The Delegated Buyer

The Delegation Thesis

This study examines the demand side of machine-mediated commerce. Its companion, the Machine-Readiness Atlas, measured the supply side: how much of the world's small-business web a machine can actually read. This report asks the mirror question. As consumers and organizational buyers hand discovery and selection to answer engines and conversational assistants, what are they handing over, how far does the handoff go, and how deep is the dependency becoming? The two studies are meant to be read together, because the commercial consequence of machine mediation only appears where a rising demand-side behavior meets a supply that a machine can or cannot assemble into an answer.

The central argument is that delegation is not a novel behavior produced by generative AI. It is the newest expression of a constraint the decision sciences described more than fifty years ago. Herbert Simon's account of bounded rationality holds that no real decision-maker has the time, information, or mental capacity to weigh every option and select the optimum, so people satisfice: they accept the first option that clears a threshold of good enough and stop. A synthesized machine answer meets a buyer already predisposed to stop and gives that buyer a cleaner reason to stop sooner. Read this way, delegation is bounded rationality operating at the scale of a market, and generative AI is the surface on which it has become measurable.

To organize the evidence without overclaiming, the study uses a proprietary lens, the Delegation Ladder. It distinguishes four ascending rungs by how much of a decision a buyer hands to a machine. On Rung 1, Consult, the buyer asks a machine for information but retains the judgment. On Rung 2, Shortlist, the buyer lets the machine assemble the consideration set and accepts the field it presents. On Rung 3, Decide, the buyer accepts the machine's single synthesized recommendation without independently comparing alternatives. On Rung 4, Transact, the buyer authorizes the machine to complete the purchase. Ascent up the ladder is not uniform. It is governed by two gates: a stakes gate that slows the climb for high-consequence and credence-good decisions, where buyers retain verification behavior longer, and a trust gate that is sharpest between Decide and Transact, where willingness to let a machine finish a purchase collapses even among heavy users.

The ladder reframes the headline adoption numbers. Reach measures how many buyers stand on Rung 1, but the commercial consequence lives on Rungs 2 and 3, where the answer removes the visible list of losing options and anchors the choice on whoever is named first. The model is descriptive and testable, not a claimed law, and every rung in this study is measured against dated evidence and our own first-party audits.

The table below sets out the ladder and what governs movement up it.

The evidence standard for this study is stated at the outset. This is a synthesis of dated, attributed third-party research and Raveneye Global first-party audits, not a census. Every figure carries an evidence tier: established where independent sources converge, emerging where a figure is recent or single-source, and contested where credible parties disagree on magnitude. The central asymmetry the study returns to is that the supply floor is measured directly while the demand shift, though real and fast-moving, is documented through instruments that still disagree on exact size. The direction of travel is well corroborated; the precise magnitude for any single market must be measured, not inferred.

Reach measures how many buyers stand on the first rung, but the commercial consequence lives higher up, where the answer removes the visible list and anchors the choice on whoever is named first.

Adoption and Usage: Reach Faster Than Any Prior Cycle

The reach of AI answer engines has expanded faster than any prior consumer information technology, and the pace is now documented across independent instruments rather than vendor marketing. OpenAI reported that ChatGPT reached 800 million weekly active users by October 2025, announced by Sam Altman at DevDay, and crossed 1 billion global monthly active app users by June 2026. Google's answer layer scaled even more widely by virtue of default distribution: Alphabet reported AI Overviews at over 2 billion monthly users by mid-2025, and its conversational AI Mode passed 1 billion monthly active users following an October 2025 global expansion. On the population side, Pew Research Center places US ChatGPT use at 44 percent of adults in 2026, up from 18 percent in 2023, and syntheses of Stanford HAI and NBER data estimate that generative AI reached about 53 percent of the world's population within roughly three years of launch, a diffusion pace steeper than the personal computer or the internet in their own eras.

A single caveat frames every one of these figures. Authoritative US surveys differ by more than twenty percentage points depending on definition and instrument, so no single adoption number should be quoted without its source. The table below sets the principal instruments side by side, not to average them but to show that they measure different things: weekly versus monthly users, one product versus a whole category, a claimed behavior versus a tracked one.

