Agentic search: how AI agents browse the web on behalf of buyers

AI agents as buyers

AI agents take over a person's research to build a shortlist and sometimes make the purchase too. ICSC and McKinsey surveyed 3,004 US consumers and estimate that agentic commerce in US business-to-consumer retail could reach $1 trillion by 2030. The same survey found that 68% of consumers had used at least one AI tool while shopping in the previous three months.

Adobe Analytics recorded that visits to US retail sites from generative AI tools rose 693% year over year over the 2025 holiday season. The real change is that the agent filters options before the human shows up. If your brand drops off the shortlist, nobody on the buying side ever weighs it.

How much people hand over depends on the purchase. A B2B buyer may let an agent compile a vendor comparison and still run the demo calls in person. A consumer reordering coffee filters may let it check out. So start by listing which of your buyers' tasks they could hand to an agent, from research through to payment.

How agentic search works

The agent begins its search loop by splitting the goal into sub-questions and running a search for each one. Then it reads the results and checks for gaps. If it finds a gap, it searches again before it answers or acts. Google described this step at its I/O 2025 launch of AI Mode, which it said works by "breaking down your question into subtopics" and issuing a multitude of queries at the same time. Google calls the technique "query fan-out."

The depth of the loop depends on the mode. In a Seer Interactive test of 501 prompts, Google's model produced 10.7 fan-out queries per prompt on average, with a range of 3 to 28. Deep research modes go much further. In an Ahrefs test, ChatGPT Deep Research ran 420 searches for a single shopping question.

For a brand, the buyer's original question is the start of this search process. The agent can find you through a sub-question about pricing tiers or an integration. It can also find you through a "versus" comparison the buyer never typed. Check whether your site answers those narrower questions in plain text, because the agent judges you on them.

What agents do on websites

On a website, an agent reads pages and follows links. It can also fill in forms and compare prices. With permission, it can log in and buy. OpenAI's ChatGPT agent, launched on July 17, 2025, has two browsers: a visual one that clicks through the interface the way a person would, and a text one for quick reads. OpenAI's launch post says the model "can choose to open a page using the text browser or visual browser". It can also download a file and run commands on it in a terminal.

Google showed the transactional side at the same I/O event. When asked for baseball tickets, AI Mode would look across sites to analyze "hundreds of potential ticket options" with live pricing and inventory, then fill in the forms itself.

An agent that picks a text fetch sees only the raw page. A visual browser sees what a person sees, but it gets stuck on cookie walls and CAPTCHAs. Prices hidden behind a "contact sales" button also stop it. Give an agent your most common buyer task on your own site and watch where it stalls.

Agentic search vs RAG

Retrieval-augmented generation (RAG) retrieves documents once and writes an answer from them. Agentic search repeats retrieval as often as it needs to, and it can act on what it finds. RAG is a single pass. An agent runs many passes, and each one can use RAG.

Under RAG, a single retrieval decides whether your page is used. Under agentic search, you get more chances to appear, because a later sub-query can surface you. You also face more tests, because the agent compares you against rivals at each step and can drop you at any of them.

Agentic SEO fundamentals

Agentic SEO means you make your pages findable for the sub-questions agents ask and usable once an agent arrives. Ranking still matters, because agents send their queries to search indexes. Surfer studied 10,000 keywords and found that pages ranking for both the main query and its fan-out queries were 161% more likely to be cited in Google's AI Overviews than pages ranking for the main query alone.

In practice, you write for the comparison step. Agents look for facts they can check against other sources. They need prices and limits, along with details about which integrations the product supports. They also need to know delivery times and who the product is for. A page that says "flexible pricing for every team" gives the agent nothing to compare, so it moves on to a rival that lists $12 per seat per month.

Consistency matters as much as coverage. If your pricing page says one thing and a two-year-old blog post says another, the agent trusts neither because it can't tell which is true. One fact should sit in one canonical place, and every other page should repeat it exactly. Before you publish, ask whether an agent could pull a specific, checkable answer from the page in a single read.

Making your site agent-ready

To be agent-ready, your site needs its key content in the HTML the server sends. It also has to let in the agents you want and describe products in structured data. A visitor must be able to finish a task without human help.

