What Is Agentic Search and Agentic Commerce?
Agentic search is AI that does the research itself: it runs many searches, opens pages, compares options and returns a recommendation. Agentic commerce goes one step further, letting the AI start a purchase for the user. For a business, it means being chosen by software that reads your site before any human visits.
- Agentic search means an AI keeps searching and reading until it can answer; one question can trigger many searches.
- Agentic commerce adds checkout: the agent helps the buyer pay, while the merchant stays the seller.
- Today most agent features are early, regional or waitlisted, so prepare the basics rather than rebuild your shop.
- Agents choose from what they can read: clear pages, accurate product data and open access.
What is agentic search in plain words?
Agentic search is a search where the AI decides what to look up, runs the searches, reads the results and decides whether to keep going. The user asks once. The software does the legwork.
OpenAI's developer guide describes this directly. It lists a mode called agentic search, where a reasoning model can search as part of its chain of thought, analyse results and decide whether to keep searching 1.
Older search handed you ten links to read. Agentic search hands you a finished answer, with links to the sources it used.
How does an AI agent research a question?
An agent loops. It reads the prompt, decides whether the web can help, writes search queries, reads results, then writes more queries if the picture is incomplete. Google's Gemini documentation says the model generates and runs one or multiple search queries, then processes the results into a grounded answer 2.
Effort scales with the question. Anthropic's documentation says simple factual questions typically use 1 to 3 searches, while comparative or multi-entity research can use 10 or more 3. OpenAI says its deep research models use actions such as search, open page and find in page, and can synthesise hundreds of sources into a cited report 4.
Bigger agent systems split the work. Anthropic describes a lead agent that starts 3 to 5 subagents in parallel, and it reports that such systems use about 15 times more tokens than ordinary chats 5. The practical lesson is that comparison questions, such as best provider for a given need, trigger the most reading.
Do AI agents actually visit your website?
Yes, and some do not behave like search crawlers. OpenAI says ChatGPT-User may visit a page when a user asks ChatGPT or a custom GPT a question, and that robots.txt rules may not apply because a user triggered the visit 6.
Perplexity documents a similar split. Its Perplexity-User fetches a page when a person asks a question and generally ignores robots.txt for that reason 7. Google's July 2026 guide adds that browser agents may access your site to gather the data they need to complete tasks 8.
So blocking by robots.txt is not a full control, and an agent will meet whatever your page shows at that moment. Pages that need scripts, logins or pop-ups to show prices may look empty.
What is agentic commerce?
Agentic commerce means an AI assistant helps a buyer complete a purchase, not just research it. Two live examples show the shape, and both are still early.
OpenAI's Agentic Commerce Protocol documentation says OpenAI is not the merchant of record. Merchants bring their own payment provider, ChatGPT shows the checkout and collects details, and the actual checkout and payment processing happen on the merchant's systems 9. Product feeds are how merchants share price and availability.
Google announced Universal Cart on 15 October 2026 for the US, in Search and the Gemini app. Checkout is powered by the Universal Commerce Protocol, shoppers can pay with Google Pay or finish on the merchant's site, and merchants remain the seller of record 10. Google's merchant page says UCP access is by waitlist and builds on existing Merchant Center feeds 11.
What does agentic search mean for being chosen?
An agent picks from what it can read, compare and verify. If your price, service area or return policy is unclear, the agent either skips you or guesses, and a guess can be wrong.
No source reviewed publishes how agents rank or choose products. The only documented input is accurate feed data such as price and availability 911.
Comparison queries are the main risk and opportunity. Because they trigger the most searching, they also give many chances for your pages to turn up, if those pages answer the sub-questions directly.
No published study measures how often agents buy from a given site, so any claim about conversion lifts from agents should be treated as unproven.
What should a site have ready for agents?
Prepare the basics now. None of these are agent-only features: they are things people also need.
| Need | Why an agent cares | Quick check |
|---|---|---|
| Facts in plain HTML text | Agents read text, and some do not run scripts | View the page with scripts off |
| Clear prices, ranges or 'from' prices | Comparisons need numbers to compare | Is a price or range on the page? |
| Specific service area and hours | Local and service picks depend on fit | Is it written, not only in an image? |
| Accurate product feed (shops) | Feeds drive shopping surfaces in ChatGPT and Google | Merchant Center and feed match your site |
| Policies on one page (returns, warranty) | Agents summarise risk before recommending | Linked from product and footer |
| Open access for user-triggered bots | User-initiated fetches may ignore robots.txt anyway | Check logs for ChatGPT-User and Perplexity-User |
What does an agent journey look like?
The example below is illustrative. It shows how one buyer prompt could become several searches and where a small business could lose or win a place.
- Prompt: Find a bookkeeper for a 5-person design studio in Leeds under a set monthly budget.
