Three audiences are leaving the marketing room at the same time, and most CMOs are still optimising for the one that’s already gone.
The first is the human searcher. The second is the human buyer. The third — and this is the one almost nobody is honest about — is marketing’s own imagined audience for itself.
Picture a Q1 marketing review in May 2026. The slide says click-through is down 40% year on year, conversion is up 12%, attribution is “to be investigated.” The CMO opens the floor. Nobody asks the only question that matters: who, exactly, are we marketing to now?
That’s the through-line of 100 weeks of Chronicles of Change, compressed into a sentence. The data on each of the three exits has shifted faster in the last six months than in the preceding two years, so Issue 100 is as good a place as any to put the whole argument on the table.
Where we were, and how quickly
When this newsletter started, “machine customer” was a Gartner footnote and “agentic” had not yet escaped the engineering forums. Today, machine customer is a Mastercard product line, “agentic” is a category in Visa’s annual report, and GEO — for which I just published an operating manual — is a software segment that grew more than 2,000% on G2 in 2025. None of this was supposed to happen this fast. The S-curve I sketched in the early issues — Visibility over Value, then ROI-Driven Deployment, then Embedded Infrastructure — has compressed by roughly eighteen months against my original timing.
Take the three exits in order.
The searcher
The human searcher is the easiest to count. Similarweb’s median zero-click rate on Google AI Overview-triggered queries reached 80% in early 2026, averaging around 83%. Ahrefs tracked position-one click-through for informational keywords without an AI Overview from 7.6% in December 2023 to 3.9% in December 2025 — a halving in twenty-four months. Penske Media’s February 2026 antitrust filing presented internal evidence that AI Overviews had cut click-through by 58% across major properties under its umbrella. HubSpot, until recently the textbook case of content-driven SEO, is estimated to have lost 70 to 80% of organic traffic. Business Insider lost 55%, and cut 21% of its workforce in May 2025.
The under-reported part of this exit is that visibility itself has fragmented. An Averi analysis of 680 million citations across ChatGPT, Perplexity, and Google AI Overviews found that only 11% of cited domains appear on more than one of those three platforms. Superlines documented a 615x variance in citation volume for the same brand across engines. Citation share, in other words, is not a number. It’s a basket. The brand that dominates Perplexity may be invisible on ChatGPT, and a marketing team measuring one of them is choosing — without realising it — to ignore 89% of the surface where their buyers are now getting answers.
Benedict Evans, in his May 2026 deck, drew the comparison I think is the right one. Global mobile data traffic grew more than thirty-fold between 2010 and 2025; over the same period, the global telco index went sideways. Infrastructure rarely captures the value it carries. The same logic now applies to the search engine itself: the carriers of intent are not the owners of value. The value sits with whichever brand is sufficiently legible to be cited inside the answer.
The buyer
The human buyer is leaving more slowly, but with much larger consequences. McKinsey’s QuantumBlack study, published in October 2025, projects $3 to $5 trillion in agent-orchestrated global retail spend by 2030, with up to $1 trillion of that in US business-to-consumer alone. Morgan Stanley estimates that roughly half of online shoppers will use AI shopping agents by 2030, accounting for 25% of their spend. Adobe Analytics measured a 4,700% year-on-year jump in generative-AI traffic to US retail sites between July 2024 and July 2025. Mastercard’s Agent Pay reached every US cardholder in November 2025 and is rolling globally through 2026. Visa’s head of growth says the company expects 2026 to be the mainstream-adoption year. Google’s Chrome browser — north of 70% global share — now supports the Universal Commerce Protocol natively. Amazon’s Rufus went from 65,000 SKUs in its Buy for Me service at April 2025 launch to over 500,000 by year-end, and Amazon’s Q4 2025 disclosure put incremental annualised sales through Rufus at roughly $12 billion.
Pause on the asymmetry. The infrastructure for machine buyers is being built across every layer of the stack simultaneously — payment rails, browsers, retailer agents, and protocol standards — while marketing teams in most enterprises are still trying to work out whether to fund a GEO pilot. By the time the average marketing function has a defensible position in machine-readable commerce, the protocol will have settled. Visa, Mastercard, and Chrome aren’t waiting for marketing to catch up. The protocol conversations are happening without marketing at the table.
There’s one data point that complicates this picture, and it’s worth holding open. OpenAI’s own Wrapped report, published in early 2026, shows that 80% of ChatGPT users sent fewer than 1,000 prompts during all of 2025 — less than three a day, on average, for four-fifths of the user base. Consumer agentic habits exist, but they’re shallow. Bain’s September 2025 consumer-search-preference survey shows that “mostly or always GenAI” is still in single digits across every age cohort except under-30s, where it’s barely into double digits. The audience is leaving — but it’s leaving the old room faster than it’s forming durable habits in the new one. That puts marketing in the worst part of the curve: the metrics for the old buyer are dying, and the metrics for the new buyer aren’t yet stable. The temptation, for any CMO under pressure, is to defer. Deferral, in this environment, is the most expensive decision available.