Put key content in the initial HTML

Most AI crawlers don't run JavaScript. Vercel and MERJ analyzed crawler traffic across Vercel's network in December 2024 and found that ChatGPT's crawlers requested JavaScript files on 11.50% of requests but never executed them. Anthropic's ClaudeBot behaved the same way. Gemini and AppleBot were the exceptions in that study. If your prices or specs load in the browser after the page arrives, these agents see an empty page. Render critical content on the server.

Decide which AI bots get in

AI companies run separate bots for separate jobs, so you control access per bot in robots.txt. OpenAI documents four separate crawlers. GPTBot handles training, while OAI-SearchBot handles search. ChatGPT-User fetches pages live when someone asks a question. OAI-AdsBot is the fourth crawler. You can block training and still allow the search and user agents that bring buyers. Blocking everything by default removes you from agentic results too.

Mark up products and offers

Schema.org Product and Offer markup tells an agent what a number means by separating a price from a discount. It also marks stock status and shipping terms. Adobe reported in April 2026 that a large share of US retail sites lag in AI visibility because their pages aren't machine readable.

Keep tasks completable

Agents fail where people would hesitate. Label every form field and write error messages in text. Show prices before login. A demo request that needs four screens and a pop-up is a task an agent abandons.

Agentic commerce signals

Agentic commerce signals give an agent machine-readable facts about price and availability before it buys. The agent also needs delivery and returns information, along with a checkout it can use. In September 2025, OpenAI and Stripe released the Agentic Commerce Protocol. It powers Instant Checkout in ChatGPT, which began with US Etsy sellers, and Stripe said over a million Shopify merchants would follow. Stripe CEO Patrick Collison wrote at the launch: "It's clear that internet purchasing modalities are going to change a lot."

These signals matter because traffic from agents now converts. Adobe found that AI-referred visits to US retail sites converted 42% better than other traffic in March 2026, a year after they had converted 38% worse.

How an agent reaches you varies. Some merchants sell through a protocol and a product feed. Others rely on agents browsing their checkout. Either way, the price and stock shown in the feed have to match those on the page. The cart has to carry the same values. When they don't, the agent records a mismatch and picks a seller whose data lines up.

How AI agents cite brands

AI agents cite the brands they find mentioned repeatedly across sources, and they link the pages that supplied a specific fact. Ahrefs studied 75,000 brands and found that branded web mentions had a 0.664 correlation with visibility in Google's AI Overviews, against 0.218 for backlinks. In the same study, brands in the top quarter for web mentions averaged 169 AI Overview mentions. Brands in the next quarter averaged 14.

So what other people write about you counts for more than what you publish yourself. Reviews and comparison articles show an agent that independent sources know your brand. Forum threads and analyst lists do the same. Your own pages earn the citation link when they hold the exact figure the agent needed.

A citation doesn't guarantee a visit. Cloudflare measured that Anthropic's crawlers fetched 38,000 pages per referral in July 2025, while OpenAI's fetched 1,091. Search for your category in an assistant and note which third-party pages it cites, because those are the pages that shape its picture of you.

Measuring agentic visibility

You measure agentic visibility at two levels: what agents say about your brand, and what traffic they send you. To track what agents say, write a fixed set of prompts your buyers would hand an agent. Run them in each assistant every month. Then record whether your brand appears and where it ranks on the shortlist. Note which page gets cited.

For traffic, OpenAI says ChatGPT adds utm_source=chatgpt.com to referral links, so those visits show up as their own source in analytics. That count misses visits because the parameter isn't always added, as Seer Interactive found in June 2025. Cloudflare also notes that native apps such as Claude's send no Referer header, so their visits land in "direct."

Live-fetch agents such as ChatGPT-User identify themselves in server logs, so a spike on your pricing page means agents are checking it for buyers. Compare those fetches with your prompt results. Then you'll know whether agents read you and still pass you over.

Next steps

Pick the ten prompts your buyers are most likely to give an agent and run them this week. Then fetch your three most important pages with JavaScript turned off. Those two checks show the gap between what you think your site says and what agentic search actually reads. If prices or specs vanish without JavaScript, fix that first, because no amount of content or mentions helps an agent that can't see the page.

Snoika audits websites and tracks brand visibility in Google and AI assistants, and its team reviews the results. To see how agents read your pages today, run a Snoika website audit.