What Agentic Buying Changes for a Founder

I would change the evidence a third party can quote before I change the copy on the site. Agents are already comparing brands and recommending products, and the buyer on the other side of that recommendation may never open a homepage. The spend that matters is the spend that leaves a sentence an agent can lift, with a name, a date, and a source attached.

The comparison I use for that call is Agentic Commerce Bypasses Your Website. Here. That page is the deeper resource. This essay is how I would use it as a founder: which numbers I would put in a budget meeting, which ones I would refuse to treat as a forecast I must hit, and what would make me reverse the sequence.

The commerce figures I would put on one page

McKinsey's estimate, as carried by Forbes Councils, is that AI agents will mediate $3 trillion to $5 trillion in global consumer commerce by 2030. I would write that range on the page as a range. A founder who picks $5 trillion and builds a plan to it has already left the source.

The same write-up carries the nearer measurement. eMarketer projects AI platforms will account for $20.9 billion in retail spending in 2026, nearly quadrupling 2025. Adobe reported that AI-driven traffic to U.S. retail sites grew 805% year over year on Black Friday 2025. J.P. Morgan estimates agentic commerce could account for 25% of all U.S. online sales by 2030.

On the business-buying side, Gartner projects that 90% of B2B purchases will be handled by AI agents by 2028, with $15 trillion flowing through automated exchanges. I would keep the consumer figures and the B2B figures in separate lines. A company that sells to a procurement team is looking at a different clock than a company that sells a product a shopper can hand to an assistant.

I would fund work that shows up inside the window I can still observe, the 2026 retail figure and the Black Friday traffic jump, and I would label 2028 and 2030 as the horizon those sources named. A horizon is a reason to build the evidence now. It is a poor reason to promise a revenue number the sources did not give me.

What the agent is grading

A human buyer can be moved by a homepage, a demo, and a conversation. An agent evaluating a brand is working from what BCG's July 2026 research calls observable performance: the brand promise has to line up with the experience a third party can see. The tagline is weak material. A press release that restates the tagline is weak material. A specific claim in a publication the agent already retrieves is material it can use.

That changes the order of a marketing plan. The sentence "we are the leading platform in our space" gives the agent nothing to check. The sentence that names a growth figure, a source, and a date can be lifted. I would put the second kind of sentence on the list I am willing to pay for, and I would stop paying to produce the first kind at volume.

The same test covers consistency. If the site, a review page, and a reported article describe three different products, the agent has three partial records. I would rather have one description repeated in three independent places than three campaigns that each sound sharper on their own.

Recency sits in the same test. An old superlative loses to a recent, bounded fact. The Agentic Commerce Protocol, co-developed by OpenAI and Stripe and released as open source in September 2025, is already in checkout flows for ChatGPT and Microsoft Copilot. If the catalog, the price, and the feature list are not in a form that protocol can read, the agent that is ready to buy cannot see the offer. I would ask engineering whether the catalog is readable before I ask marketing for a new narrative.

How I read the CMO survey

The Open Future Forum CMO AI Leverage Report, published August 2026 from 230 respondents in that network, found that 81% of marketing and growth leaders were past exploring agentic AI. The split inside that 81% is the part I would take to a meeting: 36% building agentic AI products, 26% piloting agents in a function or two, 20% running agents in production, and 19% still exploring.

I would not read "81% are already doing this" as a reason to buy a stack. The largest single bucket is building products, which is a different job from making a brand legible to someone else's agent. A founder whose buyer will use an outside assistant needs the evidence layer. A founder whose own product is the agent needs a product roadmap. Mixing those two budgets because they share the word "agentic" is how a quarter gets spent twice.

The useful question under the survey is whether the brand's proof is findable, consistent, and specific enough that an agent can act on it. For a company whose public record is a homepage and a launch post, the survey is a clock, and the record is the work.

The citation ratio I would set next to the forecast

Muck Rack's Generative Pulse, covering more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries, found earned media driving 84% of AI citations and paid or advertorial content at 0.3%. That ratio held across three editions of the study from July 2025. I would tape those two percentages next to the McKinsey range. The range says the channel can be large. The ratio says the channel is reading earned sources, and advertising is close to invisible inside it.

Adobe's figures on the traffic that does arrive are the quality side of the same point: AI-referred visitors spent 48% longer on retail sites, browsed 13% more pages, and converted 42% better than other traffic. I would treat that as a reason to earn the referral, and I would refuse to treat it as a conversion rate I can paste onto a forecast. The sample is retail sites, the comparison is non-AI traffic, and a B2B company still has to measure its own arrivals.

I would use that pair of findings in a specific order inside the company. The commerce forecasts tell me the channel is large enough to deserve a named owner. The citation ratio tells me the owner is wasting the quarter if the work product is another campaign about ourselves. Adobe's 48%, 13%, and 42% figures tell me the sessions that do arrive behave differently, so the cost of being absent is not a missing impression. It is a missing high-intent session whose source was an agent that had already made a short list.

