Agentic Commerce: Getting Recommended by AI Shopping Agents

Agentic commerce is when AI agents take a goal like "best running shoes under $120 by Friday" and do the shopping themselves: research the options, compare specs and price, read the reviews, and increasingly place the order on your behalf. It is already running inside ChatGPT, Perplexity, Amazon's Rufus, and Google. The shopper states intent once, and the agent collapses a hundred tabs into a short list or a single buy. So the brand that gets picked is the one whose product is clean, well reviewed, spec clear, properly structured, and an obvious match for the stated intent and budget.

That is the whole game in two sentences. The shopper delegates the comparison, and the agent rewards whichever product it can read and trust fastest. Your job stops being "rank on page one" and becomes "be the option an agent can confidently hand to a person who already decided to buy."

Now let me be straight with you before we go further. The eye-popping numbers in this space are forecasts, not bank statements, and I will label them as such every time. Hard data on exactly how agents weight one product over another is still thin, because the companies running these agents do not publish their ranking logic. So treat the projections as direction and the tactics as the durable part. The tactics are durable because they are just good commerce hygiene that agents happen to reward.

Key takeaways

• By 2030, roughly half of online shoppers are forecast to use AI shopping agents, accounting for about 25 percent of their spending, per Morgan Stanley (via commercetools). That is a forecast, not current revenue.
• Gartner projects AI "machine customers" will replace 20 percent of human storefront interactions by 2028, and that up to 80 percent of internet traffic could be agent-driven by 2035 (via Human Security). Also forecasts.
• Adoption is real but early. Around 35 to 40 percent of consumers use AI in some part of shopping today, yet only about 10 to 15 percent completed a purchase directly after an AI referral, per MetaRouter.
• Agents reward readability and trust: clean structured data, accurate specs, recent reviews, in-stock availability, and a price that matches intent. The product an agent can parse with confidence wins.
• Transparency is limited. No major agent publishes its product-ranking logic, so the honest move is to measure where you actually show up rather than guess at the algorithm.

Here is the long version, because "agents are coming" sounds soft until you see how they actually pick.

What agentic commerce is (and how it differs from old AI shopping)

Let's get the definition tight, because people use "AI shopping" to mean three different things and it muddies everything.

Old AI shopping was a recommendation engine. You browsed, the site suggested "you might also like," and a smarter search box surfaced products that matched your query. You still did all the real work: comparing the options, weighing the trade-offs, opening fifteen tabs, and clicking buy yourself. The AI was a helpful assistant standing next to you. It pointed. You acted.

Agentic commerce flips the roles. You hand the agent a goal and a constraint, and it does the pointing and the acting. "Find me a quiet office chair under $300 with good lumbar support that ships by next week." The agent goes off, reads product pages, compares specs, scans reviews for the word "squeaky," checks stock and delivery dates, and comes back with a short list or, in the newer flows, just buys the thing. You delegated the task instead of getting suggestions for it.

The difference that matters is action. A recommendation needs a human to finish it. An agent finishes it. That single shift, from suggest to act, is what changes the rules for brands, because now there is a software layer between your product and the shopper, and that layer has opinions about what counts as a good match.

If you want the formal definition with the surrounding concepts, we keep one in the agentic commerce glossary entry, and the broader idea of an autonomous software buyer lives in the AI agent glossary entry. The short version: an agent is a model that can take a goal, make a plan, use tools (search, browse, checkout APIs), and act with some autonomy. Point that at shopping and you get agentic commerce.

Where agents already live

This is not a someday thing waiting on some future product launch. The surfaces exist now:

• ChatGPT browses, compares products, and has been rolling out shopping and checkout features.
• Perplexity answers buying questions with cited product picks and has tested its own buy flow.
• Amazon's Rufus is a shopping assistant living inside the world's biggest store, which I dug into in the Amazon Rufus breakdown.
• Google surfaces products inside AI Overviews and AI Mode, and is wiring agentic checkout into its shopping graph.
• Gemini answers product questions and pulls from Google's commerce data.

