How People Search with AI: Long Prompts, Questions and Decisions

With AI, people write full questions, with context and constraints, and keep the conversation going until they decide. According to Google, the average search in AI Mode is three times the length of a traditional search. For anyone producing content, that means covering situations and not just keywords. The big picture is in SEO, AEO and GEO: the difference.

Quick summary

• Prompts are long and describe a situation: who the person is, what they need and what limits them. A bare keyword becomes the exception.
• Google reports that the average AI Mode search is three times longer than a traditional one, and that use for planning and idea generation is growing.
• A study of ChatGPT usage (NBER, 2025) shows that practical guidance, information seeking and writing add up to almost 80% of conversations.
• One question can turn into several searches behind the scenes: Google confirms that AI Overviews and AI Mode may use query fan-out.
• What changes in content: cover context variables, answer in sequence, compare honestly and keep data current.
• No one knows for sure how AI systems choose the sources they cite. Treat recommendations as hypotheses to test.

How to read this: if you are just starting, read the sections on prompts versus keywords, the decision journey and the step-by-step for discovering your audience's questions. If you already do SEO, go to the sections on fan-out, what is still unknown and the use of Search Console data. For terms such as prompt, LLM and fan-out, see the SEO, AEO and GEO glossary.

How do people search with AI today?

They search by conversing. Instead of typing "women's running shoes", they write out a whole situation and then refine it with follow-up questions. Public data confirms the direction, although each source looks at a different platform and period.

In Google's AI Mode, according to Google's own report on usage in the United States, the average search is three times the length of a traditional search, and planning and idea queries are growing faster than searches overall. Google also cites the growth of phrases such as "where" and "ideas for". According to specialized press coverage, the data spans May 2025 to April 2026 and refers to the US market. This source has no equivalent figure for other countries.

For ChatGPT, the study How People Use ChatGPT, published in September 2025 by the NBER with researchers from OpenAI and Harvard, tracked usage from November 2022 to July 2025 using a privacy-preserving method. It concludes that practical guidance, information seeking and writing account for almost 80% of conversations. In other words, people use AI to learn and decide as much as to produce text.

In plain English: searching traditional Google is like consulting a library index: you name the subject and get a list of books. Searching with AI is like talking to a librarian: you explain your problem, they hand back a summary and you answer "what if I have a small budget?".

What changes between a keyword and a prompt?

The length, the context and the expectation change. A keyword is a short subject label; a prompt is a request with a situation, constraints and an answer format. Someone searching a keyword expects a list of links; someone writing a prompt expects a ready answer, with sources when the platform shows them.

	Traditional search	AI search
Format	Loose words or a short phrase	Full question, with context
Example	"crm for small business"	"I run a store with 3 salespeople, we use text messages and a spreadsheet, and I want a CRM that costs no more than $30 a month. Which should I pick?"
Result	List of pages	Synthesized answer, sometimes with cited sources
Continuation	New search from scratch	Follow-up question that inherits the context
Click	The main way to reach the site	Optional; the answer may be enough
Measurement	Position and clicks	Citations, mentions and presence in answers

For SEO professionals: the prompt often contains several intent attributes at once (type of business, size, budget, tool already in use, decision criterion). That pushes demand into the extreme long tail: many variations with negligible individual volume, which keyword tools tend not to show. Search Console data on queries with many words is the best picture available of that tail.

What does the decision journey look like in an AI conversation?

The journey that used to be spread across several searches and tabs now fits in a single conversation. The person starts with a broad question, refines it with constraints, asks for comparisons, asks for a recommendation and sometimes asks for the next step. What used to be "research, comparison and decision" becomes a sequence of turns.

• Discovery: "What is out there for managing inventory for a small shop?"
• Refinement: "It needs to work with my card reader and run on my phone."
• Comparison: "Compare the two that are left on price, support and learning curve."
• Decision: "Which would you pick for a team of three with no IT staff?"
• Action: "Give me a checklist for migrating my current spreadsheet."

Google notes a technical detail that affects this. The documentation on AI features in Search says AI Overviews and AI Mode may use the query fan-out technique, issuing several related searches across subtopics to build the answer. In practice, a page can be consulted for a sub-question you never targeted directly. Understand the mechanism in query fan-out: why one search becomes many questions.

In plain English: it is like asking an assistant to "organize my trip". They do not run a single search: they look up flights, hotels, weather and activities separately and put it all together. AI does something similar with your question, and your page may be the source for just one of those parts.

For SEO professionals: there is no stable "position 1" in AI answers. The answer varies with the prompt wording, history and location, and changes between runs. Measure presence across a set of prompts, run several times and with the date recorded, not position on one isolated prompt.

As for how much these people actually click through to check a recommendation, and at what point, we found no reliable public data. Treat any number claiming to show it with suspicion.

What changes for people who produce content?

