AEO Updates reference report · Supported by Dig Insights
State of AEO 2026
Updated 17 October 2026 · 34 min read · Ian Ash
A living, source-led report on how brands are discovered, represented and chosen in AI-mediated decisions, and what changes when assistants build interfaces, review customer relationships and request authority to act.
October 2026 update
This October update reviews the later evidence for interfaces, merchant readiness, permissioned agents, customer advocacy and adviser trust. It preserves the August adoption and platform-weighting baselines rather than presenting them as refreshed October estimates.
Review scope: a frozen 117-article archive, including 26 dated candidate articles from 29 August through 17 October 2026. The earlier Machine Positioning evidence was already covered. This is an editorial evidence review, not a representative market survey or a count of independent studies.
Representative decisions
Prompt coverage, account state and human-machine gap research now sit beside explicit denominators. August baselines are retained, not relabelled as current estimates.
Evidence operations
A canonical claims library, qualifications, owners and persistence tests turn the evidence agenda into a repeatable operating cadence.
Interfaces and commerce
AI-built comparisons and Google/Shopify routes broaden the readiness question. Recommendation, checkout, payment and merchant responsibilities remain distinct.
Permissioned agents
Discover, Compare, Prepare and Act is a task diagnostic. Agent recognition is not customer authority; financial-service plans remain explicitly future-tense.
Customer advocacy
Existing-customer bills, renewals and service value become research questions. Proactive tools and expert forecasts are not measured switching or retention.
Advertising and trust
Personal Intelligence advertising and task-specific experiments add sponsorship, data-use and adviser-integrity questions without a universal trust claim.
Edition history
- 17 Oct 2026 · Current edition · From visibility to decisions, permissions and loyalty. Adds prompt and gap research, evidence operations, AI-built commerce interfaces, customer advocacy and paid-advice trust. Keeps August planning baselines dated and separate.
- 29 Aug 2026 · Fourth August revision · Machine Positioning and the evidence environment. Adds a brand-strategy diagnostic plus separate Ahrefs, Semrush and Webflow visualisations with visible samples, definitions and causal limits.
- 28 Aug 2026 · Third August revision · Adoption, maturity and weighting. Adds the 55–65% intentional-adoption working range, a seven-stage maturity diagnostic and separate robustness and probable-exposure views.
- 28 Aug 2026 · Second August revision · Measurement, evidence and ownership. Added the Measurement Declaration, website-handoff evidence, evidence functions, source survivorship, paid influence and accountable ownership.
- 16 Aug 2026 · Full August edition · From visibility to an operating model. Elevated eligibility versus preference, decision impact, retrieval architecture, commerce readiness and evidence integrity.
- 24 Jul 2026 · Original reference guide · The first State of AEO baseline. Established the standalone report with ten sections covering the category, platforms, measurement and implications for brands.
State of AEO update articles
Read the dated briefings alongside this living reference report. Future published briefings will be added here.
October 17, 2026 · Report edition:
State of AEO 2026: the 17 October update
A four-minute briefing on the October edition: AI representation, human choice, merchant readiness, customer-authorised action and adviser trust.
Eligibility versus Preference diagnostic
Access and eligibility make a brand observable. Inclusion, preference and human choice remain different questions. Public evidence and customer-authorised private actions require different permissions.
Executive Summary
The State of AEO has widened beyond getting a brand into an answer. The evidence reviewed through 17 October now concerns how AI represents an offer, builds a comparison interface, prepares a purchase, advises an existing customer and requests permission to act. Those are related decisions, not interchangeable measures. [51] [43] [45] [50]
The August 2026 adoption and platform-weighting baselines remain dated planning inputs. This October edition does not replace them with a new market census. Later product announcements, surveys and experiments add evidence about particular capabilities and tasks, not a comparable measure of all AI-search use. [40]
Five additions sharpen the operating agenda: representative prompt panels and human-choice research; AI-built interfaces and merchant readiness; task-specific agent permissions; customer advocacy and loyalty; and advertising, personalisation and adviser trust. Evidence operations now require a canonical claim, an accountable owner, current qualifications and repeatable testing, not simply a larger content inventory. [42] [41] [49]
The working conclusion is that teams should make their brand's advantages clear, defensible and accessible, then test whether those advantages survive the relevant decision contexts. Use correctly designed human-machine gap research to connect the questions. Use secure, customer-authorised handoffs to connect permitted actions. Neither an answer, a signed request nor a completed order establishes customer loyalty. [41] [50] [46]
The Category Has Changed
Baseline date: the adoption working range and maturity estimates below are the August 2026 cross-study synthesis, not October prevalence estimates. They remain separate planning approximations with their original uncertainty; this update supplies no matching new adoption denominator.
