How AI Agents Browse, Compare and Buy: What Marketers Must Prepare

AI agents browse by running many searches at once, compare by matching products against criteria they infer from the shopper's request, and buy either through a protocol the merchant integrates or by completing the merchant's own checkout. Every step runs on your product data, so that's what marketers must prepare. This guide walks through each step using Google's and OpenAI's own descriptions, then gives a prep list.

Method: we read Google's May 2025 and January 2026 shopping announcements, OpenAI's commerce docs and crawler docs, and Shopify's agentic storefronts page on 31 October 2026 (CET). Quotes are verbatim. Opinions are marked "Our view"; examples are marked "Illustrative".

On this page

• Step 1: How do agents browse?
• Step 2: How do agents compare?
• Step 3: How do agents buy?
• What must marketers prepare?
• What should you measure?
• Summary
• From MentionX data: what agents read today
• Worked example: Ledgerline prepares for agent buyers
• Which page elements matter most to agents?
• How do product feeds help?
• What should you do about policies?
• How do agents handle accounts and payments?
• How do you track agent visits?
• What should marketers prepare first?
• How do you keep facts consistent across channels?
• How do you test what agents see?
• Common questions
• Sources

Checked on 31 October 2026. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.

Step 1: How do agents browse?

Google's description of AI Mode shopping: when a shopper narrows a request, "AI Mode will start a query fan-out, which means it runs several simultaneous searches" to work out what makes a product suitable, "and then use those criteria to suggest" options (Google). It draws on the Shopping Graph, which Google said had "more than 50 billion product listings" in May 2025 (Google).

For tasks like tickets, Google said AI Mode would look "across sites to analyze hundreds of potential ticket options with real-time pricing and inventory, and handle the tedious work of filling in forms" (Google).

On ChatGPT, OpenAI's search crawler is OAI-SearchBot, and visits made for a user's request use ChatGPT-User. OpenAI says for ChatGPT-User, "Because these actions are initiated by a user, robots.txt rules may not apply" (OpenAI). Merchants can also send product feeds; onboarding "is currently available to approved partners" (OpenAI).

Our view: "browsing" for an agent means reading feeds, indexes and pages in parallel. It never sees your homepage carousel.

Step 2: How do agents compare?

Neither Google nor OpenAI publishes a ranking formula. What they do say:

• Google: Merchant Center data is used "to match your products to the right queries" (Google Merchant Center).
• OpenAI: recommended feed attributes "improve ranking, relevance, and user trust" (OpenAI).
• Google, January 2026: new Merchant Center attributes cover "answers to common product questions, compatible accessories or substitutes" (Google).

Fictional example: "Find me a carry-on suitcase under €200 that fits Ryanair's cabin limits and has a USB port." The agent infers criteria: dimensions, price, USB port, perhaps weight. It then compares candidates on those facts. A listing with exact dimensions in centimetres, a price in euros and "USB charging port" in its attributes can be checked; one with "cabin-friendly" can't.

What about reviews and reputation?

Google said the Shopping Graph has listings "each with details like reviews, prices, color options and availability" (Google). Our view: review volume and content are part of what agents compare, but no platform says how much weight they carry.

Step 3: How do agents buy?

There are two models.

Model A: the agent completes your checkout

Google, May 2025: when a shopper is ready, they "confirm the purchase details and tap 'buy for me'. Behind the scenes, we’ll add the item to your cart on the merchant's site and securely complete the checkout on your behalf with Google Pay" (Google). Here the agent uses your existing site, so checkout friction directly blocks sales.

Model B: the merchant integrates a protocol

OpenAI's Agentic Checkout Spec lets merchants "run end-to-end checkout flows inside ChatGPT while keeping orders, payments, and compliance on their existing commerce stack" (OpenAI). Payment details are shared via a delegated payment spec, and charges are "restricted by the delegated payment’s max amount and expiry" (OpenAI).

Google's Universal Commerce Protocol (January 2026) "will soon power a new checkout feature on eligible Google product listings in AI Mode in Search and the Gemini app". Google says "Retailers remain the seller of record" (Google).

Shopify connects stores to these channels. It lists "ChatGPT, Google AI Mode and Gemini, Microsoft Copilot, and Meta" among AI channels (Shopify).

More detail: agentic checkout protocols.

What must marketers prepare?

1. Product data that can be checked

Exact specs, prices, availability and variants in your feeds. Google asks that feed price "match with the price from your landing page, structured data, and at checkout" (Google Merchant Center). See Merchant Center feeds for AI shopping.

2. Pages agents can read

Google asks for "important content" in "textual form" and structured data that "matches the visible text on the page" (Google). Put delivery times, return windows and warranty terms in plain text on product pages.

