How AI Agents Are Changing Marketing: How to Stay Discoverable

1. Blog
2. Discoverability & Content

By Saroj Jha · March 16, 2026 · 11 min read

AI assistants now shortlist, compare and, in a growing number of US shopping flows, buy for your customers. What changed, where agents read your facts, which checkout standards matter, and a 30-day plan to stay in the running.

1. 01Ranked for humans
2. 02Retrieved by agents

Two different qualifying rounds.

Two qualifying rounds

Agents show three to five brands, not twenty links.

For twenty years the deal was simple: people searched, brands optimized, and Google ranked the results. That deal still covers a lot of buying. But a growing share of customers now ask an assistant instead, and since late 2025 the assistants have started doing more than answering. They compare prices, call stores, fill carts and, in a growing number of US shopping flows, check out.

The audience is large. OpenAI said in February 2026 that ChatGPT had 900 million weekly users^[1]. In July 2026 Google said AI Mode in Search had passed 1 billion monthly users and the Gemini app had 950 million^[2].

This post is about the part after the answer: agents acting for your customers, and what a small business has to publish so an agent can find it, compare it and buy from it. For how AI answer engines choose what to cite, read What Is GEO? What the Evidence Actually Says About AI Citations.

AI agents now shortlist, compare and, in more and more US shopping flows, buy for your customers. They work from facts they can read: product feeds, your Business Profile, your prices and reviews on other sites. To stay in the running, publish those facts, keep them the same everywhere and give an agent a path it can finish without a phone call.

What changed: assistants that act, not only answer

People haven’t stopped searching. What changed is that the assistant now does part of the shopping. Between September 2025 and October 2026 the largest platforms shipped the pieces an agent needs to act:

• Google started agentic checkout in US Search in November 2025: a shopper sets a budget, and when the price drops Google can buy the item on an eligible merchant’s site with Google Pay, after asking permission. The same month it began calling local stores to check stock and prices for shoppers^[5]. In January 2026 it released the Universal Commerce Protocol, an open standard co-developed with Shopify, Etsy, Wayfair, Target and Walmart, to run checkout in AI Mode and the Gemini app^[6].
• OpenAI published the Agentic Commerce Protocol, whose first payment implementation is Stripe’s, so ChatGPT can show a merchant’s products from a feed and pass orders to the merchant’s own systems^[7]. In March 2026 it said it would let merchants use their own checkout and focus ChatGPT on product discovery^[8],[9].
• Perplexity made shopping free for all US users in November 2025, with checkout inside the chat through PayPal^[10].

In each of these the retailer stays the merchant of record: the agent passes the order, and the business still takes the payment, ships and handles returns^[15],[16]. For a small business that is the part that matters. You don’t give up the sale; you give the agent the facts it needs to send it your way.

Shoppers are following. Adobe, which measures more than 1 trillion visits to US retail sites, found traffic from AI assistants up 693% year over year in the 2025 holiday season^[17]. In March 2026 those visitors converted 42% better than other traffic; a year earlier they converted 38% worse^[3]. The volume is still modest next to search, email and social, Adobe notes, but it no longer behaves like idle browsing^[18].

Three jobs an agent does, and what each one reads

A search engine ranks pages and leaves the choice to a person. An answer engine reads a few sources and writes one answer. An agent goes further: it builds a shortlist, compares the options on facts, then acts by buying, booking, calling or filling in a form. Each step reads different data, and you can drop out at any of them without ever seeing it happen.

• Shortlist. The agent has to know you exist and what category you’re in. It needs a clear category, one sentence on what you sell and to whom, and the same story on your site, your Business Profile, marketplaces and review sites. If a person can’t explain your business in one sentence, an agent won’t either.
• Compare. The agent ranks the shortlist on the facts the customer asked about: price, availability, delivery time, returns, specifications, location and hours. A price it can’t read is a price it can’t compare, and the competitor whose price it can read takes the slot. Why doesn’t ChatGPT recommend my company? works through this gap on a real audit.
• Act. The agent needs a step it can complete: a checkout connected to its platform, a booking page, a trial or a form that works without a call. If the only next step is a sales call, the agent can tell the customer to phone you, but it can’t finish the job, and a competitor’s checkout can.