The early AEO conversation focused on whether brands were cited by ChatGPT, Gemini, Perplexity and other answer engines. The current question is more demanding: what does the model believe about the brand, which claims does it repeat, in which decision contexts does it surface the brand, and does that representation create commercial value? That shift moves AEO beyond monitoring. It requires an evidence and information architecture that a model can retrieve, interpret and qualify correctly.
Institutional validation has continued. HubSpot's free AEO Grader and paid monitoring product signal that AI visibility has entered the mainstream marketing stack. Pure-play tools continue to add multi-model monitoring and agentic workflows. Holding companies have embedded AI visibility capabilities into their proprietary platforms. But the market has also exposed a gap: dashboards show mentions, while few operating models explain how a brand should improve the claims and evidence that determine those mentions. [1] [2]
Competition is also becoming observable inside retrieved content. The GEO-Flag preprint estimated optimisation signals on 8.90% of 10,095 webpages retrieved for 1,000 real-user queries, rising to 16.36% among pages modified in 2026. Those estimates apply to the researchers' released Google Search and Gemini-grounded retrieval sample, not the whole web, and the work remains a preprint. Its strategic implication is narrower: basic retrievability will not preserve advantage when competing sources are also adapting. [28]
Adoption now appears to have crossed into the marketing mainstream, but the label covers a wide maturity range. AEO Updates' current working estimate is that approximately 55–65% of marketing organisations are doing something intentional about AI search. Roughly 20–25% may have integrated AEO across strategy, execution and reporting, dedicated capability sits near 10%, and mature cross-functional practice is probably rarer still. These stages are not measured by one instrument, are not necessarily nested and should be used as a diagnostic rather than a league table. [33]
The range is anchored primarily to HubSpot's 58% broad-marketer finding. Minuttia and Growth Memo, Semrush, Webflow and McKinsey provide distinct contextual frames measuring implementation stages, strategy change, observable website conditions and systematic tracking. Their samples and thresholds should not be pooled. Uncertainty increases toward the mature end of the diagnostic. [33]
The category is also moving closer to brand strategy. Machine Positioning gives a name to the gap between what a brand intends to stand for and what an AI system actually says when a decision becomes specific. It is not a separate identity for machines. It is a diagnostic for Brand Representation, Consideration and Recommendation within a defined audience and decision territory. [34]
Machine Positioning Becomes a Strategic Measurement Layer
Machine Positioning is the degree to which intended brand positioning is faithfully and advantageously represented in AI-mediated decisions. It asks whether named AI systems connect the brand with the intended audience, attribute, use case and decision territory, then preserve that position as constraints become more specific. It is a working AEO Updates definition, not a separate brand strategy for machines and not a validated score. [34]
Ahrefs' screened study of 75,000 established domains found that YouTube mentions had a 0.737 Spearman rank correlation with AI visibility, YouTube mention impressions 0.717 and branded web mentions between 0.66 and 0.71 across measured AI surfaces. For ChatGPT, branded search volume was 0.352, Domain Rating 0.266 and website page count 0.194. The sample excluded many smaller or low-demand brands, and the association does not show that adding a channel or mention will cause a visibility gain. [35]
