The new ecosystem: a fourth player between brand and buyer
For twenty years, ads were built to catch a human eye. In agentic shopping, the first "viewer" of your ad is software, and software does not care how pretty your banner is.
Emily is training for a half marathon. On a Sunday night she types into her AI assistant: "Find me trail running shoes under $150. Good grip in the rain. Delivered before Saturday." In the old world, Emily would open a search engine, scan sponsored links, click five product pages, read reviews, compare prices in three tabs, and maybe give up. Every one of those steps was an ad opportunity. In the new world, her agent does all of that in a few seconds. It checks catalogs, compares prices, confirms stock and delivery dates, and comes back with two or three picks. Emily may only see the final answer. She may even tell the agent: "Just buy the best one."
Now think about a shoe brand. Let's call it Stickler — a made-up brand. Stickler has great shoes, a strong ad budget, and beautiful creative. The question for Stickler is simple and scary: how do we get picked when a machine does the shopping? That is what this article is about.
The classic ad world had three main players: the advertiser (Stickler), the publisher (a search engine, a website, an app), and the human who sees the ad. A lot of ad tech sat in between: DSPs, SSPs, exchanges, measurement vendors. Agentic shopping adds a fourth player that changes everything: the agent. It sits between Emily and every brand. It reads, compares and decides before Emily sees anything.
The new ecosystem has five layers:
- The shopper's agent. ChatGPT, Gemini and Google AI Mode, Amazon's Alexa for Shopping, Perplexity, and many smaller assistants. This is where Emily's request starts.
- The brand's data layer. Product catalogs, live price, live stock, promotions, reviews. This is what the agent reads to decide. Some brands will also run their own brand agent that talks to the shopper's agent.
- The ad layer. Paid placements inside the agent's answer, or next to it. ChatGPT now shows sponsored listings below answers, Google shows ads inside AI Mode, and Amazon runs sponsored prompts in its shopping assistant.
- The checkout and payment layer. New protocols let an agent place an order and pay safely, using a token instead of Emily's real card number.
- The measurement layer. Pixels, Conversions APIs and attribution partners that tell Stickler whether the ad actually led to a sale.
The big shift: the ad's first job is no longer to catch Emily's attention. It is to be understood, trusted and chosen by the agent, or to show up right next to the agent's answer at the moment of decision.
This is not a future idea. Amazon folded its Rufus assistant into Alexa for Shopping in May 2026, and its "Buy for Me" feature can buy from other brands' websites on the shopper's behalf. Google's Universal Commerce Protocol already powers checkout inside AI Mode and Gemini. OpenAI began testing ads in ChatGPT in February 2026 and has since expanded to more countries.
How brands get discovered by agents
Emily's agent can only pick Stickler if it can find Stickler, understand the shoe, and trust the facts. That comes down to four things.
A product catalog written for machines, not just for people. A human reads "Conquer any trail!" and feels something. An agent reads it and learns nothing. Emily asked for three things: grip in the rain, under $150, delivered by Saturday. If Stickler's catalog only says "premium outsole," the agent cannot match it to "grip in the rain." If it says "Vibram Megagrip outsole, rated for wet rock, 4 mm lugs," the agent can.
What a good agent-ready catalog looks like:
- Specific attributes, not adjectives. Weight, drop, lug depth, waterproof or not, sizes, widths.
- Use cases in plain words. "Good for: wet trails, half marathons, wide feet." Agents match intent to these words.
- Clean structure. Standard fields (schema.org markup on the website, complete product feeds to Google Merchant Center and other platforms), not facts hidden inside images or PDFs.
- Honest comparisons and FAQs. "How does this compare to our Model X?" Agents love direct answers.
- Reviews the agent can read. Ratings, review counts and common themes ("runs small") help the agent judge fit.
MCP and agent-facing APIs. A product feed is a snapshot. It can be hours old. Agents increasingly want to ask a brand's systems directly. MCP (Model Context Protocol) is an open standard that lets an AI agent call tools and fetch data from outside systems in a common way. Think of it as a universal plug. If Stickler exposes an MCP server with tools like search_products, check_stock, get_price and get_delivery_estimate, any MCP-aware agent can ask: "Do you have size 9 wide in stock near Emily's ZIP code?" and get a real answer. Google's Universal Commerce Protocol takes a similar idea further for checkout. A business publishes what it supports in a standard file on its own website, so agents can discover its capabilities without a custom integration.
