How Search Intent Is Evolving With Conversational AI Assistants

The labels most SEO teams use are older than that shift. Informational, navigational and transactional come from a 2002 paper based on work at AltaVista, when a search box returned a list of links. Those labels still help, but they were built to sort single queries, and assistants now handle whole conversations.

This post covers the history and the data: where the classic search intent types came from, what dated usage studies show about how people talk to assistants, and what changed for content teams. To turn it into a content plan, read our guide to AI search intent and the four conversational buyer stages. For a plain definition of the term, see our glossary entry on search intent.

Where the Classic Search Intent Types Came From

In 2002, Andrei Broder published "A taxonomy of web search", based mostly on work at AltaVista. Classic information retrieval assumed every search was an "information need." Broder argued that on the web, the need behind a query is often something else.

He proposed three types. A navigational query wants to reach a particular site. An informational query wants to learn something. A transactional query wants to do something on a site: shop, download a file or find a map. He measured the mix two ways, with a pop-up survey of AltaVista users and a hand-labeled sample of 400 queries from the daily log.

Two lines in the paper read differently in 2026. Broder wrote that "transactional queries are satisfied only indirectly" by search engines of the time. He also described a "third generation" of engines, then emerging, that tried to "blend data from multiple sources" to answer "the need behind the query." Assistants that answer, compare and book are one version of what he described.

Google's version: Know, Do, Website, Visit-in-person

Google teaches its human raters a parallel set of intents. The Search Quality Rater Guidelines (edition dated 11 September 2025) say it helps to think of queries as having one or more of four intents: Know (including Know Simple), Do, Website and Visit-in-person. Our post on how AI search interprets user intent walks through each one with Google's examples.

Most SEO tools use a fourth type, commercial, to separate comparison research from the purchase itself. Our search intent glossary entry covers those four types and how to spot them.

The three assumptions baked into the old model

Every version of the taxonomy assumed the same three things:

1. One query carries one goal. A query gets one label.
2. The query is short. Intent is read from a few words and a modifier such as "best" or "buy."
3. The engine returns places to go. The searcher does the reading, comparing and acting.

Conversational assistants break all three. The rest of this post shows the evidence, in order.

A Timeline of Search Intent, 2002 to 2026

Each step below changed what an engine could understand, or what people felt they could ask.

Sources: Google's ranking systems guide and its post on how AI powers search; OpenAI's ChatGPT launch post; Microsoft's February 2023 Bing announcement; Google's AI Overviews launch; OpenAI's ChatGPT search launch; Google's AI Mode Labs post and US rollout post.

What Usage Studies Show About How People Talk to Assistants

For two decades, research on search intent leaned on search engine logs. Now the assistant makers publish their own usage data. Each study below uses its own taxonomy and its own product, so read them side by side rather than adding them up.

People ask for help deciding, not just for facts

An OpenAI-led team of economists studied about 1.1 million sampled ChatGPT conversations from May 2024 to June 2025. In the NBER working paper "How People Use ChatGPT" (September 2025), they built a new intent scale. "Asking" means seeking information or advice "that will help the user be better informed or make better decisions." "Doing" means asking for an output, such as an email or code. "Expressing" means sharing views or feelings.

About 49% of messages were Asking, 40% Doing and 11% Expressing. Asking and Expressing both grew faster than Doing: by late June 2025, the split was 51.6%, 34.6% and 13.8%. The topic "Seeking Information" rose from 14% to 24% of all use in a year, and the authors describe it as "a very close substitute for web search." Broder's informational type is still there, but much of it now serves a decision.

Assistants get the harder tasks

Microsoft researchers compared Bing Chat with Bing Search using 80,000 conversations sampled from May to July 2023. In their study, 72.9% of chat conversations were knowledge work, against 37% of search sessions. Using a standard scale of cognitive complexity, 37.0% of chat conversations fell in the higher levels (apply, analyze, evaluate, create), against 13.4% of search sessions.

People didn't just move their old searches into chat. They brought harder work with them.

