AI Search Intent: The 4 New Conversational Buyer Stages

This shift in AI search intent is well past early adopters. In G2's March 2026 survey of 1,076 B2B software buyers, 51% said they now start software research in an AI chatbot more often than in Google, up from 29% eleven months earlier. Google says the average AI Mode query is triple the length of a traditional search, and that follow-up queries in AI Mode grew by more than 40% a month on average in the US.

SEO still files every keyword under one of four types: informational, navigational, commercial or transactional. A buyer talking to an assistant can pass through all four in five minutes, without typing a keyword at all. This guide sets out the four stages, the dated evidence for each, and the content that wins each turn. For how engines decode a single question, see how AI search interprets user intent. For memory, location and chat history, see how AI search uses user intent and context. For the history and the usage data in depth, read how search intent is evolving with conversational AI assistants.

Why AI Search Intent Folds Four Intent Types Into One Chat

The classic search intent types assume one query carries one goal. "Invoice software" is commercial. "Tallyfold login" is navigational. "Tallyfold pricing" is transactional. Each query gets its own results page, and each page gets its own keyword.

AI search intent doesn't work that way. The buyer explains a problem, asks for options, questions one option, then asks for the next step, all in one thread. The engine keeps the earlier turns in view. OpenAI says that when you ask follow-ups, ChatGPT "will consider the full context of your chat". At I/O 2026, Google said that when you move from an AI Overview into AI Mode, "your context stays with you".

The prompts also stop looking like keywords. In Semrush's clickstream study of US ChatGPT use from October 2024 to February 2026, 65% to 85% of prompts couldn't be matched to any keyword in its 27-billion-keyword database for most of the period. The same study found sessions getting longer: average prompts per session rose 50% in four months, to 1.75 by February 2026.

That last number is also a warning against overstating the change. An average of 1.75 prompts means many chats are still one or two turns long. Buyers often spread the stages across several sessions and channels. Google's own AI Mode report says shoppers "often begin their journey with traditional Search and click into AI Mode to dive deeper." So AI search intent can collapse into one dialogue, but you should plan for each stage to happen anywhere.

Rankbox's Four-Stage Model of AI Search Intent

Rankbox built this model of AI search intent to plan content, not to describe how any engine works inside. No AI product labels conversations with these names. They sort buyer turns by what the buyer needs next, which tells you what page you need published.

The stages map loosely onto the old types. Exploration is mostly informational. Synthesis is commercial. Interrogation mixes commercial and navigational, because the engine often goes to your own domain. Handoff is transactional. The difference is that in AI search intent, one conversation carries all of them, and the buyer's context rides along from turn to turn.

Stage 1: Prompt Exploration

In Prompt Exploration, the buyer describes a situation rather than a product. They don't know the category name yet, or they don't care. "Our clients keep paying late and I lose a day a month chasing them" is a Stage 1 prompt. At this point, AI search intent is broad: the buyer wants to understand the problem before choosing a fix.

What the usage data shows

People now bring the whole problem to the first prompt. Google's one-year AI Mode report, covering May 2025 to April 2026, says people are "no longer worrying about the 'right way' to formulate their questions." Brainstorming queries in AI Mode grew 30% faster than AI Mode queries overall since launch, and searches starting with "where to", "where should I" and "ideas for" are growing. For how AI Mode differs from the classic results page, see Google AI Mode vs. traditional search.

OpenAI's own study points the same way. In the NBER working paper "How People Use ChatGPT" (September 2025), about 49% of messages were "Asking": seeking information or advice "to inform a decision." Seeking Information grew from 14% to 24% of all usage in a year, and the authors call it "a very close substitute for web search."

What the engine does

Two things happen, and you can influence both. First, the engine decides whether to search at all. In Semrush's data, ChatGPT ran a web search on 34.5% of prompts in February 2026, down from 46% in late 2024. An answer written from memory can only name brands the model already knows.

Second, when it does search, it splits the problem into sub-searches. Google calls this query fan-out: "issuing multiple related searches concurrently across subtopics and multiple data sources," as its AI Mode launch post put it.

The asset that wins Stage 1

A problem guide. It names the causes of the problem, the options for fixing it, and the trade-offs, with real numbers. It mentions your product only where it fits. The guide wins because each fan-out search lands on one slice of the problem, and your guide has a section for each slice. Our guide to optimizing content for AI search covers writing those sections so they stand alone.

Stage 2: Solution Synthesis

In Solution Synthesis, the buyer asks the engine to turn the problem into a shortlist. "Which invoicing tools do that for a small agency?" is the classic Stage 2 turn, and it's where AI search intent turns commercial. This is the stage most GEO advice already targets, and it's where brands first get named or left out.