The behavioral question, how far this reach has migrated into actual decisions, is where the demand-side picture sharpens. On search, the answer increasingly replaces the visit. Analysis of SparkToro and Similarweb clickstream panels found 68.01 percent of US Google searches ending with no click in early 2026, up from 60.45 percent two years earlier, and the effect concentrates precisely where a synthesized answer appears. Pew Research Center's tracked-browsing study of 900 US adults across 68,879 real searches is the strongest single anchor: when an AI summary was present, users clicked a traditional result in only 8 percent of visits against 15 percent otherwise, and just 1 percent clicked a source cited inside the summary itself. A pre-registered field experiment by Agarwal and Sen independently measured a 38 percent drop in organic clicks and a rise in the zero-click rate from 54 to 72 percent on affected queries. The convergence of a consumer-behavior researcher, a clickstream panel, an SEO vendor, and an academic experiment on the same directional finding makes the click collapse on answer-bearing queries the most established result in the study.

The migration into buying decisions is now measurable on the demand side, though still largely through proprietary methods. NielsenIQ found that 42 percent of US consumers used at least one AI tool to shop within the prior month in early 2026, and an Envision Horizons survey reported 63.1 percent had used AI for product research, up from 49 percent a year earlier. At the local level, BrightLocal recorded use of AI platforms for business recommendations rising from 6 to 45 percent of surveyed consumers in a single year, becoming the third most-used recommendation source, while Google's recommendation share fell from 83 to 71 percent. On spend, Adobe Analytics measured AI-referred traffic to US retail sites growing about 693 percent year over year during the 2025 holiday season with a 31 percent conversion lift. These figures establish that a real and large redistribution toward the answer is underway.

Two disciplined qualifications protect the strand's credibility. First, usage is not trust and is not resolved market share: the Reuters Institute found weekly AI-chatbot news use at only 10 percent globally, with trust at 20 percent against 37 percent for news overall, and reported AI-search share figures still vary roughly five to ten fold by denominator, from about 12 percent of US queries to 92.4 percent of trackable large-language-model referral traffic. The behavior therefore sits in a band on the diffusion curve, past the early adopters and entering but not saturating the early majority, rather than at a single point. Second, the aggregate does not describe any single business. Every dataset here measures publisher panels, national consumer surveys, or keyword databases; none measures foot traffic, calls, or bookings for an individual local operator. The defensible conclusion is that adoption of AI answers has moved faster than any prior cycle and has begun resolving inside the answer for a growing share of research and buying decisions, while the exact magnitude for a given market remains something a business must measure directly.

The Psychology of Delegation

The delegation of buying decisions to machines is best understood not as a technology story but as a behavioral one, because the mechanisms that drive it were described long before generative AI existed. The foundation is Simon's bounded rationality. People satisfice rather than maximize, accepting the first option that clears a threshold of good enough, and, in Simon's own framing, a wealth of information creates a poverty of attention. Pew Research Center's browsing panel supplies the behavioral fingerprint: with an AI summary present, session abandonment rose from about 16 to about 26 percent, click-through to any result fell from 15 to 8 percent, and a source cited inside the summary drew a click only about 1 percent of the time. A synthesized answer does not create the disposition to stop; it arrives to a buyer who already had it and lowers the cost of acting on it.

The mechanism that makes delegation feel costless is cognitive offloading, the transfer of memory and effort to an external store. Sparrow, Liu and Wegner's 2011 Science study established that when people expect future access to information, they recall the information itself less well and instead encode where to find it, treating the network as a form of transactive memory. Recent work extends the concern from recall to reasoning: a 2025 survey of 666 participants reported a negative correlation between frequent AI-tool use and critical-thinking scores mediated by offloading, and an MIT Media Lab EEG study by Kosmyna and colleagues found that essay writers using a language model showed the weakest neural connectivity across conditions and that 83 percent could not quote a sentence they had just produced. Both later findings are held at emerging tier, being single-source, self-reported or preprint, but the direction is consistent: offloading lowers the felt cost of accepting a machine's output without independent checking.