Reputation runs through all three. In a G2 survey of 1,076 B2B software buyers in March 2026, 51% said they now start research in an AI chatbot more often than in Google, citations from software review sites were the signal that most raised their confidence in an AI answer, and 69% had chosen a different vendor from the one they planned because a chatbot recommended it^[4]. G2 sells review and visibility services, so read these as a vendor’s numbers, but they point the same way as Adobe’s.

Do this

Ask three assistants (ChatGPT, Gemini and Perplexity) the question your best customer would ask, with their real constraints: budget, location, delivery date. Note whether you make the shortlist, which facts the answer compares on, and where your facts are missing or wrong. That list is your backlog.

Where agents read your facts

Agents don’t read your brand story. They read a handful of sources, and most of them are feeds and profiles you can fix in an afternoon.

• Product feeds. ChatGPT’s shopping results come from merchant feeds: a regularly refreshed file with identifiers, descriptions, price, availability, images and fulfillment options, which OpenAI recommends sending in full once a day with updates in between. Feed onboarding is open to approved partners, who apply at chatgpt.com/merchants^[19]. Google’s guidance says Merchant Center feeds and Business Profiles can help products and services appear in its AI responses^[20], and in January 2026 it added dozens of feed attributes for conversational shopping, such as answers to common product questions, compatible accessories and substitutes^[6].
• Your Business Profile. When a US shopper asks Google to call nearby stores, an automated call that identifies itself as Google asks about stock, price and offers, and Google requires merchants to keep an accurate phone number on their Business Profile^[21]. Hours, address, phone and services are the facts a local agent checks first.
• Your own pages. An agent that browses reads what loads. Adobe scored US retail sites with its own checker and found product pages only 66% readable by machines on average, the lowest of the page types it tested^[3]. Prices, specifications and delivery terms drawn in by JavaScript are the usual gap.
• Your commerce and payment platforms. Many small stores reach agents without a custom build. PayPal’s store sync made PayPal merchants’ catalogs discoverable in Perplexity^[15], and Shopify’s catalog supplies product data to ChatGPT^[9].
• Other people’s pages. Reviews, marketplace listings, directories and press are where buyers, and the agents working for them, check what you say about yourself. Consistency matters more than volume.

Running case: Example Co. Fictional

Example Co. is a nine-person US online shop selling coffee grinders and espresso machines. It has 240 products, good reviews on its own site and a Merchant Center feed set up years ago. When its founder asked ChatGPT, Gemini and Perplexity for “a quiet burr grinder under $300 that arrives by Friday”, none named Example Co.; two named a competitor whose feed listed delivery times. Example Co.’s feed had no shipping or returns data, its prices loaded by JavaScript, its Business Profile showed an old phone number and its reviews lived only on its own pages. All details are illustrative.

Do this

List the eight to ten facts customers compare you on: price, stock, delivery, returns, hours, location, key specifications, who it’s for. Check each one in four places: your product feed, your Business Profile, your own page with JavaScript switched off, and your main review or marketplace listing. Fix every mismatch before you add anything new.

Give an agent a path it can finish

Being compared is half the job. The other half is the next step: if an agent can’t complete the purchase or booking, the sale goes to someone whose path it can complete. For a shop that now means one of three checkout standards; for a service business it means a booking, trial or price an agent can act on.

• OpenAI’s Agentic Commerce Protocol has three parts: a product feed, checkout endpoints the merchant runs, and a delegated payment, in which a one-time token with a maximum amount passes through the merchant’s own payment provider. Stripe’s Shared Payment Token was the first compatible implementation, and OpenAI is not the merchant of record^[7].
• Google’s Universal Commerce Protocol has a business publish a manifest at /.well-known/ucp listing what it supports, such as checkout, discounts and fulfillment, and which payment handlers it accepts; Google’s own checkout also needs an active Merchant Center account^[22]. It works with Google’s Agent Payments Protocol, which records what a shopper authorized an agent to buy in signed “mandates”^[11]. On October 15, 2026 Google launched Universal Cart in the US, one cart across retailers in Search and Gemini with checkout through the protocol^[14]. Canada and Australia come next, then the UK^[16].
• Perplexity runs checkout through PayPal. Merchants join through PayPal’s store sync and “agent ready” services, using their existing PayPal setup^[15].

