How User Search Intent Evolves Within a Conversation With an AI Assistant

Most chats are still short. Public chat logs from 2023 and 2024 average two to two and a half turns per conversation, and Microsoft said in 2023 that most Bing chat users found their answer within five turns. But the thread is getting longer. A Microsoft Research analysis of Copilot users found messages per conversation rose about 1.5 times over six months of 2024. Every extra turn is another point where your page is either in the answer or dropped from it.

This post zooms in on those turns: how search intent evolves from one message to the next, what the assistant does with a follow-up, and how to write pages that survive the second and third question. For the long view, from ranked links to written answers, read The Death of 10 Blue Links, our timeline of search from 1998 to 2026. For how intent changed over the years, see how search intent is evolving with conversational AI assistants.

What Conversation Data Shows About Follow-Ups

No AI company publishes a full log of how its users move between turns. The best evidence comes from a vendor statement, two open datasets and two 2026 studies. Each measures something slightly different, so read them side by side.

Most conversations are short, but getting longer

When Microsoft capped Bing chat in February 2023, it explained the numbers behind the cap. Its update post said "the vast majority of you find the answers you're looking for within 5 turns and that only ~1% of chat conversations have 50+ messages."

Open datasets tell a similar story. LMSYS-Chat-1M, a million conversations collected in 2023, averages 2.0 turns. WildChat, about a million ChatGPT conversations from April 2023 to May 2024, averages 2.54. Averages hide how many chats are a single exchange. Counting both sides of the conversation, the ThoughtTrace authors found that one question and one reply made up over 60% of WildChat conversations and 67% of LMSYS-Chat-1M conversations. Treat WildChat with care, too: the Microsoft Research team found it is "significantly skewed towards highly proficient 'power' users."

The trend runs toward more turns. In the Copilot data from 2024, both messages per conversation and message complexity grew about 1.5 times. Google's and Semrush's reports on longer AI sessions are covered in our history of search intent, so we won't repeat them here.

Follow-ups mostly build on the last answer

The most direct evidence on how search intent evolves within a thread comes from ThoughtTrace, a 2026 dataset from researchers at Johns Hopkins, MIT and Google Research. Participants chatted with AI models about their own tasks and noted, turn by turn, why they sent each message. An AI model then labeled how each message related to the one before.

• 57.0% of user messages extended, deepened or built on the prior request.
• 25.2% were first requests that opened a conversation.
• 12.5% were completely new requests.
• The rest, about 5%, were re-attempts or variations of the prior task.

The reasons people gave shifted as the chat went on. Task Motivation, the goal itself, drove opening turns. Later turns were driven by Task Continuation, Context Grounding (adding constraints) and expectations about content or style. In the authors' words, "user intent shifts from goal-setting to refinement as interactions progress." Even when people were unhappy with an answer, they mostly extended the task rather than abandoning or retrying it.

One caveat: participants were recruited through Prolific, and their conversations ran long, with a median of eight turns counting both sides, against two in the open datasets. The split between move types is the useful part. The length is not typical.

How Search Intent Evolves: The Follow-Up Move Map

The data says follow-ups mostly refine. It doesn't say how search intent evolves from one turn to the next. The Follow-Up Move Map is Rankbox's model of the six moves a follow-up usually makes. It's a planning tool for content, not a label any assistant uses.

Moves 1, 2, 3 and 6 are what ThoughtTrace counts as extending the prior task. Move 4 often follows a doubtful answer, and move 5 matches the study's "new variation." The order isn't fixed, but a typical thread runs from exploring (deepen, narrow) to comparing (compare, verify) to deciding (act). That arc is the backbone of our four conversational buyer stages, which plans the pages each stage needs. The Move Map works one level down, at the single turn.

What the Assistant Does With a Follow-Up

People type follow-ups as fragments: "what about in Canada?" or "is that on the cheaper plan?" Neither makes sense as a search on its own. So before it searches, the assistant has to rebuild the full question from the conversation.