How much credit a buyer extends to the machine is genuinely split rather than settled. Dietvorst, Simmons and Massey documented algorithm aversion, the tendency to lose confidence in an algorithm faster than in an equally fallible human after seeing it err. Logg, Minson and Moore documented the opposite, algorithm appreciation, in which people weight advice more heavily when it is attributed to an algorithm, strongest for objective, quantifiable judgments. The reconciling frame, held here at emerging tier because it rests on synthesis rather than a single controlled test, is that aversion dominates for high-stakes, subjective decisions while appreciation dominates for low-risk, objective, repeatable ones. This split has direct commercial force: it predicts that a buyer will delegate a restaurant or local-service choice far more readily than a high-consequence medical or financial one, and our Trust in the Machine study finds exactly that asymmetry in the field data.

Automation bias is the specific failure mode of over-delegation, and it is among the better-documented effects in human-factors research. Skitka, Mosier and Burdick distinguished commission errors, following an automated directive despite contradictory evidence, from omission errors, failing to act because the aid did not flag a problem. Parasuraman and Manzey's 2010 review in Human Factors established that automation complacency appears in novices and experts alike, is driven by how attention is allocated under load, and cannot be reliably trained away. Applied to answer engines, this literature reframes the first-answer problem: the risk is not that buyers consciously decide a machine is trustworthy, but that a plausible synthesized answer suppresses the verification behavior that would surface its errors.

The final mechanism is anchoring. Tversky and Kahneman showed that the first value a person encounters exerts disproportionate pull on later judgment even when it is known to be irrelevant. Our Answered Once, Chosen Once model argues, as an explicit extrapolation rather than a measured law, that when an engine returns one named business instead of ten links it does not remove this choice architecture but concentrates it: the single answer becomes an anchor with no visible competitor beside it. The inference deserves its caution, because the anchoring findings were built on lists and prices, not single-answer AI surfaces, and the choice-overload meta-analysis by Scheibehenne, Greifeneder and Todd, which found the mean effect statistically indistinguishable from zero, is a standing warning that confident universal claims about choice tend to break. The synthesis is that delegation concentrates rather than eliminates the biases that govern choice, and that the correct response is to measure standing on the answer surface directly rather than to assert it.

Consequences for Buying Choices

Machine-mediated discovery does not merely change how buyers find businesses; it changes which businesses are eligible to be chosen at all. The starting condition is often misread. Buyers were never comparing the full shelf: field research on real consideration sets found people seriously weighing only 3.3 to 4.0 brands per category even where dozens were available, as Hauser and Wernerfelt reported in the Journal of Consumer Research in 1990, a ceiling anchored in bounded rationality and in the working-memory limits Miller described in 1956. What a synthesized answer removes is not the size of the shortlist but its visibility. The buyer shifts from picking among a visible list to accepting or rejecting a single handed-down recommendation, and the gap in the evidence is that no published study yet counts how many alternatives an engine names per query, category by category.

The collapse is compounded by a citation gap and by concentration inside whatever list survives. Where anyone has measured it, generative engines name local businesses far less often than Google's map pack: SOCi's 2026 Local Visibility Index found ChatGPT recommending a given location 1.2 percent of the time and Gemini 11 percent, against 35.9 percent for the local three-pack, and Local Falcon found 83 percent of restaurants entirely invisible to ChatGPT. These local-visibility figures are held at contested tier, drawn from commercial tools running non-standardized samples, but two independent reads point the same way. Visibility that is granted then concentrates heavily, so the practical field a buyer sees is narrower than the citation rate alone suggests.

A newer and underappreciated mechanism is anchoring on the first named option. A growing body of algorithm-audit work shows that large language models exhibit strong position bias. In the Fragile Preferences study by Yin, Vardi and Choudhary, position effects reached 92.8 percent in resume comparisons and favored the first high-quality option as much as 98 percent of the time. For a business, this means that being named first inside an answer is not a cosmetic advantage; it can be close to the whole decision. The finding is held at emerging tier, because it is measured on model behavior in controlled tasks rather than on live buyer outcomes, but it converges with the local-concentration data toward the same conclusion: the answer both shortens the list and orders it in a way that rewards the top-named entry disproportionately.

On the business-to-business side, the ratio of self-directed research to sales contact has inverted, and that research now runs through an AI layer. Forrester reports that 94 percent of business buyers used AI in their most recent purchase, up from 89 percent, and that buyers are twice as likely to name generative AI as a more meaningful source than vendor websites or sales reps. G2's 2026 Answer Economy report found 54 percent naming AI chatbots as the single largest influence on their vendor shortlist, ahead of review sites at 43 percent, though several of these B2B figures rest on single-vendor commissioned surveys and are tiered contested accordingly. Semrush ranks agencies and service providers as the single most AI-researched B2B category. The consistent implication is that a prospect can form a firm opinion of a business before any human on the seller's side knows the buyer exists, and that the deciding evidence must already sit where a machine can retrieve it.