Marketing’s audience for itself
The third exit is the one I want most marketers to sit with. It’s internal.
MIT’s NANDA initiative published The GenAI Divide in August 2025. Its headline finding — that 95% of enterprise generative-AI pilots show no measurable P&L impact — has been disputed for methodology, fairly in places, but its load-bearing claim has not. More than half of all enterprise generative-AI spend is being directed into sales and marketing tools, while the function with the highest measured ROI is the back office. RAND’s late-2025 meta-analysis found an 80.3% AI-project failure rate, confirmed in slightly different framing by Gartner’s April 2026 report on AI in IT and operations. Gartner’s June 2025 forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027 — for cost, unclear value, or inadequate risk controls — is now standard reading. Gartner also reckons that only 130 of the thousands of vendors claiming agentic capability are legitimate; the rest is what they euphemistically call agent washing.
The Atlanta Fed’s December 2025 CFO survey adds the layer Evans flagged in his deck. The categories where CFOs are seeing AI returns — productivity, decision-making insights — are easy to deploy and hard to measure. The categories where CFOs expect returns in 2026 — cost-saving, new revenue — are hard to deploy and easy to measure. The first is where the spend has actually gone; the second is where the spend was justified. The gap between the two is precisely the gap that destroys most marketing AI cases at next year’s budget round.
What this adds up to is uncomfortable. Marketing is spending the largest share of its GenAI budget on the function with the lowest documented return, in support of a buyer who is delegating the decision to an agent, with metrics that don’t yet exist. The audience marketing imagines it is writing for — a human reading marketing’s output — is the audience that is, in real time, being abstracted away.
A16Z’s enterprise-spending data from March 2026 makes the irony plain: coding alone accounts for around $3 billion of annualised enterprise AI spend; every other category — legal, support, medical admin, search, real estate, finance — sits at half a billion or less combined. Where AI is being paid for, it’s being paid for by engineering. Where AI is being marketed as transformational, it’s being marketed by marketing.
What survives
Here’s what I think survives this, and where the three projects I’ve put into the public over the last twenty-four months actually live.
The CMO who survives the next thirty-six months does three things, and none of them is glamorous. The first is to know which stage of the S-curve the organisation is actually on, not which stage they would like it to be on, and to budget accordingly. Visibility-stage organisations should not be running portfolio-grade governance. Embedded-stage organisations should not be running pilots. Most marketing teams I see are operating at stage one with stage-three expectations.
The second is to manage AI deployments as a capital book — with position sizing, variance, governance load, dry powder, retirement gates, and a do-nothing benchmark — rather than as a tool stack. The 70% of transformations that fail do so because medieval org charts cannot host autonomous technology. The portfolio frame is not metaphor. It’s the operating model.
The third is to architect the brand for machine readability, because the audience that is forming habits the fastest does not read prose. Authority — measured through co-citation, third-party validation, and entity consistency — outweighs schema markup by a factor of roughly 3.5 in driving citation share, but schema is still where most organisations spend their hygiene budget. Get the basket right.
Evans’s other useful provocation in his deck applies here. He asks what becomes possible when something that used to be expensive becomes free. “Listen to one customer call and tell me if the customer sounds strange” was a 1990s task. “Listen to a million customer calls a day and tell me what you notice” is now a Tuesday. The marketing functions that get this right will not be doing the same job with fewer people. They will be doing a different job, against a buyer they previously could not address.
Seven bets
Seven things I’m willing to be wrong about in public by the end of 2027.
By the end of 2026, machine-readable product catalogues will be a board-level discussion at every Fortune 500 retailer, and the phrase “API quality” will appear in at least three CFO earnings calls.
Citation share will appear as a standing metric on a Fortune 500 CMO’s quarterly scorecard before the end of 2026, sitting alongside organic ranking and share of voice.
Pay-per-crawl, Cloudflare’s default-block since 1 July 2025, will be a standard 2027 enterprise marketing line item, alongside paid search.
The 50/30/20 GEO budget split — organic GEO, merchant-feed engineering, paid AI placement — will become the default reference, and roughly 30% of marketing spend will be quietly reclassified as supply-chain work.
ChatGPT-cited brands will trade at a measurable premium on B2B procurement RFPs by year-end 2027. Procurement teams will start asking vendors for citation share as a qualifying signal.
Portfolio-management experience will start appearing as a preferred qualification in Fortune 500 CMO job descriptions by 2028, the way “digital transformation” appeared in postings a decade earlier.
The first agent-to-agent commerce protocol will move from pilot to production in B2B procurement before the end of 2026 — and it will be Stripe, Visa, or Mastercard who ships it, not the AI labs.
If your CFO asked you tomorrow what proportion of your marketing spend was going to a buyer who no longer exists, what would your honest answer be? And which of the three audiences is your marketing organisation still primarily optimised for — the one that already left, the one that’s leaving, or the one that’s quietly forming new habits while you’re not looking?
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.