What the usage data shows

Google says searches beginning with "which" grew 40% faster than AI Mode queries overall in the six months to April 2026, with "which of" and "which one" growing most. Its reading: "people are increasingly using AI Mode to help decide."

For software, G2's April 2026 release found AI chatbots were the top source shaping which vendors make a buyer's shortlist, and that 69% of buyers had picked a different vendor than planned because of a chatbot's advice. Comparing vendor strengths and weaknesses was the top use of chatbots in software research, at 41%. G2's July 2026 Buyer Behavior Report, a separate survey of more than 1,000 buyers, then found review sites (38%) had edged past AI chatbots (37%) as the top shortlist source. G2 runs a review site, so weigh its framing with that in mind. Either way, chatbots now sit at the center of the shortlist: in the July survey, 82% of buyers had sourced software recommendations from an AI chatbot in the prior two years.

What the engine does

The engine assembles a list from pages that already compare options: roundups, review sites, comparison pages and vendor pages. OpenAI's March 2026 shopping update now shows products "side by side with key details like price, reviews, and features." Microsoft's Copilot usage report, which studied 37.5 million conversations from 2025, even has its own intent label for this job: "Shopping and Product Research," defined as learning about and comparing products.

How that list is built and ordered is covered in how to rank on ChatGPT and how AI models rank brands in search results.

The asset that wins Stage 2

Pages that make comparison easy for the engine and fair to the buyer: a comparison page against your two closest rivals, use-case pages ("invoicing for design agencies"), and current profiles on the review sites your category uses. Our comparison page formula shows how to write one that engines quote.

Stage 3: Dealbreaker Interrogation

In Dealbreaker Interrogation, the buyer questions one or two shortlisted products until something breaks or nothing does. The questions are narrow and personal: "Does it take pounds and euros? What will five users cost? Do my clients need an account to pay?" AI search intent is at its most specific here, and this stage decides the deal. It's also where a site with only blog posts has the least to offer.

What the usage data shows

Follow-ups are growing fast in Google's AI Mode data: follow-up queries rose by more than 40% on average per month in the US. When people shop in AI Mode, the top attributes they look for are price, location, color, brand and availability, in that order. Among follow-ups about stores, "in stock" and "car dealerships with financing" both make Google's top 10.

In B2B software, G2's July 2026 report found that evaluation is now the longest stage of buying, "surpassing research for the first time." Evaluation is where buyers "scrutinize pricing, assess security, pressure-test implementation." IT security review was the biggest source of delay (39% of buyers, 50% at enterprises), ahead of budget approval (32%) and implementation planning (25%).

What the engine does

It carries the conversation forward and runs narrow searches, often aimed at your own site. Nectiv's August 2026 study re-ran about 4,000 prompts and found ChatGPT used a site: search in 64% of its fan-out queries, with "official" among its most common words. Software prompts averaged 10.7 fan-out searches each, which the author guesses is because engines must confirm features, pricing and capabilities for complex purchases.

If your site doesn't state the answer, the engine fills the gap from somewhere else, or guesses. In G2's April survey, 64% of buyers said they see inaccurate AI chatbot recommendations "often or very often." A missing price is one common version of this, covered in our post on hallucination by omission.

The asset that wins Stage 3

A set of plain, factual pages that answer dealbreakers directly:

• A pricing page with worked totals. "Three users included, $12 per extra user" is good. "Five users: $63 a month" is better, because the engine doesn't have to do the math.
• Integration and compatibility docs. One page per integration, with what syncs and what doesn't.
• A security and compliance page. Who holds card data, which certifications you have, where data lives.
• A limits page. Who the product isn't for, and what it doesn't do. It feels risky, but it answers the question the buyer is about to ask anyway.

To find the dealbreakers AI answers already raise about you, run the prompt set in our defensive GEO guide.

Stage 4: Action Handoff

In Action Handoff, the buyer asks the engine to take the next step: start a trial, book a demo, reserve a table, buy the thing. The engine either links out to the page where that happens, or an agent does part of it for them. This is where AI search intent turns into action. For most B2B products today, the handoff is still a link.

What the usage data shows

Action is growing, but trust lags. Google says AI Mode planning queries grew 80% faster than AI Mode queries overall in the six months to April 2026. Yet in G2's July 2026 survey, only 9% of software buyers were comfortable letting an AI agent make purchases within approved guardrails, and 2% without pre-approval. Most (47%) wanted agents to research and recommend while humans decide.

Agent logs tell a similar story. A Perplexity and Harvard study of Comet Assistant use from July to October 2025 found shopping was 10% of agentic queries. Productivity (36%) and learning (21%) led by a wide margin.