Two consequences bracket the funnel. At the top, machine mediation narrows the field before the buyer sees it and anchors the choice on whoever is named first. At the bottom, once a buyer reaches checkout, the same honesty-and-sequencing logic governs whether the sale closes. The Baymard Institute puts cart abandonment near 70.22 percent across 50 pooled studies, driven not by price or weak persuasion but by late cost disclosure, about 39 percent, and forced account creation, about 24 percent. The through-line is that the businesses that get chosen are increasingly those that are retrievable, first-named, and clean at checkout, rather than merely well-ranked.

Being named first inside an answer is not a cosmetic advantage. Where the machine returns one recommendation instead of ten links, it can be close to the whole decision.

Segmentation: An Uneven Migration

The delegation of discovery and buying to machines is not a single, uniform migration. It is a set of overlapping shifts that move at different speeds, in different directions, for different buyers, and segmentation is the only honest unit of analysis. Four axes structure the unevenness: the buyer's generation, the category of good being bought, whether the purchase is a consumer or an organizational decision, and the digital maturity of the market. On every axis the delegation is real, but its shape and its measurability change so much that a single headline adoption figure conceals more than it reveals.

Generation is the strongest predictor of baseline adoption, but the relationship inverts under decision stakes, and that inversion is the finding most likely to mislead a strategist working from intuition. Pew Research Center's 2026 survey shows chatbot use falling monotonically with age, from 66 percent of adults 18 to 29 to 23 percent of those 65 and older. Yet the picture reverses for high-consideration, credence-good decisions: iLawyer Marketing's 2026 survey found that 57 percent of consumers aged 45 to 60 would use ChatGPT to research a lawyer, against only 35 percent of those aged 18 to 29, and overall willingness to use ChatGPT for that task rose from 9 percent in 2023 to 41.9 percent in 2026 while stated Google use for it fell 14.8 points in a single year. Younger cohorts lead the habit; older, higher-purchasing-power cohorts lead its application to consequential spending. Generation predicts whether a person uses these tools at all, not whether they use them to choose a surgeon, a lawyer, or a contractor.

Category is the axis where the economics are clearest and the delegation most divergent. The search, experience, and credence trichotomy set out by Nelson in 1970 and Darby and Karni in 1973 explains why. Search and experience goods, most consumer retail, delegate to machines quickly and measurably, which is what the Adobe holiday figures capture. Credence goods, legal, medical, dental, and aesthetic services, whose quality a buyer cannot verify even after purchase, show slower and trust-gated delegation, because the AI answer has not closed the old verification gap so much as relocated it, and the answers themselves remain unreliable in exactly these fields. The Stanford RegLab study of purpose-built legal research tools found roughly a third of queries hallucinating. The result is two categories moving in opposite ways: in retail the machine is increasingly the transaction surface, while in credence markets it is a first, partly distrusted opinion that a buyer still cross-checks, when the fading habit of a second source allows.

The business-to-business axis is where delegation is most collective and most invisible to the seller. The deciding research is done by a group, typically six to thirteen stakeholders, before a vendor is ever contacted. Gartner reports that 67 percent of B2B buyers now prefer a rep-free experience, Forrester found 94 percent used AI in their most recent purchase, and G2 found a majority beginning research in an AI chatbot rather than Google. Unlike the consumer case, where a delegated decision belongs to one person, the B2B delegation is distributed across a committee that never announces itself, so the visible surface has to satisfy many independent functional concerns at once, and no single relationship can compensate for a machine-readable presence that fails any of them.

Market maturity produces the most structurally distinct form of delegation, because in mobile-first and conversational-commerce markets the machine consulted is often a private messaging thread rather than a public answer engine. In India, the great majority of online adults message a business weekly and hundreds of thousands of firms pay to advertise into that chat, yet the catalogs and threads where this commerce happens are not crawlable at web scale, and platform censuses that omit WhatsApp miss it entirely. Latin America shows the same conversational pattern with sharp country-level variation, Instagram reaching 69 percent of Brazil's population against 40.5 percent of Mexico's, so a single regional assumption fails. These markets are delegating discovery to machines as much as any, but through surfaces the generative answer layer largely cannot read, which means the visibility problem in a conversational market is not the same problem as in a search-first one.

