Are AI agents actually completing purchases yet?

Not in any way a company has been willing to disclose. OpenAI walked back Instant Checkout to a discovery-only role in March 2026 without ever publishing an order count. Etsy's own CEO called agentic traffic less than one percent of the total. Visa has disclosed hundreds of agent-initiated transactions, no dollar figure attached. No retailer, card network, or platform has published a real transaction count or GMV figure for a completed agentic purchase.

Does publishing an llms.txt file get you cited by ChatGPT or Google?

No, and this is now on the record from the source. Google's own Search Central documentation states that Google Search ignores llms.txt files entirely. Google's John Mueller has said no AI system currently uses it. Ahrefs checked 137,210 domains and found 97 percent of published llms.txt files got zero requests in a month, and of the small remainder that did, most of the traffic was coding agents, not search engines.

What actually changes an AI shopping agent's decision?

Research from Bayes Business School and King's Business School tested standard marketing levers against real shopping agents. Star ratings consistently swayed the agent's choice. Price reliably worked against the higher-priced option. Scarcity cues, countdown timers, and strike-through pricing, the tools built for a human's fear of missing out, mostly did nothing, and sometimes backfired. A separate study found agents penalize sponsored tags while rewarding platform endorsements, the reverse of what works on a human.

What should a marketer actually do to prepare for AI agents as buyers?

Two different jobs need two different files. For agents that reason and answer, the leverage is in structured data, named authorship, and a fast, clean, crawlable page, the same things that already drive answer-engine visibility. For agents that transact, the leverage is a real machine-readable capability manifest, the kind the Universal Commerce Protocol asks for at /.well-known/ucp, not a text file nobody's crawler is reading. Ship the manifest. Keep the llms.txt file if you already have one, but do not credit it with anything.

Marketing to Machines: When AI Agents Become the Customer

Q: Will 85 percent of customer data come from AI agents by 2027?

That is not what Gartner said, though it is how the stat usually gets repeated. The real subheading, from a January 2025 Gartner release, is that 85 percent of customer data will be collected from automated interactions or those led by AI agents. The or is doing most of the work: it folds in ordinary automation, like forms and app telemetry, that has nothing to do with agents. Treat the pure-agent version of this number as a distortion, not a citation.

Gartner says the end of channel-based marketing is here. The receipts say nobody has sold a single product to an agent and been willing to put a number on it.

Both things are true, and the gap between them is the actual story. Gartner's own analyst, Emily Weiss, opened 2026 with a line built to be quoted: "This marks the end of channel-based marketing as we know it." Six months earlier, a separate Gartner release put a number on where this is heading: by 2028, 90 percent of B2B buying will be AI agent intermediated, pushing more than $15 trillion of B2B spend through agent exchanges. Those are real, dated, on-the-record predictions from the analyst firm every CMO's board reads.

Now the other half. I spent this week chasing every company with a reason to brag about agentic commerce, OpenAI, Etsy, Shopify, Stripe, Visa, Mastercard, Amazon. Not one has published a real transaction count or a dollar figure for a completed agent-driven purchase. The company that launched the flagship product, OpenAI's Instant Checkout, quietly walked it back in March 2026. That is the honest starting point for this piece, and it is more useful than the prediction.

The stat everyone is going to misquote

The 85 percent figure is already circulating as "85 percent of customer data will come from AI agents." That is not what Gartner published. The actual subheading, from a January 2025 release, reads: "By 2027, 85% of Customer Data will be Collected from Automated Interactions or Those Led by AI Agents." The word "or" is not decoration. It folds in ordinary automation, form fills, app telemetry, tracking pixels, alongside anything an agent actually does. No methodology or sample is attached to the number either. It is an analyst forecast, and the version worth repeating is the one Gartner actually wrote, not the tighter, scarier one that spreads faster.

The same release predicts mobile app usage will fall 25 percent by 2027, driven by AI assistants replacing apps for many functions, and, separately, by brands consolidating apps into shared partnerships to defray development costs. Notice that second cause. It is not agents. It is ordinary cost-cutting riding the same headline. A companion prediction says over a third of web content will be created for the purposes of gen-AI search by 2026, a figure that, on inspection, appears only in a subheading and is never defended in the body copy underneath it. Directionally believable. Evidentially thin. Say so.

Gartner has also published the honest hedge that belongs next to all of it: over 40 percent of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls, and Gartner's own estimate is that only about 130 of the thousands of vendors calling themselves agentic are building anything real. Analyst Anushree Verma's phrase for the rest is agent washing, the rebranding of ordinary chatbots and robotic process automation with a new label and no new capability.

What the machine actually buys, right now

OpenAI's own retreat is the cleanest evidence available. In March 2026, the company wrote that Instant Checkout "did not offer the level of flexibility that we aspire to provide," and shifted merchants back to their own checkout flows inside an in-app browser. Shopify's own product lead confirmed the purchase now completes on the merchant's site, not inside ChatGPT. Walmart replaced the integration with its own app rather than keep using it. The honest 2026 pattern is discover in the chat, buy on the site, which is a real and useful shift, and not the agent-does-the-buying story that gets sold at conferences.