Five doors, each with its own taste in products. Which is exactly why you measure each one separately instead of assuming a single strategy covers all of them, the same way we treat citations in how AI engines choose sources.

The numbers, all forecasts, labeled honestly

Time for the figures, with a warning stapled to the front: almost everything big in this section is a projection. I am going to say "forecast" enough times that it gets annoying, on purpose, because the gap between "could happen by 2030" and "is happening now" is where a lot of strategy money gets wasted.

Start with the headline. Morgan Stanley forecasts that by 2030, roughly half of online shoppers will use AI shopping agents, and those agents will account for about 25 percent of their spending (via commercetools). Read that twice. Half the shoppers, a quarter of the dollars, four years out. If that lands even partway, the agent becomes a buyer you have to sell to directly, not just a channel that sends humans your way. But it is a forecast. Nobody has that revenue in the bank yet.

Gartner goes harder and longer. By 2028, Gartner projects AI "machine customers" will replace 20 percent of human storefront interactions, and by 2035, up to 80 percent of internet traffic could be agent-driven (via Human Security). The 80 percent number is the one that gets quoted in every keynote, so let me put the honest frame on it: it is a 2035 projection about traffic, not sales, and "agent-driven traffic" includes a lot of bots that are not buying anything. Useful as a direction. Dangerous as a plan.

Now the reality check, which I like a lot more because it is measured, not modeled. About 35 to 40 percent of consumers already use AI in some part of their shopping, but only roughly 10 to 15 percent have completed a purchase directly after an AI referral, per MetaRouter. That gap is the whole story of where we actually are. People happily use AI to research and shortlist. Far fewer let it close the deal. The funnel narrows hard at the buy button, because handing a bot your credit card is a bigger ask than asking it for ideas.

So what do you do with a table that is half forecast? You plan for the measured numbers and prepare for the forecast ones. The 10 to 15 percent who buy through AI today are real customers you can win right now. The 50 percent forecast is a reason to get your product data in shape before the wave, not a reason to bet the quarter on it. For the wider adoption picture across AI search generally, the AI search statistics for 2026 post collects the measured ones in one place.

How AI shopping agents actually choose products

Here is the part you can control. Agents are not magic and they are not malicious. They are pattern matchers under a deadline, trying to satisfy a stated goal with the cleanest available evidence. When you understand what "cleanest evidence" means to a model, the playbook writes itself.

An agent runs a rough loop: parse the shopper's intent and constraints, gather candidate products from the surfaces it can read, score each candidate against the constraints, and present or buy the winner. You win or lose at the scoring step, and scoring runs on whatever data the agent could extract. If your product page is a mess of marketing prose with the actual specs buried in an image, the agent cannot score you well, so you lose to the boring competitor whose spec table it could read in one pass.

Let me walk the signals one at a time, because each one is a lever.

Specs. The agent matches against constraints, and constraints are specs. Budget, size, material, compatibility, color, delivery date. If a shopper says "under $120," price is a hard filter. If they say "for wide feet," width is a hard filter. Products with complete, accurate, machine-readable specs survive the filters. Products with vague or missing specs get dropped, not because they are worse, but because the agent cannot prove they fit. Incomplete specs read as a "no" to a cautious model.

Price. Not just low, but legible and matched to intent. The agent needs your current price in a form it can read, and it needs to match the stated budget. A great product at $130 loses a "under $120" query to a worse product at $115. That is not unfair, it is the literal instruction. And stale or inconsistent prices across surfaces make an agent nervous, because it does not want to recommend something at a price that turns out wrong at checkout.

Reviews. Volume and recency and sentiment, in that rough order of how much agents seem to lean on them. Reviews are the closest thing an agent has to social proof it can quote, and they often get summarized straight into the answer ("reviewers praise the battery life but mention the strap"). Thin review counts read as risk. A wall of recent, detailed reviews reads as a safe pick. You do not control the sentiment, but you absolutely control whether you are collecting reviews at all.

