How to classify a keyword's search intent

The keyword is not the intent. The results page is.

Classifying search intent means naming the job a query is doing, then checking that name against what Google is already showing. If the first page is product URLs, a 2,000-word definition will not be “more complete.” It will be the wrong shape.

The results page is the classification

Search the query in a clean window. Write down the first ten organic results as types, not as brands: guide, documentation, product page, category, video, forum, local pack.

That list is the intent. A query that returns eight vendor homepages is commercial, even when the wording is “what is”. A query that returns docs and a troubleshooting thread is informational, even if it contains a product name.

Google describes the same idea from the snippet side. In the snippet documentation, snippets preview the page content that best matches that search. Google may show different snippets for different searches. One URL, several jobs. Your label has to survive that.

Do this for the query you will put in the title, not for a synonym you prefer. Then read how to check a meta description against the page before you write the sentence under that title.

Four labels, used as a working set

Google does not publish this taxonomy. It is still the smallest set that keeps a brief honest:

• Informational. The person wants an explanation or a procedure. Results are guides, docs, and references.
• Commercial. The person is comparing options. Results are roundups, versus pages, and product pages with proof.
• Transactional. The person is ready to act. Results are pricing, signup, product, and local inventory.
• Navigational. The person wants a specific site. Results are that brand, and little else.

Many queries are mixed. “Accessibility checker” in our own field pulls tools and explainers. Call the mix. Do not average it into a fifth label called “general.”

A second split matters more than the adjective: is the dominant result a page that answers, or a page that sells? Write the draft in that format. Adding a FAQ block does not turn a product page into a guide, and it does not turn a guide into the product.

The snippet changes when the query changes

Intent is not one tag stored on the keyword forever.

Google’s generative AI guide describes query fan-out: the model issues related searches to fill in the answer. Their example starts with “how to fix a lawn that’s full of weeds” and fans out to herbicides, non-chemical removal, and prevention. Three neighboring intents, one original query. A single label of “informational” hides the split.

That is also why a pile of near-duplicate posts fails. The same guide says that making a separate page for every query variation, in order to chase rankings or AI answers, violates Google’s scaled content abuse policy. One strong URL can match a query that does not repeat its heading word for word. If two of your posts would both answer the original query, that is cannibalization, not a cluster.

Re-check intent when the results change shape. A new ads block or a forum takeover is a new brief.

A model label stays a hypothesis

A classifier is useful when you have thousands of keywords and you want the same labels every time. It is a hypothesis until someone opens the results.

TypeSafe documents Jev as a model that answers typed questions about state you send. Choice returns the selected option, a probability for each option, and a confidence value. There is no browsing step in that design. If you do not put result-page evidence into the state, the label is a guess about the string.

TypeSafe says to keep questions atomic. “What is the intent?” is acceptable only if the options and the criteria are explicit, and if low confidence routes to a person. Combining “intent” with “should we write this?” in one question is how you hide a business decision inside a probability.

On 15 September 2026 TypeSafe published end-to-end response times of 70ms to 500ms for this class of call. Fast labels are still labels. Sample the ones near your confidence cutoff against live results before the label reaches a brief. The brief’s first line should name the result type you saw, not the enum value alone.