Ask the retailers who actually ran the pilot and the caution compounds. Etsy's CEO, on an earnings call, called agentic traffic "less than 1% of our total traffic," and her CFO added the number was too small to move anything. That is traffic, not purchases, and Etsy has never disclosed a purchase figure in either direction. Shopify reports growth multiples, AI-referred orders up nearly 13 times year over year, without ever naming a base big enough to judge. The one absolute figure on the public record comes from Visa: "hundreds of secure, agent-initiated transactions have now been successfully completed," no dollar volume attached, eight months after the company's first press release on the topic. The only large, real transaction count anywhere is a machine-to-machine micropayment protocol running at an average ticket of 32 cents, which proves agents move money for compute, not that they are shopping for anything a marketer sells.

Adobe's own data makes the case better than I can. A January 2026 report showed retail AI traffic under 1 percent of total visits, in an unlabeled chart. Every report since has dropped the share number entirely while continuing to publish enormous year-over-year growth figures, 393 percent in the first quarter, 269 percent in March, off a base the company has stopped showing anyone. That is the mechanism behind most of this year's AI-traffic headlines: a real percentage, applied to a denominator nobody discloses. Salesforce's most quoted stat, agents driving 67 billion dollars in Cyber Week sales, is defined in the company's own release as coming from "personalized product recommendations and conversational customer service," which is on-site merchandising and support chatbots, not an agent completing a checkout. And Salesforce's forward prediction that 20 percent of 2026 holiday traffic will come from AI chat agents explicitly includes, in its own words, "competitor scrapers fueling algorithmic price-matching." Read the fine print before you build a plan on the headline.

The stats that do not survive a source check

Two numbers anchor most of the agentic-commerce discourse, and neither holds up. Ahrefs' claim that AI search visitors convert 23 times higher than organic traffic is measured on exactly one website, ahrefs.com, and the same post admits those visitors click through 75 percent less often than ordinary search traffic. Semrush's claim that an AI search visit is 4.4 times as valuable, introduced with the words "we have seen that," carries no sample, no date range, and no definition of what counted as a conversion. Both numbers get cited constantly as settled research. Neither is one.

The pattern repeats across the industry's own larger studies, which quietly undercut the smaller, punchier ones. Semrush's own 50,000-site study puts AI referral traffic at 0.14 percent of total visits. Ahrefs tracks a fixed cohort of roughly 75,000 sites and watched the same figure climb from 0.28 percent in March 2026 to 0.39 percent by June. Zero peer-reviewed research measures AI agent traffic share at all. Every number in circulation, on every side of this argument, is vendor telemetry with no external replication. Say that plainly before quoting any of it.

What actually changes an agent's decision

Researchers at Bayes Business School and King's Business School ran real shopping agents against the standard toolkit of persuasion: scarcity cues, countdown timers, strike-through discounts, bundles. Across four models and four product categories, none of it reliably moved the agent, and some of it backfired. Two things did work with consistency: star ratings swayed the agent's choice, and price worked in the predictable direction, a higher price reliably hurt selection. The tools built for a human's fear of missing out mostly do not translate, and that gap is worth building a campaign around.

A Columbia Business School study of live shopping agents found the same asymmetry from a different angle: agents "consistently penalize sponsored tags while rewarding platform endorsements," the exact reverse of how a human treats an ad label, and a seller can shift real market share with simple, honest description tweaks that respond to what the buyer actually asked for. The manipulation risk is real too. One study found a single deliberately polluted page can fool an agent's product judgment up to 27 percent of the time, and replacing the top three results with polluted pages pushes that past 73 percent. Another found that once every competitor in a category adopts the identical optimization tactic, the payoff per player collapses from a meaningful edge to functionally nothing, a textbook arms race, while a well-known incumbent brand still gets recommended 100 percent of the time against an identical, unknown competitor, until the challenger earns just one-tenth of a star more in reviews. Reputation, not gimmick, is what an agent actually reads.

Does llms.txt actually work? No, and it is not close.

llms.txt is the single most oversold practice in AI marketing right now, and marketers keep asking me about it. Jeremy Howard proposed it in September 2024 as a way to hand a clean, curated version of a site to an AI system. It has since been sold to marketers as an AI-search visibility play. It is not one, and the people who would actually know have said so on the record.

Google's own Search Central documentation states it plainly: "Google Search itself doesn't use them," and adds that publishing one "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them." Google's John Mueller said the same thing directly: "FWIW no AI system currently uses llms.txt." Ahrefs checked the server logs behind 137,210 domains and found that 97 percent of published llms.txt files received zero requests in a month, and their sharpest line is worth repeating in full: "Zero requests came from AI bots for llms.txt files that don't exist. They never go looking." Of the small remainder that did get hit, the largest category of requester was coding agents, not search or answer engines. Much of the adoption growth reported this year is also not organic: Shopify auto-deployed llms.txt as an alias for a different file, agents.md, to every store on the platform in May 2026, which alone accounts for the bulk of the reported adoption curve.

