# Purchase Intent: What It Is and How AI Agents Read It

> What purchase intent is, how ChatGPT and Google interpret it, how an AI agent turns it into conditions and what data your product page needs.

URL: https://clione.ai/glossary/purchase-intent/
Lang: en

"Purchase intent" is used in three places with three meanings. In market research it is the answer to a survey question ("would you buy it?"). In SEO it is often confused with [search intent](/glossary/search-intent/), which classifies a query. Here it is what a person wants to buy and under what conditions, as they tell a search engine or an assistant. It matters to a store because **assistants and [shopping agents](/glossary/shopping-agent/) work with that intent, not with your keywords**: if the person asks for a size, a price and a delivery date, what counts is whether your [product page](/glossary/product-page/) has them.

## How it works: from inferred to declared

Purchase intent has always been in searches, but the search engine had to guess it. Google explains it like this: to return relevant results, it first has to establish what you are looking for, **the intent behind your query**, and to do that it builds language models that connect the few words you type with the most useful content.[^google-how-search-works] With assistants and agents, that intent goes through four degrees, from the one that is guessed to the one that is signed.

| Degree    | How it is expressed                       | Example                                                                                     | Who interprets it                                      |
| --------- | ----------------------------------------- | ------------------------------------------------------------------------------------------- | ------------------------------------------------------ |
| Inferred  | A query                                   | "waterproof trail running shoes"                                                            | The search engine                                      |
| Conversed | A natural-language request, with context  | "I need shoes for hiking that won't get my feet wet, for less than I paid for my last pair" | The assistant, with the conversation                   |
| Delegated | An order with conditions                  | "Let me know when they drop in price and buy them"                                          | An agent, which watches and executes                   |
| Signed    | Conditions with a cryptographic signature | A mandate with store, item and maximum amount                                               | The agent, the store and whoever processes the payment |

### Inferred and conversed: what ChatGPT does

OpenAI describes it in its help center: **when your question suggests purchase intent**, for example "help me find costumes for my dog", ChatGPT may show products with images, details and links to the stores. A product appears in the carousel when ChatGPT sees it as relevant to your intent, and to decide that it takes into account the query and the context, such as memory or custom instructions. If you say your budget is 30 dollars, it pays more attention to price; if you don't, to other things. And OpenAI acknowledges the limit: **because the model interprets the intent, it sometimes gets it wrong**.[^openai-shopping-help]

To choose which products to show, ChatGPT takes into account, among other things, **structured metadata from first-party and third-party providers**, such as the price and the [product description](/glossary/product-description/).[^openai-shopping-help] That is: the person supplies the intent, but the matching depends on the data the store has published.

Google is heading the same way. In [AI Mode](/glossary/google-ai-mode/) you can describe what you are looking for the way you would tell a friend and get an answer with prices, reviews and stock information.[^google-agentic-checkout]

### Delegated and signed: what an agent does

The next step, already in [agentic commerce](/glossary/agentic-commerce/), is for the intent to be executed. On November 13, 2025, Google began rolling out its agentic _checkout_: you tell it which item you want to track, **down to the size, the color and what you want to spend**, and it alerts you when the price falls within your budget. If the store meets the requirements, Google offers to buy it for you on the store's website with Google Pay, always asking your permission first and only after you confirm the purchase and shipping details. It started in the United States with stores such as Wayfair, Chewy, Quince and some Shopify merchants.[^google-agentic-checkout]

The [AP2 (Agent Payments Protocol)](/glossary/ap2/) takes that intent one step further: it signs it. In its September 2025 announcement, Google called the record that captures the person's request an "Intent Mandate"; in a delegated task such as "buy concert tickets as soon as they go on sale", that mandate sets in advance **the price limits, the timing and the other conditions**, and lets the agent close the purchase when they are met.[^google-ap2-launch] Its official site sums it up in one principle: **verifiable intent, not inferred action**. The current specification stores those conditions in open checkout and payment mandates.[^ap2-site] Each one is an [AP2 mandate](/glossary/ap2-mandate/), and what they allow is called [delegated payments](/glossary/delegated-payments/).

## On a store

A **fictional example**: a person asks their assistant for "waterproof trail running shoes, size 10, that arrive before Friday and cost no more than what I paid last time". Each part of that intent is a condition, and each condition needs a piece of data on the product page:

| What the person asks for | The data that meets it                                  | Where it has to be                                                                    |
| ------------------------ | ------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| Trail running shoes      | The product type and the category                       | The title, the category and the [structured data](/glossary/structured-data/)         |
| Waterproof               | The membrane or the treatment, stated in words          | The description and the [product attributes](/glossary/product-attributes/)           |
| Size 10                  | That the size exists and that there is stock            | The [product variants](/glossary/product-variants/), with their availability          |
| Before Friday            | The delivery time                                       | The product page's shipping terms                                                     |
| A maximum price          | The current price, the same on the page and in the data | The product page, the structured data and the [product feed](/glossary/product-feed/) |

The row that fails most is the second. **"Waterproof" cannot be deduced from a photo**: if the product page does not say it, no assistant can match that part of the intent with your product, however well it ranks.


That purchase intent is recognized by words like "buy" or "cheap", and that putting them on the product page is enough. ChatGPT interprets intent from the query and from the context of the conversation, such as the person's memory or instructions,[^openai-shopping-help] and a request without the word "buy" can be the closest to buying. **What the product page controls is not the word, but the data that meets each condition.**


## How to check it

1. **Write five requests the way your customer would make them**, with their conditions (size, use, price, deadline), and without using the words from your product pages.
2. **For each condition, look for the data on the product page.** If a condition has no written data that meets it, that is the work still to do.
3. **Run the requests in ChatGPT** and see whether the product carousel appears and whether it respects the conditions. Repeat on different days: a single answer is a sample of one, and OpenAI itself warns that the model may misinterpret the intent.[^openai-shopping-help]
4. **Check that the [price and availability](/glossary/price-and-availability/) match** on the page, in the structured data and in the feed, because assistants read structured metadata as well as the page.[^openai-shopping-help]
5. **Review your variants**: an intent with a size or color is only met if that variant exists and has published stock.

## Related concept: search intent

[Search intent](/glossary/search-intent/) describes what a specific query is for: to learn, to go to a website or to do something, such as buy. It is a classification of searches, useful for deciding which page answers each one. Purchase intent describes the person: **what they want to buy and under what conditions**, and it can run through several searches and conversations before it becomes an order. With an agent that watches a price for days, a new search is not even needed for the purchase to happen.[^google-agentic-checkout] The [conversational query](/glossary/conversational-query/) is the format where the difference shows most: a single request can carry five conditions.

To keep going, the [shopping agent](/glossary/shopping-agent/) entry explains who carries out that intent and the [agentic commerce reading path](/learn/agentic-commerce/) puts the rest of the vocabulary in order.