What Is Few-Shot Prompting?

Few-shot prompting is the technique of including a handful of worked examples inside a prompt so the language model infers the task's pattern and applies it to new input. The term comes from the 2020 GPT-3 paper, "Language Models are Few-Shot Learners" (Brown et al.), which demonstrated that large models could pick up tasks from examples alone — no retraining required.

How it works in practice

A few-shot prompt shows the model input-output pairs before presenting the real input: two or three examples of a customer email and its correct category, then the email to classify. The examples do double duty — they communicate edge-case handling that instructions alone describe poorly, and they lock the output format so responses parse reliably. Three to five well-chosen examples typically capture most of the gain; poorly chosen ones actively teach the wrong pattern.

The GEO application: measurement rubrics

AI visibility tracking runs on machine scoring. A tracker that samples hundreds of engine answers per week cannot hand-review them, so an evaluator model scores each answer: Was the brand mentioned? Ranked where in the list? Described with what sentiment? Recommended or merely named? Few-shot examples are what make that evaluator consistent — a rubric anchored by scored specimens ("this answer is a recommendation; this one is a neutral mention; this borderline case counts as absent") produces stable data where an instruction-only rubric drifts.

The stakes are concrete: if the judge's threshold for "mentioned" wobbles between runs, your visibility trendline records rubric noise instead of market movement. Anchored scoring is a precondition for the week-over-week comparisons that feed visibility reporting.

Boundaries of the technique

Few-shot prompting steers behavior within a conversation; it does not add knowledge or persist. It also consumes context-window space, and modern instruction-tuned models need fewer examples than their predecessors for simple tasks. Its enduring niche is exactly where GEO measurement lives: subjective-but-repeatable judgments that must be made identically thousands of times. Related machinery — LLM-as-judge, eval harness — is defined in the glossary.