- Search 1: bookkeepers for small creative agencies Leeds
- Search 2: bookkeeper monthly fee small studio UK
- Search 3: reviews of the three firms found
- Reads: each firm's services page, pricing page and reviews
- Answer: shortlist of two firms with fees and one caveat each
- LOSS POINT: a firm with no public fee range is left off or described with a guess
- WIN POINT: a firm whose page states who it serves, its fee range and turnaround is quotable
What do people get wrong about agentic search?
One myth is that you must build a special agent interface now. Google's guide has a section on exploring agentic experiences and calls protocols such as UCP emerging 8. That is a watch-and-learn message, not a deadline.
Another myth is that agents will replace your site. Even in the ChatGPT flow, product feeds and merchant systems still handle the real order. Your site, data and policies remain the source of truth.
How do you apply this week?
Pick five comparison prompts a buyer would use, run them in an assistant, and read the sources it lists. Our query fan-out explorer shows the searches behind an answer, and query fan-out explains why they matter.
Then check access with the AI crawler checker and read how to check if AI bots can crawl your site. Shops should also read about AI search engines to see which assistants matter.
Frequently asked questions
Is agentic search the same as ChatGPT search?
Not exactly. ChatGPT search can answer with a quick web lookup. Agentic search is the deeper mode where a reasoning model keeps searching, opening pages and checking results until it is satisfied. Several assistants offer both, and the product names change often.
Will AI agents buy things without asking the user?
In the two launches reviewed here (ChatGPT checkout and Google's Universal Cart) the shopper completes the purchase. Google also describes other agentic features that act on a shopper's behalf, so check each platform's current terms.
Do I need to join OpenAI's or Google's commerce programmes?
Only if you sell products online and the programmes are open to you. Google's UCP access is by waitlist, and both build on product feeds. Start by keeping your feeds and site data accurate. Service businesses have no equivalent programme and should focus on clear pages.
Can I block AI agents from my site?
Blocking index crawlers such as OAI-SearchBot removes you from ChatGPT search answers, and user-triggered fetchers may ignore robots.txt (details above). Blocking is a trade-off, since agents that cannot read you cannot recommend you.
How do I know if agents are visiting my site?
Check your server logs for user agents such as ChatGPT-User, OAI-SearchBot, PerplexityBot and Perplexity-User, and use the providers' published IP lists to verify them. Logs show visits, not whether you were cited, so also run your key prompts monthly.
Sources
- Web search (tools guide) · OpenAIOpened October 20, 2026. Supports: Describes agentic search where a reasoning model can search as part of its chain of thought, analyse results and decide whether to keep searching.
- Grounding with Google Search · Google AI for DevelopersOpened October 20, 2026. Supports: The model decides whether search helps, generates and runs one or multiple queries, and synthesises a grounded response.
- Web search tool · Anthropic (platform.claude.com docs)Opened October 20, 2026. Supports: Simple factual queries typically use 1 to 3 searches; comparative or multi-entity research can use 10 or more. This is guidance for setting a cap, not a measured average.
- Deep research (API guide) · OpenAIOpened October 20, 2026. Supports: Deep research models are agentic, use search, open_page and find_in_page actions, and can synthesise hundreds of sources into a report with inline citations.
- How we built our multi-agent research system · Anthropic EngineeringOpened October 20, 2026. Supports: Lead agent spins up 3 to 5 subagents in parallel; multi-agent systems use about 15 times more tokens than chats (June 2025).
- Overview of OpenAI Crawlers · OpenAIOpened October 20, 2026. Supports: ChatGPT-User may visit a page when users ask a question; because the action is user-initiated, robots.txt rules may not apply; opting out of OAI-SearchBot removes sites from ChatGPT search answers.
- Perplexity Crawlers · PerplexityOpened October 20, 2026. Supports: Perplexity-User fetches pages when a user asks a question and generally ignores robots.txt because the fetch is user-requested.
- Optimizing your website for generative AI features on Google Search · Google Search CentralOpened October 20, 2026. Supports: AI agents are autonomous systems that perform tasks for people; browser agents may access your website; protocols like UCP are emerging; guide last updated 10 July 2026.
- Key concepts (Agentic Commerce Protocol) · OpenAI DevelopersOpened October 20, 2026. Supports: OpenAI is not the merchant of record; merchants bring their own payment provider; checkout and payment processing occur on the merchant's systems; product feeds drive display.
- An intelligent new shopping cart · Google (The Keyword blog)Opened October 20, 2026. Supports: Universal Cart launched 15 October 2026 in the US across Search and the Gemini app; checkout powered by UCP and Google Pay; merchants remain seller of record; presented as a tool for confidence and control.
- Universal Commerce Protocol (UCP) for merchants · Google for DevelopersOpened October 20, 2026. Supports: UCP enables agentic actions in AI Mode and Gemini starting with direct buying; access via waitlist; builds on existing Merchant Center feeds.