The Supply-Demand Mirror

The central argument of this study rests on a mirror. On the demand side, consumers and business buyers are handing discovery and selection to machines at a pace that now shows up in behavioral data rather than survey sentiment alone. On the supply side, the businesses those machines are meant to name remain, in the great majority, unreadable to them. Raveneye Global has measured both halves as first-party research, and the two halves fit together with unusual precision.

The supply-side anchor is the Machine-Readiness Atlas of July 2026, which applied one fixed method across 205,103 independent small businesses in six economies and found that the share fully machine-readable ranges roughly 7.6-fold, from 32.1 percent in the United States down to 12.7 percent in Europe, 7.3 percent in Japan, and about 4.6, 4.5, and 4.2 percent in Southeast Asia, Africa, and India. Because every business sampled already held an active Google Business Profile, each figure is a floor: the true share a machine can read is lower still. The single sentence that carries this study is that as buyers delegate more to machines, most of the supply they would delegate over cannot be assembled into an answer. The table below sets each economy's machine-readable share against its web-presence rate, the two supply filters a demand signal has to pass through before a business can be named.

The demand side is documented across our behavioral studies with equal specificity. The clearest instrument is the click: Pew found that with a Google AI summary present, users click a traditional organic result in only 8 percent of visits against 15 percent without, and end the session outright 26 percent of the time versus 16. Our Consideration-Set Collapse study connects this to a settled behavioral floor, the three-to-four-brand shortlist, and shows that what the engine changes is not the size of the shortlist but who assembles it and whether the losing options are ever seen. The delegation is now visible in money as well as attention: Adobe recorded AI-referred retail traffic up about 693 percent year over year across the 2025 holiday season, with those shoppers converting about 31 percent better than other traffic.

Where these two bodies of evidence meet is the thesis. Every demand signal presumes a readable supply the machine can name. The Atlas shows that supply is mostly missing, and missing cheaply: the human-facing basics, mobile-friendliness and security, are near-universal in every economy, while the machine-facing finish, a few lines of local-entity markup, is skipped almost everywhere. That combination, high and rising importance, low cost, and low prevalence at once, is the profile of a large and durable advantage for the businesses and markets that close the gap first.

One caution travels with the material and is stated plainly. Machine-readability is eligibility, not ranking. Google Search Central is explicit that structured data enables richer features and page understanding but does not on its own lift position, and the evidence that it directly increases AI-answer citations is not settled. The claim the study defends is therefore narrow: a business a machine cannot read cannot be named. That is a necessary condition, not a sufficient one. The asymmetry that makes the mirror the durable backbone of the study is that the supply-side gap is measured rather than modeled, while the demand shift, though real and corroborated across independent sources, is still moving in magnitude.

Counter-Evidence and Limits

The strongest single fact behind the delegation thesis is also its most contested one. Pew Research Center's browsing-panel measurement, the study's backbone finding, showed click-through to a traditional result falling from 15 to 8 percent the moment a Google AI summary appeared. That measurement is behavioral rather than surveyed, but it is disputed by the party with the most to lose. Google publicly rejected the study as flawed methodology and a skewed queryset not representative of Search traffic, and noted that the design compared queries captured in one month against re-run searches in a later month. The wider industry does not dispute the direction, only the magnitude, and the magnitude estimates diverge widely, from roughly 15.5 percent at one end to about 61 percent at the other, a spread too large to state as a single number. The commercial counterpoint sharpens the caution: Alphabet reported first-quarter 2026 Search revenue up 19 percent year over year to 60.4 billion dollars and stated that AI features are driving Search growth, which sits awkwardly against any story of imminent search collapse.