Questions get longer and less like keywords

Google's one-year AI Mode report (May 2026) says the average AI Mode search is triple the length of a traditional query. Its longer PDF version adds that people are "no longer worrying about the 'right way'" to phrase a question. The most common first words in AI Mode queries are What, How, I, Is and Can.

Third-party clickstream data agrees. Semrush's study of US ChatGPT activity from October 2024 to February 2026 could not match 65% to 85% of prompts to any keyword in its database for most of that period. That's a problem for any research process that starts and ends in a keyword tool.

Conversations run longer

Follow-ups are how intent gets refined now. Google says follow-up queries in AI Mode grew by more than 40% a month on average in the US. In ChatGPT, Semrush saw the average session hold at 1.16 to 1.21 prompts through most of 2025, then climb to 1.75 by February 2026. Most chats are still short, but the trend is toward more turns, and each turn can carry a new intent.

Personal and advice-seeking use is rising

Microsoft's Copilot Usage Report 2025 classified 37.5 million conversations from January to September 2025. Searching was the top intent, with advice second and rising. Health was the top topic on phones at every hour, while work topics led on desktops during office hours. The same need looks different by device.

Anthropic sees a similar spread on Claude. Its March 2026 Economic Index report found personal use rose from 35% to 42% of Claude.ai conversations between November 2025 and February 2026. The report ties the change to more personal queries about sports, "product comparisons" and home maintenance. Its June 2026 report adds a weekly rhythm: personal use runs around 35% of conversations on weekdays and just under 50% at weekends.

Read the studies with care

Each company measures its own product with its own automated classifiers. OpenAI's "Asking" and Microsoft's "Searching" are different labels, and Google reports growth rates, not counts. The direction is consistent across all of them. The exact percentages don't transfer from one product to another.

The Intent Shift Ledger: What Changed for Content Teams

This ledger is Rankbox's summary of the studies above. Each row pairs an assumption from the old model with what the 2024 to 2026 data shows, and the change it calls for on your pages.

The fourth row gets a full treatment in our buyer stages guide, which verifies the checkout and booking features vendor by vendor. The fifth row connects to context signals such as memory and location, covered in how AI search uses user intent and context.

A quick example of the ledger in use

Take a made-up invoicing app, Tallyfold, and one keyword from its old plan: "invoice reminder software," tagged commercial. The ledger turns that single row into three questions a buyer might actually ask an assistant, in order:

1. "Clients keep paying us late. What can a small agency do?" (asking, to decide on an approach)
2. "Do automatic reminders actually get invoices paid faster?" (asking, to judge one option)
3. "Which tools send reminders and sync with QuickBooks?" (asking, to build a shortlist)

One keyword became three pages' worth of questions, and none of them contains the keyword. That's the practical meaning of search intent evolving: the unit you plan around is the question and the decision behind it.

How to Update Your Search Intent Research

You don't need to throw out your keyword research. Add a conversational layer on top of it.

1. Keep the classic labels as a first pass. They still tell you what kind of page a query wants.
2. Collect real questions. Mine sales calls, support tickets, demo forms and long queries in Search Console. Our free AI question generator can suggest more.
3. Write down the decision behind each question. "What is X?" often means "should I use X?" Say what the reader is trying to decide.
4. Add the next question. For each question, note what the person would ask next. That's the follow-up your page should also answer.
5. Check the live answer. Ask the question in ChatGPT, Perplexity and Google AI Mode, and note what shape the answer takes and which pages it cites.
6. Re-check twice a year. Most of the studies above came out in the past 12 months, and the numbers are still moving.

Writing for these questions is covered passage by passage in our guide to optimizing content for AI search.

Where Rankbox Fits in Intent Research

Rankbox's Answer-Space Research maps the questions buyers ask ChatGPT, Perplexity and Google in your category and scores each one for volume, difficulty and search intent, using model estimates rather than measured counts. Its Citation-Ready Writer then researches the live web and drafts source-backed articles of 2,000 to 3,500 words for the questions you approve. Rankbox doesn't track AI citations today. See pricing.