Semrush and Growth Memo observed 1,094 US ChatGPT categories monthly from January to June 2026. Under their five-prompt consistency and margin test, 15.2% had a clear owner, 31.2% an emerging leader and 53.7% remained unsettled. Topic owners had higher branded search volume in 55.7% of matched pairs, higher organic traffic in 48.4% and a higher Authority Score in 52.5%; only branded search was statistically significant. Broad SEO strength therefore did not consistently identify the topic owner, but the study did not isolate the missing cause. [36]
Webflow's cross-sectional assessment of 2,000 US websites placed companies into a proprietary 123-gate maturity system. The Level 3 group had 2.3 times the mention rate and 3.7 times the citation rate of the Level 1 group. Webflow reported on-site authority as the strongest correlate of mention rate, followed by third-party recognition. These are differences between groups, not measured gains from companies moving through the levels. [37]
Yext's analysis of 17.2 million citations provides a different boundary. Citation-source patterns varied by model, sector, industry and query class. Listings represented 54.53% of distinct URLs, while websites produced more citation occurrences per URL; Claude and SearchGPT also showed distinct source preferences in the studied sectors. Citations are not brand recommendations, and the Q4 2025 snapshot does not reveal a universal source recipe. [38]
The cross-study architecture is therefore a hypothesis: Human Brand Equity may create a wider Public Evidence Footprint, which may support Machine-Readable Brand Representation and Decision Fit before a brand enters AI Consideration or receives a Recommendation. Access and Visibility make the brand observable. They do not prove accurate representation, preference or downstream human choice. [34]
Eligibility Comes Before Preference
The most useful current distinction is between eligibility and preference. Eligibility asks whether an answer engine has enough reliable, current and retrievable information to treat a brand as a credible candidate at all. Preference asks whether the available evidence gives the system a defensible reason to place that brand ahead of alternatives. These are related, but they are not the same competitive margin.
A complete-market audit of 4,776 Bali hospitality businesses gives the distinction an empirical basis. In that setting, own websites, review volume, listed pricing and third-party mentions were associated with entering the answer, while higher ratings predicted first position among businesses already recommended. The study is observational and category-specific, so its exact rates should not be generalised. Its structural lesson is stronger: teams should diagnose a documentation and evidence gap differently from a comparative preference gap. [12]
That changes the measurement sequence. The decision chain is eligibility, inclusion, recommendation, choice and commercial outcome. A brand may be technically eligible but absent from a particular prompt domain, included but not preferred, or recommended without any measurable downstream action. Treating all of those outcomes as visibility conceals the intervention required to improve them.
The Commercial Signal
The commercial case is now stronger than the early traffic debate suggested. Reuters reported Adobe Analytics findings that visitors referred by AI services generated 41% higher revenue per visit than shoppers arriving through traditional channels. The same dataset found that 41% of US consumers used generative AI for online shopping in June. Ulta Beauty reported that shoppers arriving through Gemini and ChatGPT showed double the conversion and intent. Walmart, Ulta Beauty and Wayfair are among the retailers optimising for AI shopping visibility. [5]
These findings do not make every AI referral intrinsically more valuable, and they should not be treated as a universal benchmark. They do, however, explain why visibility is becoming a finance question rather than solely a communications question. AI-assisted shoppers may arrive after an assistant has interpreted the need, compared alternatives and narrowed the decision. AEO measurement therefore needs to connect representation to qualified sessions, conversion, revenue per visit and assisted pipeline, not only citation share.