Availability. In stock and shippable by the deadline, or you are invisible for any time-bound query. "By Friday" is a constraint, and an agent will not recommend something it cannot get there in time. Out-of-stock is the hardest no in commerce, and agents enforce it without mercy.

Structured data and feeds. This is the quiet kingmaker. Schema markup, clean product feeds, and a well-formed merchant catalog are how an agent reads your product without guessing. Structured data turns your page from prose the model has to interpret into facts the model can simply ingest. We go deep on the markup side in schema markup for AI search, and the mechanism is the same here: machine-readable beats human-pretty when the reader is a machine.

Entity clarity. The agent needs to know who you are with confidence, the same way AI engines do for citations. A clean, consistent brand entity (one name, one identity, consistent details everywhere) helps the agent trust that the product it found is really yours and not a knockoff or a confused listing. This is plain old entity SEO pointed at commerce. Confusion about your identity is friction, and friction loses to clarity.

Under the hood, a lot of this runs on retrieval: the agent pulls candidate product data into context before it reasons, which is the same retrieval-augmented generation pattern behind AI search. Clean, structured, retrievable data is the fuel. If the fuel is dirty, the engine misfires no matter how good your actual product is.

Selection signals and your move

Here is the cheat sheet. Left column is what the agent reads, right column is what you do about it.

None of this is exotic. It is commerce hygiene that you probably half-did already. The difference is that a human shopper forgives a missing spec field and an agent does not. The bar for "good enough data" just went up, because the reader changed from a forgiving person to a literal machine.

What brands should do now

So what do you actually go do on Monday? Four moves, in priority order, because doing them out of order wastes effort.

1. Fix your structured product data first. This is the foundation and everything else sits on it. Audit your product schema and your merchant feed for completeness and accuracy. Every spec field filled. Real values, not "see description." Prices and stock that update in real time. If you sell through Shopify, the platform-specific version of this is in AI visibility for Shopify, and the broader ecommerce playbook lives on the ecommerce AI visibility page. Get the data clean before you touch anything else, because an agent reading bad data will skip you no matter what else you do.

2. Pour fuel on reviews. Volume and recency are the levers you control. Set up a real review collection flow if you do not have one, and keep it running so your review base stays fresh instead of frozen in 2024. Recent reviews signal a live, trusted product. Stale or sparse reviews signal risk, and a cautious agent routes around risk.

3. Be present and consistent across surfaces. An agent might read you through Google's shopping graph, through Amazon, through a general web crawl, or through a retailer's catalog. If your product looks great in one place and broken in another, the agent gets a mixed signal and trusts the worse one. Make your name, specs, price, and identity consistent everywhere. This is search everywhere optimization applied to products: show up the same way in every room the agent might walk into.

4. Lock down your entity. One brand name. One consistent identity. Consistent product identifiers (GTINs, SKUs, model numbers) across listings. The cleaner your entity, the more confidently an agent can say "yes, this is the product the shopper wants, from a brand I can identify." Confusion is friction, and in a delegated purchase, friction is fatal.

And the move under all four: measure where you stand before and after. You cannot fix what you cannot see, and "are we getting recommended by agents" is a measurable question, not a vibe. Run the buyer questions a real shopper would ask ("best wireless earbuds under $80," "durable hiking boots for wide feet") across the engines and record when your brand shows up, how it is framed, and who beats you. That is exactly what AI Citation Monitor does across the five engines it tracks, ChatGPT, Perplexity, Gemini, Google AI Overviews, and Microsoft Copilot, with confidence intervals so you are not fooled by one lucky run and a competitor share-of-voice view so you can see who is eating your lunch. If you are starting from zero, AI brand monitoring covers the discipline, and if your products are simply absent, why a brand does not show up in ChatGPT is the troubleshooting path.

The honest caveats

I would be doing you a disservice if I left you thinking this is all settled science. It is not, and pretending otherwise is the kind of hype this whole post is trying to avoid.