Vendors describe this step in their own words:

• Microsoft, 2023. Bing's Prometheus system generated "a set of internal queries" to answer the user's question "within the given conversation context."
• OpenAI. ChatGPT search rewrites prompts into one or more targeted queries before it sends them to search providers. Our post on how AI search interprets user intent walks through OpenAI's own example.
• Microsoft, 2025. Its agentic retrieval engine for developers sends "chat history along with the original query" to a model that plans the searches.

Researchers treat this as a standard step. Apple's QReCC dataset, 14,000 conversations with 80,000 question-answer pairs, splits conversational answering into question rewriting, passage retrieval and reading. A 2025 study of real chat logs adds that rewrites get better in longer conversations, where there is more context, and that models "make plausible assumptions" about what users want.

Here is why this matters for your pages. The search that finds your page isn't the fragment the person typed. It's a rebuilt query carrying the product name, the constraint and sometimes a rival from earlier turns. Context from memory and location can join in too, as our guide to how AI search uses user intent and context explains. As search intent evolves across the thread, the rebuilt queries get more specific, and fewer pages match them.

Worked Example: Tallyfold's Carry-Forward Test

Tallyfold is a fictional invoicing and payments app for agencies. Brindlework is a fictional rival. The buyer, the turns and the rebuilt searches below are invented to show the method. None of it was captured from a real assistant.

A studio owner opens with: "What's a good invoicing tool for a small design agency?" The assistant names Tallyfold among others. Then come five follow-ups, and here is how search intent evolves for this one buyer. For each, we write the search an assistant might rebuild, and ask one question: does a single Tallyfold passage answer that search on its own?

Tallyfold answers 2 of 5 follow-ups, a 40% carry-forward rate. The first turn went well. The losses come in the middle of the thread, where search intent evolves from "what is this?" to "does it fit my case?"

The fixes are small and specific:

1. Turn 3: add one sentence to the currency page: "Tallyfold adds late fees to invoices in CAD, USD, GBP and EUR."
2. Turn 4: rename the pricing row from "Automations" to "Automatic reminders and late fees," and mark which plans include it.
3. Turn 5: publish a fair Tallyfold vs Brindlework page that covers late fees and price.

With those three changes, each of the five rebuilt searches has a passage that answers it: 5 of 5, or 100%, on paper. Whether assistants then use those passages is something to test, not assume.

How to Write Pages for Every Turn

You can't write a page for every follow-up, and Google warns against trying. Its AI optimization guide says that creating separate content for fan-out queries mainly to manipulate AI answers breaks its scaled content abuse policy. The goal is fewer, stronger pages that hold up as search intent evolves.

1. List the follow-ups for each key page. Run the six moves against it. Our free AI question generator can suggest questions to start the list.
2. Put the name and the constraint in one sentence. "Tallyfold adds late fees on the Starter plan" survives a rebuilt search. "It also handles this" doesn't.
3. Answer the next question on the same page. Then link the one after it.
4. Date the facts people check. Prices, limits and plan contents are the classic verify turns.
5. Name rivals fairly where buyers compare. If you don't publish the comparison, the assistant builds it from someone else's page.
6. Test with multi-turn scripts. Write three to five turns per scenario, run them in clean sessions, and note the turn where your brand drops out. Our guide to measuring GEO covers how many runs you need before a change is real.

Where Rankbox Fits

Rankbox's Answer-Space Research maps the questions buyers ask ChatGPT, Perplexity and Google in your category and scores each for volume, difficulty and intent. The volumes are model estimates, not measured counts. That list is a good start for mapping how search intent evolves in your category, step 1 above. Its Citation-Ready Writer then drafts 2,000 to 3,500-word, source-backed articles for the questions you choose, delivered through its API. Rankbox doesn't track AI citations or run your multi-turn tests. See pricing.