The Demand-Side Disruption

The funnel collapses into an answer

For two decades the marketing funnel rested on a referral economy: a user typed a query, a search engine returned a ranked list, and the click that followed was both the unit of attribution and the unit of value. That mechanism is eroding from the top. SparkToro’s clickstream analysis found that in the first four months of 2026, 68.01% of U.S. Google searches ended without any click to the open web, up from roughly 45% in 2016 and 60.45% across full-year 2024³.

The drift toward zero-click predates AI, and the mechanism behind it is visible in the click data. SparkToro’s 2024 study, run on a separate clickstream panel, found that clicks to Google-owned properties already accounted for roughly 30% of all U.S. Google clicks, and that only 360 of every 1,000 U.S. searches sent a click to the open web⁴. Google was already keeping a large share of the clicks for itself before AI summaries existed. Conversational summarization is steepening a curve the platform had already bent.

How much steeper is now measured. Pew Research Center instrumented the browsing of 900 U.S. adults across 68,879 Google searches in March 2025, the most direct measurement in this corpus of what an AI answer does to the click. When an AI-generated summary appeared, users clicked a traditional result link in just 8% of visits, against 15% when no summary was present, and clicked a link inside the summary itself only 1% of the time⁵. The summary also ended sessions. Users ended their browsing session after 26% of results pages carrying a summary, against 16% of pages with only traditional results⁵. When the answer satisfies the query, the click that carried attribution never fires, and the session-abandonment gap is that failure measured directly. The other studies confirm the pattern rather than extend it. SparkToro’s panel finds AI Overviews on more than 20% of searches, cutting click-through by “nearly 60%“³, and Gartner had forecast the direction as early as February 2024, predicting traditional search volume would fall 25% by 2026 as generative AI became a “substitute answer engine”².

The behavior shift has the scale to matter. More than 20% of Americans now use an AI tool ten or more times a month³, and Stanford’s 2025 AI Index reports that generative AI reached mass adoption faster than the PC or the internet²⁰. The conversational interface is mainstream, and discovery is migrating into it.

From keyword SEO to Generative Engine Optimization

If the surface is now a generated answer, the optimization target changes with it. The foundational treatment is Aggarwal et al., “GEO: Generative Engine Optimization,” accepted to KDD 2024, which formalizes “generative engines” — systems that synthesize and summarize across multiple sources — and introduces GEO as “the first novel paradigm to aid content creators in improving their content visibility in generative engine responses”¹. The authors name the asymmetry that makes the discipline necessary: given “the black-box and fast-moving nature of generative engines, content creators have little to no control over when and how their content is displayed”¹.

Their empirical result inverts traditional SEO instinct. Measured against two visibility metrics — Position-Adjusted Word Count and an LLM-judged Subjective Impression — the best methods improved visibility by up to 41% and 28% respectively over baseline¹. The tactics that worked were adding quotations, statistics, and cited sources, and writing fluently with authoritative framing. The tactic that actively hurt was keyword stuffing, the load-bearing move of legacy SEO, which reduced Position-Adjusted Word Count by 8.7%¹. Effectiveness also varied by domain, and the gains accrued disproportionately to lower-ranked sources. A rank-5 source could gain over 100% visibility from adding citations, while a rank-1 source could lose visibility from the same edit¹. Generative surfacing, in other words, partly resets the incumbency advantage that PageRank entrenched.

The practical implication is that reputation becomes the ad strategy. What an LLM can quote, cite, and corroborate about a brand (third-party reviews, structured data, authoritative mentions) determines surfacing more than purchased keywords do. Gartner’s read of the post-search environment points the same way: “content utility and quality still reigns supreme,” with rising emphasis on authenticating high-value content through E-E-A-T signals (Experience, Expertise, Authoritativeness, Trustworthiness)².

The same power invites capture. A 2026 position paper, Wen et al., “Generative Engine Optimization Creates Underexamined Risks,” argues that because GEO targets “LLM answer engines’ evidence pool and generation,” it concentrates influence over what gets surfaced, invites undisclosed commercial influence embedded in the answer’s reasoning, and outruns existing SEO governance¹⁷. The authors call for “answer-level governance and measurement: stronger contestability, high-precision disclosure, black-box auditing of material influence”¹⁷. If visibility is gameable and concentrated, the answer layer can be captured, and a captured answer, unlike a sponsored link, does not announce itself.