3. Access for the right bots

Allow Googlebot and OAI-SearchBot in robots.txt and at your CDN. Google's guide asks that crawling be "allowed in robots.txt, and by any CDN or hosting infrastructure" (Google).

4. A checkout an agent can finish

Our view: guest checkout, labelled form fields, no surprise fees at the last step, and risk-based rather than blanket CAPTCHAs. These help humans too.

5. Policies agents can quote

Return, shipping and price-match policies in plain text, ideally with matching Organization or merchant markup. A brand facts page is a good home for company-level facts.

6. Offers agents can find

Google's Direct Offers pilot lets advertisers "present exclusive offers for shoppers who are ready to buy" in AI Mode (Google). It's a Google Ads pilot, so check eligibility before planning around it.

What should you measure?

• Share of category prompts where your products appear, by surface
• Accuracy of prices and stock shown in answers
• Orders from agentic channels, if your platform reports them
• Checkout completion in agent test runs

Our view: run a monthly test where an agent tries to buy one product from you. It's the fastest way to find blockers.

Summary

From MentionX data: what agents read today

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

Agents and AI answers draw on the same public pages. In our sample, cited URLs spread across these types:

Brands were named most often when buyers asked comparison questions:

Our view: comparison is where agents and AI answers name brands most, so the facts agents compare on, such as price, features, delivery and returns, need to be easy to find and consistent everywhere.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.

Worked example: Ledgerline prepares for agent buyers

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

Ledgerline sells subscriptions, not physical goods, but buyers increasingly ask AI assistants to compare AP tools and start trials. The team walks through the three steps.

Browse: it checks that product, pricing and integration pages are public, fast and readable without scripts. Compare: it adds clear plan tables with prices, limits and included features, and makes sure the same facts appear on review site profiles. Buy: it reviews the trial sign-up flow and removes a step that requires a phone call for the smallest plan. It then checks server logs for agent user agents and runs a set of comparison prompts monthly to see how assistants describe its plans.

Which page elements matter most to agents?

Agents look for the facts that decide a purchase: price, availability, variants, delivery times and costs, return policy, warranty and reviews. Put these in visible text near the top of product pages, and keep them consistent with your structured data and feeds. If shipping costs only appear at checkout, agents comparing total cost may skip you or estimate wrongly.

How do product feeds help?

Feeds give shopping engines and agents structured, current product data. Keep titles, prices, availability and identifiers accurate, and update feeds whenever prices or stock change. Mismatches between feed and page are a common reason for products being excluded or described wrongly.

What should you do about policies?

Write return, shipping and warranty policies in plain language on their own pages, with clear headings and dates. Link them from product pages. Agents asked to find a product with free returns need to verify that claim quickly. Vague or buried policies make that hard.

How do agents handle accounts and payments?

Many agents still hand off to the user at checkout, especially for payment and sign-in. Some platforms are testing protocols that let agents complete purchases on a user's behalf with permission. Whatever the method, a short, reliable checkout with guest options and clear errors helps both people and agents. Our view: fix the basics of checkout before investing in agent-specific protocols.

How do you track agent visits?

Check server logs for user agents that identify AI assistants and agents, and for fetches triggered by users. Group requests by page type and see which pages agents read most. Watch for errors and blocked requests. If agents can't fetch your pricing page because of bot protection, they'll compare you on incomplete information.

What should marketers prepare first?

Start with an audit of your top products or plans. For each, confirm the price, availability and policies are visible in text, consistent across your site, feed and marketplaces, and readable without scripts. Then test a few comparison prompts in assistants and see what they get right and wrong.

How do you keep facts consistent across channels?

Use one source of truth for prices, stock and policies, and feed every channel from it: your site, feeds, marketplaces and partner listings. Check a sample of products each month across channels. Inconsistencies confuse agents and erode buyer trust.

How do you test what agents see?

Ask a few assistants to find, compare and recommend products in your category, and note what they say about yours. Check prices, features and policies against your pages. Record errors and trace each one to a source. Repeat the test monthly, because agent behaviour changes quickly.

Common questions

Do agents read structured data? Some do, and shopping engines rely on feeds and markup. Visible text matters too.

Should I block AI agents? Blocking them may remove you from comparisons. Decide deliberately and document the choice.

Will agents replace shopping sites? Unclear. Prepare for both people and agents using the same pages.

Sources

All accessed 31 Oct 2026; pages may have changed since.

• Google, Shop with AI Mode (20 May 2025); AI Mode updates from I/O 2025 (20 May 2025); New tech and tools for retailers (11 Jan 2026)
• Google, Merchant Center product data specification; Google Search Central, AI features and your website
• OpenAI, Overview of OpenAI crawlers; Get started; Key concepts; Agentic Checkout Spec; Delegated Payment Spec
• Shopify, Agentic storefronts