What it means to "cover a topic" changes. Instead of one page per keyword, you need to answer the same question across the context variations in which it appears. Five practical changes:

1. Cover the decision variables. Who it is for, who it is not for, price range, business size, prerequisites. People writing prompts put these variables in the request.
2. Answer in sequence. Put the direct answer first and then the questions that usually follow. A FAQ that AI cites with real questions is the simplest way to do this.
3. Compare honestly. Tables with clear criteria, and stated limits of the product or service, serve decision-makers better and are easier to reuse in an answer.
4. Keep data current and dated. Price, lead time and features change, and a synthesized answer built on stale information hurts whoever trusted it.
5. Write self-contained passages. Each section should make sense outside the page's context, as explained in how to structure content that LLMs can cite.

Google reminds readers, in its guide to optimizing for generative AI, that these are SEO principles and not separate tactics: unique, helpful content, a technically accessible page and a good experience. It also says you do not need to fragment text into tiny pieces. Clarity helps; tricks do not.

How do you discover your audience's prompts?

Take the questions your customers already ask, write them as prompts and test. You do not need an expensive tool to start. The steps below take an afternoon:

1. Collect real questions. Chat messages, support emails, sales questions and reviews say more than any keyword list.
2. Look at the long queries in Search Console. In the performance report, use the query filter with a regular expression (regex) to bring up only those with 7 words or more.
3. Rewrite each question as a prompt: situation, constraints and request ("I am X, I have Y, I want Z, what do you recommend?").
4. Run the prompts on the platforms your audience uses, note whether your brand appears, which sources are cited and what the answers say about you.
5. Group by theme and turn each group into a page, a section or an FAQ.
6. Repeat every month, because answers change. The templates in prompts to test your brand's visibility in AI help standardize it.

For the step 2 filter, use an expression like:

^(\S+\s+){6,}\S+$

It selects queries with seven words or more. Adjust the number to your site's volume.

Practical example: choosing a CRM

Hypothetical example: a company that sells a simple CRM to small businesses analyzes how its audience is likely to ask. It lists four probable turns and defines what content answers each one:

• "Which CRM works for a shop with 3 salespeople who sell through text messages?" A "CRM for small teams" page with the direct answer in the first lines and who the product is for (and who it is not for).
• "How much does it cost and what is included?" A plan table that is current and dated.
• "How do I migrate from a spreadsheet?" A short step-by-step, with the accepted files and the estimated setup time.
• "Does it integrate with my card reader?" A list of available integrations, without promising what does not exist.

The result is not "an article about CRMs" but four useful blocks, each able to answer a sub-question on its own. That does not guarantee being cited, but it raises the chance that the page is useful at any point in the conversation. Note: none of the data in the example is real; it is there to show the method.

What is still unknown about AI search?

There are four important gaps. First: no one knows how each AI chooses the sources it cites, because vendors do not disclose the criteria and answers vary between runs and users. Second: the best public statistics come from the platforms themselves, which have an interest in the subject, and they concentrate on the United States. Third: click behavior after an AI answer is poorly documented. Fourth: habits change fast, and a 2025 figure may be old by 2026.

So the prudent path is to do the basics well, measure what shows up and adjust. If the question is how your site loses or gains visits from answers without clicks, read zero-click and AI Overviews.

Frequently asked questions

Have people stopped using Google?

There is no reliable public data showing that. Google integrates AI into its own results, with AI Overviews and AI Mode, and traditional search keeps existing alongside it. What has changed is that some questions now happen in conversations with AI, inside and outside Google. Plan for both forms of search.

Do keywords still matter?

They matter as a starting point for understanding the topic, the vocabulary and the demand. What changes is that a keyword alone does not describe the context of the person asking. Use keyword research to map topics and the post on search intent to understand what lies behind them.

Do I need to write my content in question-and-answer format?

You do not need to turn everything into questions, but question-style headings with a direct answer right below make a section clearer for people and for systems that extract passages. Google reinforces that there is no required markup or format. What matters is that the answer is correct, complete and easy to find.

How do I find out what people ask AI about my industry?

There is no complete public source, because the conversations are private. Use proxies: customer questions, long queries in Search Console, "people also ask" terms and manual prompt tests on the platforms. Combine the sources and treat the result as a hypothesis, not an exact measurement.

Does this apply to every business?

It applies to any business whose customers research before deciding, but the weight varies. It seems reasonable to assume that complex purchases, services and software are compared more in conversation, and that very local or impulse-purchase businesses depend more on maps and reviews, but we know of no public data that confirms this. Test the prompts in your market before investing.

Next step

Build a list of the 10 to 20 most important questions from your audience and test each one with the prompts to test your brand's visibility in AI. For ongoing tracking, Contentor's AI Visibility records, on the questions you choose, whether ChatGPT cites your brand (ChatGPT only). Use the result to decide which pages to reinforce first.

Read also:

• SEO, AEO and GEO: what's the difference and where to start
• Search intent: types, examples and how to serve each one
• Query fan-out: why one search becomes many questions