You don’t need to build any of these yourself. Most small merchants will reach them through Shopify, PayPal, Stripe or their commerce platform, and the standards are still moving: OpenAI’s March 2026 step back from its own in-chat checkout shows the plumbing will change again. The work that lasts sits upstream of it: clean feeds, published prices and a checkout on your own site that works for a stranger.

For service and B2B businesses the same logic applies without a protocol. An agent asked to book a plumber or shortlist bookkeeping software can act on a public price, a booking link, a free trial or a form it can submit. It can’t act on “call for a quote”. The Agentic AI Marketing Playbook covers the B2B side.

Do this

Try to buy or book from yourself through an assistant: ask it to find your product or service and take the next step. Where it stops, find out why: no feed, no price, no checkout it can reach, a form it can’t submit. If you sell online, check which agent channels your commerce and payment platforms already offer, and switch on the ones that keep you as merchant of record.

Free tool

Check whether an agent can buy from you

The free Agent Buyability Check runs fifteen checks across five weighted dimensions: whether an AI agent acting for a buyer can find you, compare you with alternatives and act.

What to skip, and what is still unsettled

The agent shift attracts the same overselling SEO did. Three things are worth knowing before you spend money.

• llms.txt is optional. It is a 2024 proposal by Jeremy Howard for a plain-text guide to a site for language models^[23]. Google says Google Search ignores such files, so publishing one neither helps nor harms you there^[20]. The proposal’s own site lists AI companies that publish an llms.txt for their developer documentation, not engines that read one to decide what to recommend. It takes ten minutes; it isn’t worth a retainer.
• There is no special AI schema. Google says structured data isn’t required for its generative AI features and there is no special schema.org markup to add, though it still recommends structured data for rich results^[20]. For shopping agents, a complete product feed does more than extra markup.
• The rules aren’t settled. Amazon sued Perplexity in November 2025 to stop its Comet browser’s assistant from shopping on Amazon for users, and won a preliminary injunction in March 2026^[12]. On August 4, 2026 the Ninth Circuit vacated it, finding that the user, not Perplexity, accessed Amazon’s systems, and said its ruling doesn’t set rules for agentic AI in general^[13]. Large platforms can still try to keep agents out, and the standards above are young.

Measurement is the other open question. AI referrals are still a small share of visits, and an agent that checks out inside an assistant may never send a visit at all. Stop chasing the click: how startups can win with zero-click marketing covers measuring what never clicks.

Do this

Set AI assistants up as their own referral channel in your analytics, ask every new customer where they first heard of you, and turn on Merchant Center’s AI performance insights when they reach your market. Spend nothing on llms.txt or “AI schema” packages until your feeds and prices are right.

A 30-day plan

Four weeks is enough to get the facts right and test the path. Here is the plan, with what Example Co. did each week.

Running case: Example Co. Fictional

In the repeat test after four weeks, Gemini named Example Co. for the grinder question and Perplexity linked its product page; ChatGPT still didn’t, because its merchant application was pending. It doesn’t yet know how many sales that brings, so it now tracks AI referrals as a separate channel to find out.

Do this

Put the four weeks in your calendar with one owner each, and re-run the same three-assistant test every month so you can see whether anything moved.

What hasn’t changed

Agents haven’t replaced marketing. They’ve moved part of the decision into a conversation you can’t see, run on facts you publish. The fundamentals still decide who wins: a clear position, honest proof and an offer that’s easy to say yes to. What’s new is that a machine checks them first, and it doesn’t fill gaps with goodwill.