No credit card needed ✦ 14-day trial on all plans
AI Platforms
Learn what ChatGPT is, how it generates responses, and why it has become one of the most widely used AI assistants.
AI Platforms
Explore Claude, Anthropic's AI assistant, and learn how it delivers conversational responses using large language models.
AI Platforms
Learn about Gemini, Google's AI platform for conversational search, content generation, and AI-powered assistance.
Prompt sensitivity is the tendency of AI systems to produce different answers or citations based on small wording changes in user prompts.
Prompt sensitivity is how much an AI model's answer changes when you make a small change to the wording, structure, or emphasis of the prompt.
If you remember one thing: prompt sensitivity tells you how much an AI answer can change when the question changes slightly; it does not tell you which answer is correct.
Prompt sensitivity occurs because AI models interpret language probabilistically and are influenced by subtle cues in the input:
Prompt parsing. AI analyzes the exact wording, order, and phrasing of the input.
Context interpretation. Small changes can shift perceived intent or emphasis.
Token influence. Each word or token can significantly affect the probability of the next token.
Output variation. The model generates different responses based on these small differences.
Consistency testing. Repeated tests with slight variations reveal the level of sensitivity.
You'll see this concept under several closely related names.
LLM prompt sensitivity
Query coverage: how broadly a set of prompts represents the questions users may ask.
Ranking stability: how consistently search rankings hold over time or across measurements.
Prompt sensitivity
Query coverage
Prompt sensitivity
Query coverage
Prompt sensitivity
Query coverage
Prompt sensitivity
Query coverage
Prompt sensitivity
Query coverage
Prompt sensitivity
Query coverage
Prompt sensitivity
Query coverage
You cannot eliminate prompt sensitivity from an AI model. You can, however, make your content and testing less dependent on one exact wording of a question.
Your content answers only one version of the question. Cover the underlying topic clearly and address common ways users may phrase the same intent.
Your terminology changes from page to page. Use consistent names for your brand, products, topics, and key concepts.
Important information is difficult to extract. Use direct answers, descriptive headings, logical structure, and clearly separated concepts.
You only test one prompt. Build a set of closely related prompt variants and compare how consistently the AI responds to them.
No, prompt sensitivity means that the output can change when the prompt changes. A different answer does not by itself show that either answer is unreliable or incorrect.
Small changes can alter how a model interprets the user's intent, which context it considers most relevant, and how it generates the response. The exact causes and degree of sensitivity vary by LLM model and platform.
It can, if a small prompt change produces a different interpretation or retrieval path, the sources cited in the answer may also change. A citation difference is one observable output difference, not a complete measure of prompt sensitivity.
No universal sensitivity level should be assumed across large language models. Different models and platforms can produce different outputs for the same set of prompt variants, so sensitivity should be tested rather than inferred.