Handoff features you can verify, as of September 2026

Every row below comes from the vendor's own page. These features change often, so check the linked page before you plan around one.

Sources: OpenAI's product discovery update, shopping help page, search help page and agent help page; Google's agentic AI Mode post, UK booking post, I/O 2026 Search post and Universal Cart post; Microsoft's Copilot Checkout post; Perplexity's Instant Buy help page, with the launch date from PayPal's announcement.

Notice the pattern. Most of these features end on someone else's page. Google's booking "links you directly to the booking page, so you can easily take the last step," and OpenAI explained its checkout change by saying the first version of Instant Checkout "did not offer the level of flexibility that we aspire to provide." Even Copilot Checkout keeps the merchant as merchant of record. And every checkout program above is built for consumer goods, so for a B2B product the handoff usually lands on your own site.

The asset that wins Stage 4

A next-step page that works for a person arriving mid-decision, and for an agent acting for one:

• A trial or signup page with the steps in plain text, no required sales call, and the price restated.
• A demo booking page with a live calendar, not a "contact us" form that starts an email thread.
• Clear links from every Stage 3 page to the next step, so a buyer who lands on your pricing page from an answer can act in one click.
• Product feeds, if you sell goods. Shopping results come from merchant feeds, as the how to rank on ChatGPT guide explains.

A Tallyfold Buyer Conversation, Turn by Turn

Tallyfold is a fictional B2B invoicing and payments app for agencies. Brindlework and Kestrelyn are its fictional rivals. The buyer, the prompts, the engine's searches and every figure below are invented for illustration. None of it was captured from a real engine.

The buyer runs operations at a 14-person design agency. They send about 40 invoices a month, bill clients in the US and UK, and use QuickBooks. Five people need access. Here is how the buyer's AI search intent moves across four turns.

Turn 3 is where Tallyfold wins or loses. Its pricing page says "$39 a month, three users included, $12 per extra user." A person can work out five users: $39 + (2 × $12) = $63 a month, or $756 a year. An engine might do the same math, or it might just quote "$39 a month," which is wrong for this buyer. Publishing "Five users: $63 a month" removes the risk.

The Turn Coverage Audit

Now score the site. List the assets each stage needs, mark each one present or missing, and divide.

The total says Tallyfold covers less than half of the buyer's AI search intent. The rows say where to start: Stage 3, at 25%. That's also the stage G2's data says decides B2B deals. So Tallyfold's next three pages are a pricing page with seat totals, a security page stating who holds card data, and a limits page ("Tallyfold has no built-in timers; teams that bill hourly may prefer a tool that does").

The published figure in Stage 1 has to come from Tallyfold itself, such as its own median days-to-payment. No writer, human or AI, can invent that honestly.

How to Audit Your Site for AI Search Intent

You can run this audit in an afternoon with a spreadsheet. It works for any product with a considered purchase.

1. Collect 20 to 30 real buyer questions. Pull them from sales calls, support tickets, demo forms and Search Console. Our free AI question generator and AI Visibility Prompt Kit help fill gaps.
2. Tag each question with its AI search intent stage. A problem with no product in it is Stage 1. "Which" and "best" are Stage 2. Questions naming your product are Stage 3. "Start", "book" and "buy" are Stage 4.
3. List the asset each question needs. Use the table in the model section as a starting list.
4. Mark each asset present or missing, and score coverage by stage, as in the Tallyfold audit.
5. Test the answers. Run each Stage 3 question in two or three engines, in a clean session. Mark each answer right, wrong or "couldn't find."
6. Fix the lowest stage first, then re-test the same questions a few weeks after the new pages are indexed.

Keep the audit honest about noise. AI answers change from run to run, so one check proves little. Our guide to measuring GEO covers how many runs you need before a change is real.

Two mistakes come up often. The first is building only for Stage 2, because "best X" lists feel like the prize. A buyer who reaches Stage 3 and finds no pricing detail goes back to the shortlist. The second is treating AI search intent as a keyword label. The same words can sit in two stages: "Tallyfold pricing" is Stage 3 if the buyer is comparing, and Stage 4 if they're ready to pay.

How Rankbox Helps You Cover All Four Stages

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 intent. Its volumes are model estimates, not measured counts. That gives you the raw list for mapping AI search intent in step 1 of the audit. The Citation-Ready Writer then researches the live web and drafts 2,000 to 3,500-word, source-backed articles, which suit Stage 1 guides and Stage 2 comparisons. Brand Voice applies the tone, audience and product details you give it, so pricing and limits stay accurate when your product comes up.

Rankbox doesn't build pricing pages, security pages or checkout flows, and it doesn't track AI citations today. Articles reach your site through its API, which a developer wires in. The Business plan is $49.50 a month with a 7-day trial. See pricing.