Agentic Commerce: When the Agent Is the Customer

Short answer

• Proven today: AI assistants send shoppers who click through and buy. Adobe's US retail data shows AI-referred traffic up 393% year on year in Q1 2026, converting 42% better than other traffic in March. Adobe's later figures show 62% growth in July and 127% in August. Adobe counts visits that arrive from AI sources. It does not measure agents buying on their own.
• Not proven: agents buying on their own at scale, what the industry calls agentic commerce. OpenAI pulled back from its own Instant Checkout in March 2026, and Google's agentic checkout, launched in November 2025, buys only after the shopper confirms.
• My argument, a prediction: whoever owns the agent holds the preferences. A user-owned agent squeezes margins on anything comparable. A platform-owned one sees the decision itself.
• The counterweight: the last big drop in search costs did not end price differences online. Brand and trust still mattered, and sellers made comparison harder.
• What to do this year: put key facts in plain server-rendered HTML, keep ranking in Google and measure AI referrals. All three rest on measured evidence.

The tracking dream, then and now

1997: the web tried to limit cross-site tracking

In February 1997, the first cookie standard, RFC 2109, restricted cookie sessions with other domains inside embedded or redirected requests. Third-party tracking runs on that kind of request. Any setting that allowed such sessions had to "default to off", for privacy.

Ten years later, on 13 April 2007, Google agreed to pay USD 3.1 billion for DoubleClick, a digital marketing technology company, saying the combination would offer better tools for targeting and analysing online ads.

2026: the promise moves to agents

On 29 July 2026, Mark Zuckerberg told investors on Meta's Q2 2026 earnings call that he expects "billions of people with a personal agent". He named five years from now as an example and left the period open. That is a prediction, from a company that sells agents.

What happened last time

This part is my argument, not data. The tracking of the 2000s did not hand the value to whoever tracked best. It went to whoever owned attention, first in search, then in the ad stack on top of it. If you pay for Google Ads and SEO today, the open question is who owns the agent.

What the data actually shows

AI referrals grow fast and convert well

Adobe measures more than 1 trillion US retail visits. In Q1 2026, traffic from AI assistants rose 393% year on year. In March 2026 that traffic converted 42% better than other traffic. In March 2025 it had converted 38% worse.

Growth is cooling, unevenly

For July 2026, Adobe reports AI traffic to US retail sites up 62% year on year, still converting 60% better. For August, its holiday report puts AI-driven traffic to retail sites up 127%. Both are well below the 269% of March 2026, but the rate does not fall in a straight line.

Read the label on this data

The data covers US retail only, and it comes from Adobe, which sells a product for visibility in AI answers. It counts visits that arrive from AI sources, which Adobe describes as shoppers following links, and it does not measure agents buying on their own. The closest Swiss baseline comes from the Swiss Federal Statistical Office. By spring 2025, 43% of people aged 15 to 88 had used generative AI, not specifically for shopping.

The 2026 survey of online retailers by the FHNW, a Swiss university of applied sciences, gives a number from the merchant side. Ten percent of the online retailers it surveyed call agentic commerce a revenue-relevant channel. That is a self-assessment, not a measurement of purchases by agents.

The flip: who owns the agent decides who wins

This section is argument, not data. Companies used to track you to steer you, and an agent that holds your preferences turns that around.

When the agent is yours

Your agent compares every offer against what you asked for. Ads lose power over a buyer who never sees them. My bet is that margins shrink on anything easy to compare.

When the agent is the platform's

If the agent belongs to Google, OpenAI, Meta or Alibaba, the platform sees the decision itself, not only the click. That is a fuller view than the 2000s ad stack ever had. It is also a logical bet.

Likely losers and winners (predictions, not findings)

Information middlemen and dropshippers

I expect the squeeze where a margin depends on the buyer not finding the original listing: comparison and affiliate sites, lead-gen brokers and, most clearly, dropshippers. I have found no margin data that shows it yet.

The signal from China

On 11 May 2026, Alibaba connected its Qwen app to the whole Taobao and Tmall catalogue of more than 4 billion products, with ordering inside the chat. Reuters had reported the plan. A price tracker in the Taobao app's assistant orders by itself at a target price. An announced capability, not proven usage.