That does not mean every machine-readable file is theater. The Model Context Protocol, the standard Anthropic shipped in November 2024 for connecting tools to reasoning agents, now runs more than 10,000 active public servers and was handed to a neutral foundation at the end of 2025. And the concrete, forward-looking move for a commerce site is a real capability manifest, the kind the new Universal Commerce Protocol asks for at a fixed address, /.well-known/ucp, describing exactly what an agent is allowed to do on the site. That file gets read, because it is built for the agents that transact, not the search engines that answer questions. llms.txt is not that file, and no honest reading of 2026's server logs says otherwise.

What I actually do, and what I do not claim it does

This site runs an llms.txt and an llms-full.txt file. I am not going to pretend that is why anything on this site gets cited, because the research above says plainly that it is not. I keep them because they cost nothing to maintain and they are genuinely useful to the one class of reader that does fetch a file like this on purpose: a coding agent or a developer tool pulling context about the site, which is exactly the use case Howard built it for in the first place.

What actually earns a citation is duller and already running on every page here: named authorship tied to a real person, not an organization, Article and FAQPage schema that validates, a fast, server-rendered page with nothing blocking first paint, and a dateModified that updates honestly when the content changes instead of sitting frozen at the publish date. That is the boring, structural work the research above says the machines actually reward, and it is the same discipline whether the reader on the other end is a person or a model deciding what to cite. I will keep publishing an llms.txt file. I will not tell you it is the reason anything works.

Gartner mobile app usage prediction, including the 85 percent customer-data subheading and the one-third gen-AI content subheading: Gartner newsroom, January 15, 2025, analyst Emily Weiss. Gartner's channel-based-marketing prediction: Gartner newsroom, January 15, 2026. The 90 percent B2B agent-intermediated buying and $15 trillion figure: Gartner newsroom, October 21, 2025. The 40-plus percent agentic-project cancellation prediction and the roughly 130 real vendors estimate: Gartner newsroom, June 25, 2025, analyst Anushree Verma. OpenAI's Instant Checkout retreat: OpenAI, "Powering Product Discovery in ChatGPT," March 24, 2026. Etsy's under-1-percent agentic traffic disclosure: Etsy Q4 2025 and Q1 2026 earnings calls, CEO Kruti Patel Goyal. Visa's hundreds-of-transactions disclosure: Visa, December 18, 2025. The x402 micropayment protocol figures: x402.org live counter, read August 4, 2026. Adobe's retail AI traffic-share chart and the conversion reversal: Adobe Digital Insights, AI-Sourced Traffic Insights, January 12, 2026, and follow-on reports in April and June 2026. Salesforce's Cyber Week and 2026 holiday-prediction wording: Salesforce newsroom, December 5, 2025 and July 20, 2026. Ahrefs' 23x conversion claim and its own caveat on click-through: Ahrefs, June 2025. Semrush's 4.4x claim: Semrush, July 21, 2025. Semrush's 50,000-site traffic-share study: Semrush, April 27, 2026. Ahrefs' 74,752-site fixed-cohort tracker, climbing from 0.28 percent in March 2026 to 0.39 percent by June: Ahrefs, May and June 2026 updates. The shopping-agent persuasion research: Jafar Sabbah and Oguz A. Acar, reported in Harvard Business Review, May 12, 2026. The sponsored-tag penalty and market-share findings: "What Is Your AI Agent Buying?", arXiv 2508.02630, August 2025. The single-polluted-page manipulation rates: "One Polluted Page Is Enough," arXiv 2606.13610, June 2026. The incumbent-advantage and GEO-arms-race collapse findings: "Incumbent Advantage," arXiv 2606.17443, June 2026. Google's statement that Search ignores llms.txt: Google Search Central, "Guide to Optimizing for Generative AI Features," updated July 10, 2026. John Mueller's llms.txt statement: Bluesky, June 17, 2025. The 97-percent-zero-requests finding: Ahrefs, June 15, 2026, 137,210 domains. Shopify's llms.txt auto-deployment as an agents.md alias: Shopify changelog, May 2026. The Model Context Protocol's server count and foundation transfer: Anthropic and the Agentic AI Foundation, December 2025.

About the author

Jeff Brokaw is a sitting CMO and Certified Chief AI Officer who ships AI in production, not slideware. He built the answer-engine optimization program behind a defense manufacturer's demand pipeline, runs the machine-readable practice described above on his own properties, and built the precision-measurement layer the category doesn't sell. He has been building AI systems commercially since 2016. He built the engine behind $185M in new business for that same manufacturer, and authored the go-to-market behind a $114M institutional raise that came together in under 30 days.