Those two claims sit in tension, and the tension is the real state of the discipline. Reputation functions as the ad strategy only while the answer layer surfaces reputation honestly. If commercial influence can be embedded in the evidence pool without disclosure, then what reads as earned authority may be purchased placement, and reputation collapses back into a buyable signal. The practical reading holds both halves. Building citable, corroborable authority is what the measured evidence rewards today¹, and whether it stays rewarded on the merits depends on whether the disclosure and auditing Wen et al. call for actually arrive¹⁷. Until then, a brand investing in reputation is betting that the answer layer stays honest enough for reputation to remain legible.

The data goes dark, and a new economic model emerges

The funnel’s collapse has a second-order casualty: the referral data that publishers and merchants used to see. When the answer satisfies the query, the click that carried attribution never fires. Digital Content Next measured 19 member publishers over eight weeks in May–June 2025 and found median year-on-year Google Search referral traffic down 10% overall — 7% for news brands, 14% for non-news — with site-level losses reaching 25% and weekly declines outpacing gains two to one¹⁵. DCN’s CEO Jason Kint called the data “ground truth” about AI Overviews’ impact, contrasting it with the platform’s quality claims¹⁵.

The responses to the broken referral contract point in different directions. Cloudflare, which says it protects roughly 20% of the web, became, as of July 1, 2025, “the first Internet infrastructure provider to block AI crawlers accessing content without permission or compensation, by default,” and launched a Pay Per Crawl marketplace letting publishers charge AI companies for access and requiring crawlers to declare whether they are training, inferencing, or searching¹⁴. CEO Matthew Prince framed it as survival: “If the Internet is going to survive the age of AI, we need to give publishers the control they deserve and build a new economic model that works for everyone”¹⁴. Perplexity took the inverse approach a year earlier with a Publishers’ Program that shares advertising revenue with cited publishers (reportedly keeping around 20% and routing around 80% to participants), launched with TIME, Der Spiegel, Fortune, and others¹⁶. One model gates extraction at the network edge; the other pays for citation inside the answer. Both concede the same premise, which is that the old free-traffic-for-indexing bargain is over.

A third response sits with the regulators. In the same DCN report, Kint noted publishers could gain relief through proposed DOJ remedies requiring Google to separate its AI crawler from its search crawler, which would let a publisher opt out of AI training without losing search visibility¹⁵. Pricing, revenue share, and structural separation are now all on the table, and each is a bid to re-price content the answer layer has been consuming for free.

The transaction itself becomes agentic

The same logic now reaches past discovery into the purchase, and that step deserves to be argued rather than assumed, because no source in this corpus measures whether agents select products the way answer engines select citations. The inference runs through the shared architecture. An agent that buys must first discover and choose, and McKinsey’s read of that choice — “discoverability by agents becomes the new strategic battleground”⁸ — is the corpus’s closest statement that agent selection will reward the same machine-legible authority GEO measures. This paper carries the link as a supported inference; nothing in the corpus measures it. What is documented is the rails. Across 2025–2026 the payments and platform incumbents built the infrastructure for agent-led checkout, and the builds share one pattern: bind a tokenized credential to a specific agent, merchant scope, and consent policy, so the model completes checkout without ever holding a raw card number.

The best-documented build is OpenAI and Stripe’s Instant Checkout, shipped in September 2025, which lets U.S. ChatGPT users buy in chat through the open-sourced Agentic Commerce Protocol (ACP), using “Shared Payment Tokens” so users transact “without exposing payment credentials to ChatGPT”⁹¹⁰. It launched with Etsy sellers, with “over one million Shopify merchants” to follow¹⁰. The payment networks and Amazon built to the same token pattern, each with a different center of gravity. Mastercard’s Agent Pay (April 2025) binds its “Agentic Tokens” to “a specific agent, a specific merchant scope, and a specific consent policy”¹⁹. Visa’s Intelligent Commerce (June 2026) adds the trust layer above the token — an Agentic Directory of verified agents and merchants, and an Agent Score measuring whether an agent can navigate a merchant’s site¹¹. Amazon’s “Buy for Me” (April 2025) pushes furthest into the open web; its agents find products Amazon does not sell and complete checkout on the brand’s own site by securely passing the customer’s encrypted details¹².

