Search Intent Classification

Search intent classification is the process of assigning an intent label, such as informational, navigational, commercial or transactional, to each query or prompt in a dataset. Where search intent is the concept, classification is the operational step: deciding which taxonomy to use, labeling thousands of queries consistently, and using those labels to segment reporting and plan content.

Search engines classify intent internally on every query to decide what to show. Google does not publish its classifiers, but its rater guidelines describe the categories human raters use to evaluate results (Know, Do, Website, Visit-in-Person). Marketing teams classify intent for a different reason: to make a keyword list or prompt list analyzable. A visibility number averaged across a whole prompt set hides the fact that a brand may be strong on informational questions and absent on buying questions.

There are three common ways to classify:

• SERP-based labeling: look at what the results page contains. Shopping units and price-led ads point to transactional intent, a single brand domain at the top points to navigational, and listicles or review sites point to commercial. Many SEO tools assign intent this way.
• Rules and modifiers: words like 'buy', 'price' and 'coupon' map to transactional; 'best', 'vs', 'alternatives' and 'review' to commercial; 'how', 'what' and 'why' to informational; brand names plus 'login' to navigational. Fast and transparent, but weak on long prompts.
• Model-based labeling: a trained classifier or a large language model reads the full query and returns a label, often with a confidence score. This handles long conversational prompts far better than keyword rules, which makes it the practical choice for AI prompt sets.

AI prompts make classification harder in specific ways. They are long (Google said in May 2026 that the average AI Mode query is three times the length of a traditional search query), often mix goals in one message, and can include tasks with no search equivalent, such as asking the model to draft an email. OpenAI's own research taxonomy for ChatGPT uses Asking, Doing and Expressing instead of the SEO categories, a reminder that the label set should fit the channel. Many GEO teams also add a branded label, because prompts that name your brand behave very differently from unbranded ones. Promptwatch, for example, tags each tracked prompt with one of five intent types (informational, branded, navigational, commercial, transactional) and classifies imported prompts automatically when no label is supplied.

Good classification practice is mostly about consistency:

Pick one taxonomy and document edge cases, such as whether 'Shopify vs WooCommerce pricing' counts as commercial or transactional. Allow a single primary label per prompt for reporting, and record a secondary label only where it changes the content you would create. Spot-check model-assigned labels on a sample of 50 to 100 prompts, since a wrong label quietly distorts every chart filtered by intent. Re-check labels when engines change: query fan-out data shows how an AI engine actually decomposed a prompt, which sometimes reveals a commercial reading of a question you labeled informational.

Once labels are reliable, they drive decisions. Compare AI share of voice by intent, map each intent to a content format, and prioritize gaps where commercial prompts show competitors but not you. Related: latent intent, prompt research and persona-based prompt tracking.