Before you read the rest of this, try something.
Open any AI assistant you use and ask it for a product in your category, the way a real customer would. Don’t mention your brand name. Something closer to “What’s a good protein powder for someone who’s lactose intolerant, under $50?”
Did it name your brand? If it did, was the description or the price right? When you clicked through, was the product in stock?
Answers shift between sessions and assistants, so treat this as a gut check rather than a measurement. Still, most founders I know run it once and go quiet for a minute. Whatever came back is increasingly what your next customer hears before they ever see your site.
In the last couple of years, online buying patterns have changed from “search, browse, compare, buy” to “ask an AI, get a recommendation, buy.”
AI shopping agents sit in that middle step. They build the shortlist before your customer does.
It’s a real shift, and a lot of founders are trying to work out what it means for them. The more time I spend on it, the more convinced I am that most of the conversation is aimed at the wrong target.
Getting recommended is not the same as getting bought
Most advice about AI and ecommerce stops at visibility. Will the assistant mention you? That matters, but it’s only the first of three tests, and the easiest to pass by accident. You want to check whether your business is:
- Discoverable. Does the agent know you exist, and does it understand what you sell and who it’s for?
- Credible. When it weighs you against alternatives, does it have a reason to trust you more?
- Transactable. Once it picks you, can the purchase happen cleanly, at the right price, with the product in stock, whether a human clicks through or the agent completes the order itself?
A brand can pass the first test and fail the other two. That’s the store that gets mentioned, gets quoted last season’s price, and loses the sale on a sold-out page.
Adobe’s data shows AI-referred traffic to US retail sites grew 62% year over year in July 2026, and those visits converted 60% better than non-AI traffic.
Another survey from Salesforce estimates AI and agents influenced about 20% of 2025 holiday online sales, around $262 billion, which includes retailers’ own recommendation engines, not just third-party assistants.
Adobe’s number is narrow and fast-growing; Salesforce’s is broad and includes retailers’ own agents, so you shouldn’t read them interchangeably. The shortlist is moving upstream, and getting on it takes months, not weeks.
What AI shopping agents read, and why one store gets picked
Picture two brands selling trail running shoes at similar prices.
A shopper asks an assistant for “a waterproof trail running shoe for wide feet, under $150, that can arrive before next weekend.”
Brand A has the better-looking site. The product page opens with a lifestyle video. Specs sit inside an image. Wide sizing only appears in a dropdown, and waterproofing is mentioned in a blog post, not on the product page. The site says $139; a marketplace listing still shows $129 from an old promotion. Reviews load through a widget that a crawler can’t read.
Brand B’s site is plainer. But its product page states, in text: wide (2E), waterproof membrane, $135, ships in two business days. The same facts appear in its marketplace listing and product feed. It has a few hundred reviews, several from wide-footed runners, and shows up in independent reviews and forum threads where those runners compare options.
An agent can verify Brand B’s claims in seconds. For Brand A, it has to guess and find two prices. It recommends Brand B, and Brand A will never know that the question was asked.
It’s a simplified example, but it’s one of the most common patterns in AI shopping. Brands rarely lose on the product. They lose because the facts about it are incomplete, inconsistent, or impossible to verify.
What agents lean on most:
- Attributes in plain text. Size, material, fit, compatibility, use case. Every missing attribute is a question the agent can’t answer about you.
- Structured data. Product schema and complete feeds with proper identifiers that the format machines read first.
- Price and availability that match everywhere. When your site, feed, and marketplace listings disagree, you look unreliable.
- Third-party proof. Reviews, independent comparisons, and community discussion. What others say about you outweighs what you say about yourself.
Design still matters. A human still lands on your page and decides, and many assistants can read images. But if your key facts live only in a video or banner, you’re betting a machine reads them the way you meant. Put them in text too.
The industry calls this generative engine optimization. The labels matter less than the discipline: make true, specific facts about your product easy to find, easy to check, and identical wherever they appear.
The last mile: from recommendation to transaction
Most AI visibility work stops at the recommendation. That’s an expensive place to stop.
Today, most AI-driven purchases still end with a person clicking through. That visitor already asked a specific question and got a specific answer. If your page contradicts it, with a different price or an out-of-stock size, you lose them faster than anyone, because they arrived with an expectation which you just broke. That makes conversion optimization for AI-referred shoppers its own job: the landing page has to confirm what the agent promised, not restart the sale.
The next stage is agents completing purchases themselves. Open checkout standards now let AI assistants buy on a shopper’s behalf, and major commerce platforms are adding support. It’s early, but the requirements are clear: real-time inventory and pricing, clean product identifiers, and payment that works without a human filling in a form. Find out whether your platform will support it.
Founders I talk to usually hesitate here, fearing a loss of control: an agent applying a discount you never approved, or selling stock you don’t have. It’s a fair fear. The answer isn’t to avoid agents but to demand guardrails. Those would be price limits that an agent can’t override, human review for exceptions, audit logs, every action visible and reversible.
Where to start, and how to know it’s working
If I were starting tomorrow, I’d do five things, in this order.
- Build a baseline. Write 20 to 30 prompts your customers would actually use. Run them monthly across the main assistants and record whether you appear and whether the facts are right.
- Fix product data on your bestsellers first. Every attribute in text, schema on every page, a complete feed.
- Make your facts consistent everywhere. One price, one shipping promise, one returns policy across your site, marketplaces, and feeds.
- Earn third-party proof. Recent reviews, honest comparison pages, presence in the communities your customers trust. It’s the slowest lever and the hardest to fake, which is why it counts.
- Prepare the transaction. Landing pages that match what an agent would say, agent-checkout readiness, and written guardrails.
Then measure in three layers.
- Visibility: Track how many of your baseline prompts mention your brand, and check whether the assistant describes you accurately.
- Engagement: Segment AI-referred sessions in your analytics, and compare their bounce and add-to-cart rates with your other channels.
- Revenue: Count the orders that come from AI-referred sessions, and add an AI option to your “how did you hear about us” survey. Attribution undercounts shoppers who research in an assistant and come back later to buy.
Expect visibility to move first, then engagement, then revenue. If visibility improves and revenue doesn’t, fix the last mile before you write more content.
The shortlist is the new storefront
The good news for startups is that agents don’t care about your ad budget. They care whether your facts are complete, consistent, and credible, and whether the sale can actually happen. A small, focused team can win that game.
So run the test again in 90 days. Will the agent recommend you then?
About the author
Rahul Amarendra is CEO of CommerceV3 (CV3), where he is building an Agentic Revenue Engine that helps commerce brands get discovered, recommended, and bought in AI-driven shopping. A three-time VC-backed founder and an alumnus of IIT Delhi and IIM Ahmedabad, he has spent his career building and scaling businesses across commerce, SaaS, and AI.
Last Updated on October 28, 2026 by isabel