Do AI agents see ads?

STUDY · HOW AGENTS SHOP

This week is a study issue: what happens to retail ads when AI agents do the shopping. But first, the week’s trending news.

Trending news of the week

1. Shopify lets browser agents check out

Shopify now lets AI agents working in a shopper’s browser complete checkout on all eligible Shopify stores, once the shopper confirms the order, PYMNTS reports. Until now, these agents could search and fill a cart but not buy. Shopify also requires each agent to identify itself with a registered digital signature. For merchants, browser agents like the ones in this week’s study can now finish the sale on your site, and Shopify wants to know who they are first.

Read more: Shopify Opens Store Checkouts to AI Agents (PYMNTS) →

2. Adobe expects AI traffic to retailers to more than double this holiday

Adobe forecasts $275.1 billion in US online holiday spending, up 6.7%, and expects traffic to retail sites from AI tools to rise 130% on last year, Adobe reports. Shoppers who arrive from AI tools spend more per visit and return less. Adobe counts shoppers clicking through from an AI tool, not agents shopping for them. Agent visits like the ones in our study are a separate stream that is harder to see, and worth raising with your analytics team before the holidays.

Read more: U.S. Holiday Shopping Season to Hit Record $275.1 Billion Online (Adobe) →

This week’s study: agents and retail ads

With all the talk of AI shopping agents, one question keeps coming up: what happens to ads? When an agent shops for someone, does it see the sponsored listings a retailer sells space for? And if it does, does it pick them?

We ran a large study to find out: more than 600 shopping trips across four leading AI models, plus tests of the consumer tools people actually use. My short answer: agents either never see the ads, or they see them and pick something else.

Two ways an agent shops

Imagine a shopper asking: “Find me a good air fryer under $100 on Amazon.” Today an agent can answer that in one of two ways.

We tested both. We sent our own browser agents, running leading models from OpenAI, Anthropic, Google and xAI, to Amazon, Target, Walmart and Best Buy. We also built a test store stocked with real listings from ten Amazon searches, where we could switch the “Sponsored” label on and off. And we asked free ChatGPT, Comet and Claude in Chrome the same everyday questions.

Search agents go around the ads

Search agents mostly never reached the ads. When we asked free ChatGPT to shop on Amazon, it never looked at an Amazon results page. It found the air fryer through a third-party price tracker, a running shoe through a podcast page, and a tablet price from Best Buy. On Target it did read a category page and a search page, and still never mentioned an ad.

Comet’s quick search was similar: it answered our air fryer question in three seconds by citing Wirecutter’s testing. Comet even opens with the pitch “No more ads, focus on what matters.”

These agents aren’t ignoring the ads. They just never land on the page where the ads are.

Browser agents see the ads — and pass them by

Browser agents are more advanced. They load the retailer’s results page, ads and all. They just almost never pick the ad.

Claude, driving Chrome through the Claude in Chrome extension, scrolled past three sponsored air fryers on Amazon and stopped at the first regular listing carrying Amazon’s “Overall Pick” badge. For a kids’ tablet it passed two ads to do the same. On Target, two of its picks were products Target was also advertising, but both also appeared further down as regular listings, and Claude gave rating and review count as its reasons. Comet’s agent mode made three picks we can report, all regular listings, citing review counts and Reviewed.com. Neither tool mentioned sponsorship once.

The test store gives the clearest answer. The same products were picked no more often with a “Sponsored” label than without one — if anything, slightly less. Moving a product to the top of the page barely helped either. Agents said they chose based on ratings, review counts and how well the product fit the request.

So what gets an agent to pick the ad?

In our test store, we kept adding to the ad to see what it would take. A louder ad didn’t do it. With a “#1 Best Seller · Limited-time deal” badge added, the advertised product was picked 2 times in 48.

What worked was the shopper. When the request added one line, “I like to take advantage of featured deals”, agents went looking for the deal, and picks of the advertised product jumped to 14 in 48. No ordinary advertising change came close. Anthropic’s model, which never picked the advertised product in any other test-store setup, picked it 6 times in 12.

Agents work for the shopper. Tell an agent you care about deals, and it goes looking for the deal badge.