Recommendation also deserves its own place between representation and commercial outcome. In a controlled pension-allocation experiment, changing the AI-generated advice changed later portfolio choices: roughly 37% of the experimentally induced difference between two recommendations passed through into final allocations. The setting was hypothetical, financial and specific to 400 employed South Korean adults, so it does not prove a universal purchase effect. It does provide direct evidence that recommendation content can influence consequential decisions without displacing human judgement. [11]
Two later agency datasets point in the same direction without establishing one universal effect. Brainlabs observed declining organic sessions across 54 clients and 140.1 million organic sessions while AI-referred visitors generated advertiser-defined key events at 1.5 times the organic-search rate. First Page Sage reported a 23.6% organic-session decline across a separate 3.4-billion-session panel of 430 websites while AI referrals reached 6.2% of traffic and converted at a higher reported rate. The panels, definitions and periods differ and should not be pooled. Neither isolates AI Overviews as the sole cause of traffic decline. Directionally, they support a website role that is becoming more transactional and authoritative as more research occurs upstream in AI-mediated environments. [16] [17]
October adds functioning and proposed buying routes, not a universal agentic-commerce market. Discovery eligibility, recommendation inclusion, cart preparation, checkout start, payment authorisation, order confirmation, fulfilment and repeat purchasing need separate eligible or authorised-attempt denominators. A merchant can remain seller or merchant of record without controlling the interface that brings the next customer visit. [44] [45]
Customer Advocacy and the Future of Loyalty
AEO is relevant to existing customers as well as prospects. A permissioned assistant can help review a renewal, price rise, recurring charge or service problem. Rocket Money describes Rowan as proactive personal-finance assistance with savings opportunities, recurring-bill and cancellation support, with human verification in the service. The broader research identifies specialist consumer-advocacy products and features embedded in familiar financial tools. Product descriptions, demonstrations and anecdotes do not establish typical savings or widespread use. [55] [46] [47]
The strategic hypothesis is that review may begin before the customer actively asks whether to leave. An opt-in alert can reduce the burden of noticing a poor deal and formulating a question; it does not remove permission, trust, verification or switching friction. The reviewed evidence does not establish population adoption or a causal retention effect. QuickBooks-style advocacy remains an embedded-tool hypothesis, not a verified consumer product feature. [46]
Gartner's January survey of 322 US consumers illustrates a narrower opening for help than for autonomy: reported willingness to let AI narrow choices reached 31% in a low-stakes category, while willingness to delegate purchase decisions topped out at 11%. Those are category-specific stated intentions, not observed behaviour. Named expert predictions and programme ambitions in the consumer-advocacy research remain forecasts, not evidence of switching. [62] [46]
An assistant may reinforce an instructed brand preference, expose convenience-based retention or conclude that the current provider offers the best whole-relationship value. Research should test informed staying, negotiation, correction, downgrade, cancellation and switching. Separate advice accuracy, authorised action, reversals, fee-adjusted value and later retention; an AI-referred order is not proof of loyalty. [47] [45] [46] [48]
Search Is Becoming a Task Interface
Google's limited homepage experiment offers a useful design signal. In the test, the classic Search button was replaced for some desktop users by Create images, Ask about files and Brainstorm actions. Google described the change as a small experiment that does not alter core search functionality. The feature itself may disappear. Its strategic significance is that Google is putting AI tasks directly in the interface that previously expressed a simple query-to-results transaction. [7]
This convergence changes what brands are optimising for. Search is becoming more conversational, generative, multimodal and increasingly agentic. A user may ask an AI system to research, compare, create, reserve or buy rather than to find a webpage. In that environment, brand visibility is not limited to ranking. It includes whether the AI recognises the brand, understands its fit for the task, retrieves valid supporting evidence and recommends it with appropriate qualifications.
For retailers, marketplaces and transaction-led services, this shift creates a capability layer alongside content and evidence. Google’s Universal Commerce Protocol is designed to connect product discovery, checkout and payment capabilities across conversational surfaces, including AI Mode and Gemini. That does not make every company an agentic-commerce business. It does mean that organisations for which the transaction itself matters should audit product data, availability, pricing, Merchant Center eligibility and the technical route from recommendation to action. [13]
AI-Built Interfaces and Merchant Readiness
The brand's website is no longer necessarily the interface in which a customer compares it. OpenAI's 7 October Intelligent UI announcement describes dynamically built task interfaces; the relevant AEO question is which product facts, qualifications and comparisons remain faithful when presentation is reconstructed. This is a documented interface capability, not evidence that customers surrender decisions or that a new layout improves conversion. [53] [43]
