compression of it.
Practical mechanics, in priority order:
Per the operating guideline, context hints belong in one of four categories:
| Category | What it covers |
|---|---|
evaluation | Hints about a moment when someone is comparing, assessing, or deciding between options |
implementation | Hints about a moment when someone is integrating, configuring, or building |
governance | Hints about a moment when someone is checking, verifying, or auditing |
monetization | Hints about a moment when someone is sourcing inventory or assessing revenue |
Each hint goes into exactly one category. The taxonomy keeps inventories portable across ad groups and prevents over-broad hints that map to too many situations to be useful.
The relevance function (described on /trust-safe-relevance) takes context hints as one of three public inputs — alongside sponsor declarations and content provenance. The function reads the hint as a description of the moment, computes overlap against the sponsor's intent descriptor, and returns a relevance score.
Three observations follow from how the function reads hints:
The mechanics of the function are on the sister architectural pages. This page is just about writing the input.
OpenAI Ads publicly frames ChatGPT advertising around conversation, context, richer context signals, and real-time decisions rather than keyword bidding. The context-hint shape mirrors that public direction: clauses describing the conversation moment, not keyword lists.
Writing hints well is a portable operator skill. An operator who learns to describe moments rather than topics is better aligned with OpenAI Ads' public framing, ABF-aware runtimes, and any system that adopts the context-hint shape.