Agentic checkout is the clearest case where the 2026 reality undercut the promise. Walmart, which offered roughly 200,000 products through OpenAI's Instant Checkout, reported that in-chat purchases converted at about one third the rate of click-outs to its own site, and moved toward a discover-in-AI, buy-on-your-own-site model. Consumer trust data explains why. Checkout.com's June 2026 six-market survey found 27 percent of consumers trusting no organization to operate an AI shopping agent and 24 percent saying they will never delegate a purchase to AI, with willingness concentrated in low-stakes categories and collapsing in consequential ones. These agentic-checkout conversion figures are held at contested tier. The reading is not that delegation is failing but that it is bifurcating: assistants are becoming a genuine discovery surface while autonomous checkout remains a narrow, trust-gated behavior. Even bullish forecasts place agentic commerce at 15 to 25 percent of US e-commerce by 2030, and eMarketer projects AI platforms directly transacting about 1.5 percent of US e-commerce in 2026, rising to 8.8 percent by 2029.

Credibility requires holding the counter-counter-evidence too. Adobe Analytics reported AI-driven traffic to US retail sites up about 393 percent year over year in the first quarter of 2026, and a conversion reversal in which AI-referred traffic converted about 42 percent better than other sources by March 2026, having converted roughly 38 percent worse a year earlier. That is real, measured evidence that delegated discovery is maturing into delegated buying at the margin. The defensible synthesis is that delegation is growing fast from a small base, converting better where it is trusted, and stalling precisely where the stakes, the money, or the accuracy risk are highest.

Accuracy is the mechanism that keeps a human in the loop, and here the external evidence is unusually strong. The European Broadcasting Union and BBC study published in October 2025, the largest of its kind, examined 3,000 responses from four leading assistants across 14 languages and 22 public-service media organizations in 18 countries, and found that 45 percent of responses contained at least one significant issue and 81 percent had some problem, with errors independent of language or territory. The reception data tracks the accuracy data: the Reuters Institute Digital News Report 2026 found weekly use of AI chatbots for news at only 10 percent globally, trust in chatbot answers at just 20 percent against 37 percent for news overall. People are using these systems while actively distrusting them, which is a brake on full delegation, not an accelerant.

Two limits bound any claim of universal delegation: population and measurement. Delegation is unevenly distributed. Pew found that while 49 percent of US adults use AI chatbots, 77 percent of those 65 and older never do, and adoption skews toward higher-income households, so aggregate figures overstate the typical buyer. And the measurement infrastructure to settle these debates does not yet exist. There is no Nielsen-grade, audited, methodologically transparent panel for AI answers; visibility is estimated by a proliferation of commercial tools running small, non-standardized query samples. Every figure in this study, including the ones that support the thesis, should therefore be read as directional evidence from an unmeasured frontier rather than as settled national accounting.

Implications: Measure, Do Not Infer

The practical conclusion follows from the study's central asymmetry. The direction of travel is well documented: buyers are delegating discovery and, at the margin, buying to machines, and a synthesized answer measurably suppresses onward clicking and concentrates the choice. The exact magnitude for any given market is not documented, because no independent census of the answer layer exists and reported AI-search share varies several-fold by denominator. The defensible action is therefore to measure a business's own standing on the answer surface directly, rather than infer it from a borrowed aggregate that describes no single market and no single operator.

What gets a business chosen is shifting in a way the ranking metaphor no longer captures. As delegation climbs the ladder from Consult toward Shortlist and Decide, the businesses that get named are increasingly those that are retrievable by a machine, first-named inside the answer, and clean at the point of purchase, rather than merely well-ranked on a page a diminishing share of buyers ever click. Each of those three properties is measurable, and each is addressable, which is what makes them a more useful target than a national adoption headline.

The durable backbone of this analysis is that the supply-side gap is measured rather than modeled. The Machine-Readiness Atlas shows, economy by economy, that the machine-readable share of the small-business web is a minority everywhere and that the human-facing basics are near-universal while the machine-facing finish is skipped. That gap is high in importance, low in cost, and low in prevalence at once, which is precisely the profile of an advantage available to whoever closes it first. A demand shift that is real but imprecisely sized, set against a supply gap that is measured, argues for acting on the part that can be measured now.

The priority that falls out of the study is to build first-party measurement of standing on the answer surface and to close the machine-readability gap that determines eligibility to be named at all. Neither step depends on resolving the open magnitude questions the study is careful not to overstate. Both can begin from a measured position rather than a borrowed one, and both are within reach of an individual business today. The rest of the answer is patience: the frontier is unmeasured, the instruments still disagree, and the correct posture toward every figure here is to treat it as directional and to verify locally what an aggregate can only suggest.