Likely winners

My guess is that the winners will be whoever owns the agent layer, sellers with better products and machine-readable offers, trust infrastructure such as payments and returns, and physical operators in logistics.

Where the hype outruns reality

Autonomous checkout has barely started

OpenAI launched Instant Checkout on 29 September 2025 for US users. In March 2026 it let merchants use their own checkout instead, saying the first version lacked flexibility. CNBC reported analysts' view that it struggled to onboard merchants and show accurate product data.

Google's agentic checkout started rolling out in November 2025 with eligible US merchants and buys only after the shopper confirms. By September 2026, Google said its checkout protocol already enables direct checkout for "hundreds of thousands of brands and retailers". New capabilities for that checkout roll out in the US first, with Australia and Canada planned for early 2027. Google does not mention Switzerland. Neither checkout is autonomous buying at scale.

Cheaper search did not end price differences

In 2000, Brynjolfsson and Smith compared book and CD prices at 41 retailers. Online prices were 9 to 16% lower than in conventional stores, yet prices across online retailers still differed by 33% for books and 25% for CDs on average. In 2009, Ellison and Ellison showed retailers on a dominant price search engine using obfuscation, practices that frustrate search, to blunt it. I expect sellers to adapt to agents the same way, with agent optimisation as the next SEO.

Agents remove information steps, not physical ones

An argument: an agent can find, compare and order. It cannot hold stock, pool risk, extend credit or buy in bulk, so the intermediaries doing those jobs stay.

Correlated agents and shaped preferences

Also argument. Agents on similar models may herd.

Preferences partly form while you shop, so an agent optimising what you said you wanted may narrow what you would have chosen.

Even Zuckerberg says it is not easy yet

On the same July 2026 call, Zuckerberg said today's agents still take effort to set up.

What a Swiss business should do this year

Put the facts where every crawler can read them

In December 2024, Vercel and MERJ found that the AI crawlers of OpenAI, Anthropic, Meta, ByteDance and Perplexity do not render JavaScript. The exceptions were Google's Gemini, which uses Googlebot's infrastructure, and AppleBot. So your prices, services, locations and booking rules belong in the initial HTML, the text the server sends before any script runs. Check a key page with View Source, not the rendered view. The same gap is why LLM benchmark sites built as JavaScript apps are invisible to the crawlers of OpenAI, Anthropic and Perplexity.

Keep ranking, skip the schema sales pitch

In a 2025 study by the SEO tool vendor Ahrefs, 76% of pages cited in Google's AI Overviews ranked in the top 10. In a 2026 Ahrefs test, 1,885 already-cited pages that added JSON-LD (structured data describing a page to machines) showed no major citation lift against 4,000 controls. Keep schema for Google's rich results and put the effort into ranking.

Measure what AI sends you

In your analytics, segment visitors referred by AI assistants such as chatgpt.com and perplexity.ai. The blind spot is an agent that reads your page and answers without a click. It leaves no analytics trace, only a server-log hit. An agent that drives a real browser is the opposite case. It can run your analytics script and pass as an ordinary visit. That is my reasoning, not a measurement.

Where my own agent gets through, and where it stops

Since May 2026 I have let an agent act in my own logged-in browser, under a written authorisation that even covers purchases. On a data API that asked only for a username and password, it signed up, created an API key and wired it into my project with no human step. On another site it read the activation email and carried on until the free trial asked for a card, then entered nothing and handed the task back to me.

The other stops came from the sites: Google blocked its first search as unusual traffic, other search engines showed bot challenges, a marketplace redirected it to a login. My rules forbid it to solve a CAPTCHA or to pay without my go. None of its first 58 logged actions, from May to September 2026, is a purchase.

My reading, reasoned rather than measured, is that an agent completes what a form can express. Payment, bot checks and logins are where a human still decides.

Start with what AI can read on your site

To check a page's AI visibility along with its SEO basics, run it through my free website check. To go through crawler access and your AI referral traffic together, start with a first conversation.

Frequently asked questions