Live price, promotions and inventory. This is where most brands will lose without knowing it. If the agent sees a stale price, it may drop Stickler as "over budget." If it cannot confirm stock or delivery by Saturday, it will pick a brand that can. Google has said as much: in AI Mode, structured product data and real-time price and availability feeds matter more than ever. Promotions also become machine-readable. Google's Direct Offers pilot lets a brand show an exclusive deal, like 20% off, right inside AI Mode when a shopper is ready to buy.
Trust signals. Agents will be careful about who they send their users to. Return policy, shipping speed, warranty, fraud history and review quality all become ranking signals. A brand that is easy to buy from and rarely gets complaints becomes the "safe pick."
Back to Emily: Stickler wins the first round only if its catalog says "wet grip," its live price shows $139, its stock API confirms size 9 wide, and it can promise Thursday delivery. No creative was involved yet.
Do ad creatives still matter?
Short answer: creatives matter much less to the agent, but they still matter to the human. They are not dead. Their job moves.
Why the agent does not care about your banner. Stickler spent $80,000 on a video of a runner splashing through a muddy trail at sunrise. It is beautiful. Emily's agent will never "watch" it in any useful way. The agent wants facts it can compare: grip rating, price, weight, delivery date, return window.
To the agent, the best "creative" is clean, complete, structured data. A spec sheet beats a slogan. This is already happening. For Google's newest AI Mode ad formats, advertisers do not build the ads themselves. Google generates them from the product information the brand submits. Your feed is your creative.
Why creatives are not fully irrelevant. There are still three moments where a human decides:
- Before the agent. If Emily has seen Stickler's ads on YouTube or Instagram for months, she might type "Find me Stickler trail shoes" instead of "find me trail shoes." Brand advertising shapes the question the agent receives. That is the most valuable spot of all.
- At the final pick. The agent shows Emily two or three options. Now a human looks at an image, a short line of text, a price. The product photo and the one-line pitch still matter here.
- After the purchase. Unboxing, emails, loyalty. Still human.
What changes for creative teams:
- From "one hero ad" to many short, factual variations an AI system can mix and match.
- From slogans to claims you can prove ("tested on wet granite," "30-day returns").
- Images that show the product clearly, because agents and AI ad systems may pick and crop them.
- Brand building moves upstream, to make shoppers name your brand before the agent starts.
The protocols: who is building the plumbing
Yes, there are protocols, and quite a few. They fall into two groups: commerce protocols (how an agent finds a product and buys it) and advertising protocols (how agents buy and sell ads). Ad people tend to watch only the second group. The first group matters just as much, because it decides who controls the purchase.
Two things stand out.
- First, commerce protocols are moving faster than ad protocols. Google, OpenAI and Stripe already run real purchases. The ad side is still debating standards.
- Second, there is a quiet split in the ad world. AdCP is a new, MCP-based approach led by supply-side companies. IAB Tech Lab prefers to extend the standards the industry already uses. Something similar happened in the early days of programmatic before OpenRTB won. Expect a few years of overlap.
Back to Emily: when her agent buys the Stickler shoe, the order might go through UCP or ACP, the payment through a Visa or Mastercard agent token, and the ad that put Stickler in front of her might have been bought by a Stickler agent talking to a seller agent over AdCP or AAMP.
What changes for data and privacy
A lot changes, and not all of it is obvious.
- The agent knows more than any cookie ever did. Emily told her agent she is training for a half marathon, runs in the rain, and needs shoes by Saturday. That is richer intent than any search keyword or tracking cookie. The big question is how much of that context advertisers get to use.
- For now, platforms are keeping conversations away from advertisers. OpenAI says ads do not influence ChatGPT's answers, and that conversations stay private from advertisers. The platform uses the context to pick the ad, but Stickler never sees what Emily typed. This is the walled garden model, made even tighter.
- Privacy laws still apply, and some are stricter. In Europe, personalized ads in ChatGPT are reported to be limited to users who opt in. Expect regulators to treat chat history as sensitive data, especially when it touches health, money or children.
- Payments get safer, not riskier. Agents pay with tokens and spending limits the user sets, not raw card numbers. Under Google's UCP, the merchant stays the merchant of record and keeps the customer data. That matters to brands that fear losing the customer relationship.