The practical test I would run before the quarter ends has four lines, and each line maps to a failure I have seen companies fund by accident. Search the problem the product solves, in ChatGPT, Claude, Perplexity, and Google AI Mode, and record the sources in the answer. If we are absent, the repair is a citable third-party sentence, not a new homepage module. Put the product facts the Agentic Commerce Protocol would need, catalog, price, and features, where an agent can read them, and treat a human-only PDF as a gap. Ask for citation rate on the buyer questions, and retire any report that can only show impressions. Keep the Machine Relations Index's engine list intact when someone offers to collapse it into one score: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, and Perplexity are different retrieval paths, and a blended number hides which path moved.

The discipline I use to connect a placement to that kind of retrieval is Machine Relations. The name I use for the third-party proof the 84% figure is pointing at is earned authority.

I would keep consumer evidence and business-buying evidence from being averaged into one "agentic" budget. McKinsey's $3 trillion to $5 trillion by 2030 and eMarketer's $20.9 billion in 2026, with Adobe's 805% Black Friday 2025 jump, describe retail. J.P. Morgan's 25% of U.S. online sales by 2030 is still an online-sales figure. Gartner's 90% of B2B purchases by 2028, and the $15 trillion through automated exchanges, describe a procurement agent. A company can face both. It cannot face them with one creative brief. The Open Future Forum split, 230 respondents, 36% building products, 26% piloting, 20% in production, 19% exploring, is how I would stop a leadership team from hearing "81%" and buying a stack. The 84% earned and 0.3% paid split, stable across three Generative Pulse editions since July 2025, is how I would stop them from buying ads and calling the purchase agent-ready.

The decision

I would do three things in the current quarter, and I would keep them inside a named scope.

I would list the ten buyer questions that already have pipeline attached, and I would run those questions, not the brand name, through ChatGPT, Claude, Perplexity, and Google AI Mode. If the brand is absent from the answer, the gap is an evidence gap. I would record the sources the answer did cite.

I would spend the next placement budget on publications those answers already use, and I would require the placement to carry one checkable claim. A mention that cannot be quoted back is a mention I would not count.

I would ask for one measurement: how often the brand is cited when a buyer asks the question the product answers, across the engines the Machine Relations Index already separates, ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, and Perplexity. An impression total does not answer that question.

Scope: this sequence is for a company whose buyer can delegate comparison, shortlisting, or purchase to an agent. It covers both the consumer figures above and the B2B projection, and it stops at the evidence an agent can retrieve. It does not tell a founder to rebuild the product, and it does not transfer Adobe's retail conversion gap onto a different category.

What would change my mind: a later edition of a study with Muck Rack's scope, or a disclosed sample of similar size, in which paid and on-site claims took a real share of citations and the 84% to 0.3% split no longer described the corpus. I would also change the sequence if the Open Future Forum pattern reversed in a later sample of similar size and the majority were still exploring, because the urgency of the evidence work is tied to buyers already delegating. I would not change it because a vendor demo showed a score moving on a prompt I do not sell against.

FAQ

Which of these commerce figures would you put in a board note?

I would put the sourced range and the sourced year next to each figure, on separate lines: McKinsey's $3 trillion to $5 trillion by 2030, eMarketer's $20.9 billion for 2026, Adobe's 805% Black Friday 2025 traffic jump, J.P. Morgan's 25% of U.S. online sales by 2030, and Gartner's 90% of B2B purchases by 2028 with $15 trillion through automated exchanges. I would label the 2026 and Black Friday numbers as observed projections or reported measurements, and the 2028 and 2030 numbers as the horizon those publishers named.

Do shopping agents read the homepage?

They can retrieve a page that is crawlable and specific. BCG's July 2026 research describes the evaluation as observable performance, which puts third-party evidence ahead of a claim the brand publishes about itself. A homepage that repeats a superlative does not give the agent a sentence it can defend. A homepage that states a checkable fact, and is repeated by an independent source, does.

What would you stop funding first?

I would stop funding urgency tactics and self-description at volume, and I would keep funding placements that produce a quotable claim in a source the agent already cites. Muck Rack's Generative Pulse put earned media at 84% of citations and paid or advertorial content at 0.3% across three editions from July 2025. That is the ratio I would use until a comparable corpus says something else.

When would you treat agentic buying as a product problem instead of an evidence problem?

When the company is itself building the agent the buyer will use. In the Open Future Forum sample of 230 leaders, 36% were building agentic products, which is the largest bucket and a different budget from making an existing brand legible to someone else's assistant. I would split those budgets in the plan, and I would keep the evidence work even for a company that is also building, because the outside agents still have to resolve the brand.