Demand is already flowing through the front of this funnel while the back end stays human. Adobe Analytics, examining over a trillion visits to U.S. retail sites, found generative-AI referral traffic up 1,200% in February 2025 versus July 2024, with 39% of surveyed consumers having used generative AI for shopping and 53% planning to⁶. By the 2025 holiday season that traffic was up 693.4% year-on-year against a record $257.8 billion in U.S. online spend⁷. AI-referred shoppers behaved like high-intent researchers, spending 8% more time on site and viewing 12% more pages with a 23% lower bounce rate, yet in the February 2025 data they were 9% less likely to convert⁶. The gap between the browsing and the buying is the finding. Discovery had moved to AI faster than the purchase had.

The trust gap that gates the shift

The binding constraint on that last step is consumer trust. Forrester’s March 2025 Consumer Pulse Survey found that only 24% of U.S. online adults trust AI agents to act on their behalf for routine purchases, even as 43% agree that in an agent-mediated future “brands will market directly to these agents”¹³. By mid-2026 Forrester’s own read was that “hype is running ahead of behavior”: most “agentic” experiences remain conversational, “humans still drive decisions and checkout in most cases,” true autonomy is rare, and the market is “visibly volatile,” with major players’ features “appearing and disappearing as they learn”¹⁸.

The payments incumbents are betting on a shape from history — a capable mechanism stalls on consumer trust, then crosses a threshold once trust infrastructure matures, as e-commerce did with escrow, ratings, and buyer protection. Mastercard’s verifiable tokens, Visa’s directory and Agent Score, and Stripe’s shared payment tokens are that infrastructure, aimed at the 24%-to-majority gap¹¹¹⁹. The e-commerce comparison is an analogy rather than a forecast. Whether and when the gap closes is something the current data cannot tell us.

Who owns the customer and the data

When an agent stands between the customer and the brand, it captures the moment of choice, and with the choice goes the data that defines the relationship. That disintermediation is the stake underneath all the rail-building, and the architecture each player ships reveals where it intends to sit. OpenAI’s design is studiously neutral on the surface. Instant Checkout charges merchants a small fee, is free to users, “doesn’t affect their prices,” and explicitly states Instant Checkout items “are not preferred in product results”⁹; merchants “retain full control over what’s sold, how their brand shows up, and how orders are fulfilled”¹⁰. Amazon’s “Buy for Me,” by contrast, inserts Amazon’s agent into transactions on rival brands’ sites, and it drew an immediate backlash from businesses that objected to being listed without permission¹². The fault line is whether the agent is a neutral conduit or a new gatekeeper.

The size of the prize explains the maneuvering. McKinsey estimates AI agents could orchestrate as much as $1 trillion in U.S. retail revenue by 2030, roughly 30% of projected B2C revenue, and $3–5 trillion globally, reframing shopping from discrete steps into a “continuous, intent-driven flow”⁸. For an incumbent whose moat was the customer relationship and its data exhaust, an intermediating agent is both the largest distribution opportunity on offer and the most direct threat to that moat. The question every brand now faces is whether optimizing for the agent means renting a customer it used to own.

Measurement erodes first, and the timeline stays open

The nearest consequence is to measurement. Attribution and free referral traffic — the measurement substrate of digital marketing — are degrading together. The majority of searches now end without the click that carried the signal³, AI summaries halve the clicks that remain⁵, and publishers’ referral logs are already recording the loss¹⁵. The dashboard goes quiet before demand does; Adobe’s data shows the demand still arriving through AI referrals even as conversion lags⁶. So the disciplines built on the click (attribution, keyword bidding, referral analytics) are losing their instrument while the behavior they measured moves somewhere the instrument cannot see.

What the corpus does not yet support is a timeline. The traffic data is real and steep; the agentic rails are shipped and named; the trust gap is measured. But whether autonomous purchasing crosses from the 24% who trust it today¹³ to the majority, and on whose terms the customer relationship ends up, remains on the present evidence an open question rather than a settled trajectory¹⁸.

References