- First-party data becomes even more valuable. If Stickler cannot see Emily's chat, its own data (past orders, loyalty, email) is how it personalizes after the sale. Expect more clean room deals between brands and agent platforms, the same way retail media works today.
- A new identity problem: is this a good bot or a bad bot? For twenty years, ad tech worked hard to block bots. Now some bots are your best customers. Brands and publishers need to tell a real shopping agent (acting for a real person, with a verified token) from a scraper or a fraud bot. This is why Mastercard requires agents to be verified before they can pay, and why IAB Tech Lab is building an agent registry and authentication into AAMP.
Bidding, winning and measuring an ad in an agent
How an ad gets into Emily's answer. There are really two separate decisions happening when Emily asks her question. Mixing them up is the most common mistake in agentic ads talk.
- The organic answer. The agent picks the best shoes for Emily based on her needs. The platforms say money does not touch this part. OpenAI's stated policy is that ads do not influence ChatGPT's answers. You win here with great data, great products and trust (section 2).
- The sponsored slot. Next to or below that answer, there is a paid placement, clearly labeled "Sponsored." You win here through an auction, much like search ads.
The auction looks familiar. Stickler sets a budget and a bid, usually cost per click today. OpenAI's self-serve Ads Manager launched with CPC bidding, a pixel and a Conversions API, and later added conversion-based bidding. Amazon lets brands bid on specific shopper questions with Sponsored Prompts. Google's Direct Offers let brands bid with a discount instead of only with money.
The winner is rarely just the highest bid. As in search, platforms combine bid x predicted relevance x quality. The difference is that "relevance" is now judged against a whole conversation, not a three-word keyword. Google says its AI Mode ads are matched not just to the question, but to the AI's answer itself.
Measuring what happened. Good news: measurement is not starting from zero. It reuses the tools ad tech already built after cookies broke.
- Pixel plus Conversions API. ChatGPT Ads uses both. Its pixel stores a click reference in a first-party cookie, and advertisers can send it back with server-side conversion events. This is the same pattern as Meta's and Google's Conversions APIs.
- Measurement partners. OpenAI works with about 20 partners, including incrementality firms, and supports click-through windows plus a one-day view-through window.
- Retail media style reporting. Amazon is adding a dedicated reporting tab for prompt ads, with impressions, clicks and visits.
The hard part is new: when the agent buys for Emily, there may be no click at all. The agent reads, decides and checks out through UCP or ACP. Stickler needs the platform to report "agent-assisted purchases," not just clicks. Expect new metrics like share of agent recommendations and agent-completed orders.
Why was this ad picked? Cost or context? Today's honest answer: both, and you mostly cannot see the details. Emily gets a "Why am I seeing this ad?" link. Stickler gets campaign reports. Neither gets a full explanation of the decision.
What advertisers should push platforms for:
- Selection reasons. "Shown because: matched 'wet grip', price under user budget, in stock, delivery by Saturday."
- Loss reasons. "Not shown because: price above stated budget" or "delivery date too late." This is the most useful feedback a brand could get, because it tells you what to fix in your catalog.
- Organic share. How often the agent recommended Stickler without any ad. This is the agentic version of SEO.
- Clear separation. Proof that paid and organic are kept apart, so trust is not damaged.
Don't agents already log this? Partly. The ad auction logs bid, predicted relevance and quality, just like any ad server, so reason codes for the paid slot are easy to share. The organic pick is harder. The agent's reasoning trace is logged, but it does not always show the true cause, because part of the choice happens inside the model itself. Platforms also hold back on purpose: Emily's chat is private, raw logs would expose rivals' bids, and brands would game the signals the way they game SEO. The likely outcome is aggregated, privacy-safe reason codes, not raw logs. This is a big open gap, and an opportunity for measurement companies.
Does every product fit agentic ads?
No. Agentic ads work best where the decision can be written as a list of needs, like Emily's shoes. They work worst where the purchase is about feelings, taste or a story.
The ad formats and product categories best suited to this model fall into clear bands. Comparative purchases — shoes, electronics, appliances — benefit from the structured-data approach. Purchases driven by aesthetic preference, cultural context or subjective experience resist machine comparison. A bottle of perfume, a piece of art, a vacation package: these categories lack the attribute-level clarity that agents need to weigh options. Brands in those spaces will need to decide whether they invest in making their offering machine-readable or rely on the human-decision moments that still exist before and after the agent's turn.