Transparency is genuinely limited. No major agent publishes how it ranks one product over another. The signals I listed (specs, price, reviews, availability, structured data) are well-grounded in how these models work and in what practitioners observe, but the exact weighting is a black box, and it changes as the models update. Hard quantitative data on agent ranking bias, the kind where you could say "a 10 percent better review score buys you X percent more recommendations," basically does not exist in public yet. So treat the playbook as well-reasoned direction, not a calibrated formula. I am confident about the what. I am honest that the how much is fuzzy.

Adoption is early, and the big numbers are forecasts. I said it up top and I will say it again here, because it is the most important caveat. The 50 percent by 2030 (Morgan Stanley, forecast) and the 80 percent of traffic by 2035 (Gartner, forecast) are projections, not realized revenue. The measured number, the one in the bank, is that only about 10 to 15 percent of consumers have actually bought after an AI referral. The trend line points up and the surfaces are real, but if someone tells you agentic commerce is half your revenue today, they are selling something.

Trust is the real bottleneck, not technology. The tech to let an agent buy for you mostly works. The reason most people still do not let it is trust: handing a bot your payment details and your judgment is a big psychological step. That gap between "I'll let AI research" and "I'll let AI buy" is where adoption is stuck, and it is a human problem, not an engineering one. It will close, but slower than the keynotes imply.

So here is the honest read. Prepare now, bet later. Get your product data clean, your reviews flowing, your entity tight, and your presence consistent, because those are durable wins that pay off in regular search too. Then measure relentlessly so that when the forecasts start turning into real buyers, you already know which agents recommend you and which ones skip you. The brands that win this will not be the loudest. They will be the cleanest, the easiest to read, and the easiest to trust, which is a deeply unglamorous and very achievable bar.

FAQ

What is agentic commerce?

Agentic commerce is when an AI agent takes a shopping goal, like best running shoes under $120 by Friday, and does the work itself: researching products, comparing specs and prices, reading reviews, and increasingly placing the order. It runs inside tools like ChatGPT, Perplexity, Amazon's Rufus, and Google. The shopper states intent once, and the agent narrows hundreds of options down to a short list or a single buy.

How do AI shopping agents choose which products to recommend?

Agents reward products that are easy to read and easy to trust. That means clean structured product data, accurate and complete specs, a healthy volume of recent reviews, in-stock availability, a price that matches the stated budget, and a consistent presence across the surfaces the agent pulls from. If the agent cannot confidently parse your product against the shopper's intent, it skips you for one it can.

Is agentic commerce actually happening yet, or is it hype?

Both, honestly. Real agents exist and people use them, but completed purchases through an agent are still early. Roughly 35 to 40 percent of consumers use AI somewhere in their shopping, yet only about 10 to 15 percent have bought directly after an AI referral, per MetaRouter. The huge numbers you see, like half of shoppers using agents by 2030, are forecasts, not realized revenue.

What is the difference between agentic commerce and old AI shopping?

Old AI shopping recommended; agentic commerce acts. A recommendation engine showed you products and you did the comparing and buying. An agent takes the goal, does the research and comparison itself, and in the newer flows completes the checkout for you. The shift is from a smarter search box to a delegate that you hand a task and a budget.

What should brands do first to get recommended by AI agents?

Start with your structured product data. Clean, complete, accurate specs and schema markup are what an agent reads first, and gaps there quietly knock you out of consideration. Then make sure your reviews are plentiful and recent, your stock and pricing are accurate everywhere, and your brand looks consistent across every surface an agent might pull from. Measure where you currently show up so you know what to fix.

Can I track whether AI agents recommend my products?

Yes, by running the buyer questions a real shopper would ask across the AI engines and recording when your brand appears, how it is framed, and which competitors show up instead. Because model output is noisy, you repeat each prompt enough times to get a confidence interval rather than trusting one run. AI Citation Monitor automates this across the five engines it tracks: ChatGPT, Perplexity, Gemini, Google AI Overviews, and Microsoft Copilot.