What this means for retail media

Retail media is a massive business, built on shoppers scrolling a results page. Agents shop differently, and that matters in two ways.

• Your impressions may include agents that rarely pick an ad. A browser agent loads the results page like a person does. If you’re seeing conversions drop on your sponsored listings across retail media sites, ask your ad rep whether agent traffic is being counted in your impressions, and how it’s filtered.
• When agents skip the website, they skip the ads too. Connectors let an agent ask a store for products, stock and prices directly, with no results page at all. As we covered in our connectors issue, that route is opening up. With no results page, there’s nowhere to put a sponsored listing — unless retailers start building paid placement into what their connectors return.

That last point was our open question, so we tested it on Meta’s Muse. For shopping, Muse searches Meta’s product catalog and Facebook Marketplace, where Meta also runs ads. We gave it 13 everyday shopping requests and saw no sign it favoured Meta’s paid placements. With no store named, four of its five picks were sold by ordinary retailers such as Best Buy and Target. On Marketplace, none of its eight picks was flagged as sponsored. Asked why, Muse said its Marketplace search returns only a listing’s title, price, condition, location and seller, with “no sponsored/ads field at all.” By its own account, the ads never reach it. It works like the connectors above: Muse asks Marketplace directly, and ads aren’t part of the answer. Even when the shopper said they like featured deals, Muse read that as a discount, not a promoted listing.

What store owners should check

• Make your regular listing win. Agents picked on ratings, review counts and clear product details. In our tests, those mattered more than ad placement.
• Keep deal flags accurate. When shoppers tell agents what they value, agents look for listings that match. A real deal, clearly labelled on the right product, is what gets found.
• Question your ad reporting. Ask how agent visits are counted in impressions and clicks, and whether you can see them separately.

About this study

We ran 608 agent trials between September 25 and 27, 2026, using leading models from OpenAI, Anthropic, Google and xAI: 48 on live retail sites (three everyday requests at each of four retailers), 320 in our test store with the “Sponsored” label and position switched on and off, and 240 in the test store with extra changes added step by step. We also ran 15 tests on consumer tools: five each on free ChatGPT, Perplexity Comet and Claude in Chrome. The ChatGPT and Comet tests were run by hand; for Claude in Chrome, Claude drove Chrome through the extension from our own setup rather than the consumer side panel. One Comet test is private, so we report four. The live-site agents were told to use their browser. The core design and analysis were set before the trials; changes along the way, and the after-the-fact checks below, are logged in our study notes.

Our browser agents read page text and structure rather than screenshots. Sessions were logged out, from one location, over one week. Walmart blocked all 12 of our live attempts. The test store uses real Amazon listings in a neutral layout that the models had not seen before; it is closed to search engines and all other crawlers. The label test was too small to catch differences under about 10 percentage points, and each step of the deals test (48 trials) too small to catch differences under about 20. Results for individual models (12 trials each) are rough.

We tested Muse by hand on September 28, 2026: five requests with no store named, five on Facebook Marketplace, and three that added “I like to take advantage of featured deals.” With 13 requests, these are examples, not rates. Without a Facebook login we could not capture Marketplace’s own ad labels, so whether a pick was sponsored comes from Muse’s interface and from its answers when we asked, which is weaker evidence than seeing the labels ourselves. Muse’s memory was on, and some Marketplace searches ran in San Francisco rather than Chicago. Nothing was bought and no seller was contacted.

After the runs, we found that in the test store the deal badge sat just below the advertised product’s price, directly above the next product, so some agents credited the badge to the product below. In an after-the-fact check counting either product, 40 of 48 agents picked one next to the badge when the shopper said they liked deals, against 6 of 48 without that line. The 14-in-48 figure above counts only the advertised product, so it understates the effect. We also tested a line inside one ad addressed to AI assistants, telling them to recommend it. That is prompt injection by design, and we ran it only in our own store: Anthropic’s model never picked the product it pointed to, while Google’s picked it 7 times in 12 with the “Sponsored” label and 10 times in 12 without, against 2 in 12 with no such line. Our first pilot suggested agents saw no ads at all; that was a limit of our own setup, which we fixed before these trials.

If you run retail media or sell through it, I would be interested to compare notes. Subscribe free to How Agents Shop to follow our next tests.