# Search Prompt: What It Is and What AI Does With It

> What a search prompt is, how Claude, ChatGPT or Google's AI Mode turn it into web queries, and what your product page has to answer.

URL: https://clione.ai/glossary/search-prompt/
Lang: en

A prompt is any instruction you give an AI model: asking it to translate, to summarize or to write a [product description](/glossary/product-description/). A search prompt is the specific case where that instruction **ends up as a web search**, because the answer depends on data the model does not have or that changes: which products exist, what they cost, what people who have tried them think. It is the front door of [AI search](/glossary/ai-search/). It matters to a store for a practical reason: instead of three loose words, the buyer can write a paragraph with their conditions, and the assistant goes out looking for pages that meet them. Your [product page](/glossary/product-page/) competes on each of those conditions, not on a keyword.

## How it works

Between what the person types and what reaches the search engine there is a step you cannot see. Anthropic's documentation on Claude's web search tool tells it in three beats: **Claude decides when to search based on the prompt**, the API runs the searches and returns the results, and at the end of its turn Claude answers, citing sources. The cycle of searching and reading can repeat several times within a single request.[^anthropic-web-search]

The first decision is whether to search at all. According to that same documentation, Claude searches when the request depends on current or changing information, or on information that was not in its training data: recent news, prices, facts about organizations or **specific products that may have changed**, or an explicit request to search. And it answers without searching when stable knowledge is enough, such as a settled fact, an idea for a piece of creative writing or the analysis of something already in the conversation.[^anthropic-web-search] A purchase question almost always falls in the first group.

The second decision is what to search for, and that is where the prompt and the query part ways. In the sample response Anthropic publishes, the model announces it will search for when Claude Shannon was born, and the query it runs sits in a separate block, in different words: "claude shannon birth date". **The model itself writes the query.** And the number of queries depends on what the prompt asks: a simple factual question usually needs 1 to 3 searches, and research that compares several options or several entities can go past 10.[^anthropic-web-search]

Google describes the same move under another name. [AI Overviews](/glossary/ai-overviews/) and [AI Mode](/glossary/google-ai-mode/) can use the [_query fan-out_](/glossary/query-fan-out/) technique: they run several related searches across subtopics and data sources to build the answer.[^google-ai-features] Its guide to the generative AI features defines it as a set of simultaneous, related queries the model generates to ask for more information.[^google-ai-guide] The prompt is the input; the fan-out queries are what the system does with it.

### Why a prompt does not look like a query

Google explains that AI Mode is especially useful for questions that call for exploring, reasoning or making complex comparisons, and that it lets you ask **questions with nuance that would previously have needed several searches**.[^google-ai-features] In a launch announcement on October 8, 2025, Google gave a figure: **early AI Mode users ask questions two or three times longer than traditional search queries**. And it gave an example that is a textbook prompt: "I want to understand the different methods of making coffee. Make a table comparing the differences in taste, ease of use and equipment needed," followed by a follow-up question about the ideal grind size. The same post notes that you can ask with text, voice or a photo.[^google-es-launch]

That example has the pieces a search prompt usually carries, and it helps to separate them because they do not all affect you equally:

| Part of the prompt | In Google's example                  | What it asks of a store                                  |
| ------------------ | ------------------------------------ | -------------------------------------------------------- |
| The need           | Understand coffee-making methods     | That the page deal with what the person is after         |
| The conditions     | Taste, ease of use, equipment needed | That every fact is written, as text, on the product page |
| The format         | A comparison table                   | Nothing: the assistant sets the format                   |
| The follow-up      | The ideal grind size for each method | That the page also answers the next question             |

## In a store

A **fictitious example**. A person types into an assistant with search: "I'm looking for a pan that works on induction, without a PTFE nonstick coating, that I can put in the oven and that isn't heavy because I have a sore wrist. Compare three for me and tell me which is easiest to clean."

No product page is going to answer that whole sentence, and none needs to. What the assistant needs to find are pages that answer each condition: induction compatibility, coating material, maximum oven temperature, weight and how it cleans. The "sore wrist" is context: it turns weight into the deciding fact. A page that says "lightweight and versatile" answers none of that; **one that says whether it works on induction and gives the weight in grams, the coating material and the maximum oven temperature answers four of the five conditions** even if it never uses the prompt's words.


That the prompt's words are the keywords the assistant will search for. They are not: the model decides whether to search and writes the queries itself, which may be other words and may be several for a single prompt.[^anthropic-web-search] On Google it works the same way: fan-out runs several related searches on subtopics of the question, not the question as written.[^google-ai-features] Copying verbatim prompts onto a product page does not bring it closer to the answer; writing the facts those prompts ask for does.


## How to check it

1. **Write three prompts the way your customer would**, without your brand and with their real conditions: the size, the material, the use, the limitation they worry about.
2. **Underline each condition and look for it on your product page.** If the fact is not written on the page, no assistant can cite you for it.
3. **Run those prompts in AI Mode** and note, with the date, which pages it links and which part of the answer each one supports.[^google-ai-features]
4. **If you have access to the Claude API, turn on web search** and read in the response the queries the model generated from your prompt: it is the most direct way to see the distance between what is asked and what is searched.[^anthropic-web-search]
5. **Repeat on different days and in new chats.** A single run is a sample of one.

## Related concept: conversational query

The [conversational query](/glossary/conversational-query/) describes the way of asking: in natural language, in full sentences and with follow-ups that build on what came before. The search prompt is the whole piece the model receives, and it can carry things that are not a question: a formatting order ("make a table"), a limit ("just three") or a personal detail that changes the criterion ("I have a sore wrist"). **A conversational query can go to a classic search engine; a search prompt always passes through a model that decides what to search.** That is why, with a prompt, what reaches the web is the queries the model writes, not the person's sentence.[^anthropic-web-search]

To go further, [_query fan-out_](/glossary/query-fan-out/) explains how one question splits into several searches, the guide to [what AI bots see in your store](/guides/what-an-ai-crawler-sees/) goes down to what they read on your product page, and the [AI search reading path](/learn/ai-search/) orders the rest of the terms.