Google's Universal Cart rollout and Shopify's shopper-agent instructions show different paths to a purchase. Google describes checkout for participating brands or a transfer to the brand's site; its original announcement retains the brand as merchant of record. Shopify describes compatible agents, a connected Shop account and confirmation for purchases. Channel, customer, market, product and merchant eligibility still matter. These routes are neither technically identical nor universally available. [45] [52] [54] [44]
Readiness is operational as well as editorial. Keep price, variants, stock, delivery, returns, warranty, eligibility and promotional conditions accurate; test the route the customer can actually use. A Buy button does not establish payment success, fulfilment performance or incremental demand. Where checkout occurs does not establish which brand, platform or retailer earns the next visit. [44] [45]
The planning implication is to begin with one eligible journey and test its failure cases, authority boundary and merchant responsibilities. Measure completed business separately from AI visibility. Recommendation ranking, customer identification, service ownership and repeat purchasing remain questions to observe, not benefits to presume. [44] [45]
Agent Access Is a Permissioned Journey
Being recommended is not permission to enter an account or act for its owner. Cloudflare distinguishes Search, Agent and Training behaviours, and a bot may have more than one. Visa's Trusted Agent Protocol describes recognition and message integrity. Neither a behaviour label nor an identifiable request by itself proves customer consent, permitted account scope, spending authority, merchant acceptance or completed payment. [59] [60] [50]
AEO Updates proposes four task permissions: Discover public facts; Compare the complete offer; Prepare a reversible draft or shortlist; and Act on a consequential commitment. These are not a sequential adoption funnel or validated score. A private read can be sensitive, and a supposedly preparatory action that creates a binding obligation belongs at the action boundary. [50]
Bank reporting illustrates the distinction. JPMorgan's reported controls concern outside use of customer credentials and money movement, not withdrawal from public AI discovery. Bloomberg reporting carried by Moneycontrol says Schwab blocked unauthorised money-moving agents and described planned secure connections for balances and research without money movement. Those connections remain reported future plans, not a generally available read-only service. [48] [61] [50]
Product, security, legal and service teams should define data scope, identity checks, customer confirmation, revocation, records, recovery and escalation. Record an accurate response, a correct policy refusal, a failed authenticated handoff and a completed authorised action as different outcomes. A successful signature or a visible recommendation is not a transaction. [50] [60]
From Documents to Claims
Most organisations manage a document estate, not a claim estate. They know which webpages, case studies and PDFs they own but cannot reliably identify the exact assertions an AI system should make about them, the evidence that supports each assertion, or the decision context in which it matters. That is a critical weakness because answer engines increasingly assemble recommendations from claims, entities, sources and constraints, rather than simply returning a single document.
The next layer of AEO is claim architecture. A priority claim must be specific enough to be cited, bounded enough to be true, supported by current evidence, linked to the relevant entity, and presented in an answer path that a retrieval system can follow. The work is not just to create more content. It is to establish a coherent route from claim to proof to answer to adjacent decision questions. [9]
The October operating extension is a canonical claims library. For each material decision claim, record the exact wording, eligible audience and market, supporting source, qualifying conditions, effective date, owner, independent corroboration where available and review date. Make accurate wording easy for customers, publishers and AI interfaces to use without broadening the underlying evidence. This is an editorial governance recommendation, not a measured ranking factor. [42] [43]
The Platform Landscape Is Fragmenting
The major answer engines should not be treated as interchangeable. ChatGPT Search rewrites prompts and uses partner search providers, including Bing and Shopify. Google combines its search infrastructure with AI Overviews, AI Mode and increasingly task-oriented interfaces. Claude's connector ecosystem makes specialist tools available inside a generalist reasoning environment. Perplexity, Copilot and Grok each have distinct retrieval, citation and user-context behaviours. The same brand may therefore be represented differently across models, even when the underlying public web is unchanged.
A model-agnostic strategy is not a compromise. It is the only durable strategy. It should use a common evidence foundation while testing how the brand is retrieved, characterised and recommended in each material platform. The objective is not to make every answer identical. It is to make the brand consistently understandable, accurately qualified and competitively credible across the decision environments that matter.
The Evidence Supply Chain
AEO depends on an evidence supply chain: original information, source, retrieval, AI interpretation, answer and consumer decision. The first two stages are under increasing pressure. A Pew Research Center study found that traditional result clicks occurred on 8% of visits with an AI Overview, compared with 15% without one, while clicks to cited sources occurred on only about 1% of AI Overview visits. The research is observational, so it does not establish causation, but it makes the relationship between citation and publisher traffic impossible to ignore. [4]
A later preregistered field experiment strengthens the evidence hierarchy. Among 1,100 U.S. adult Chrome users observed over seven days, forced AI Mode exposure reduced external click-through by 18.8 percentage points and did not improve reported trust, satisfaction or platform loyalty in that setting. The experiment forced exposure rather than observing natural adoption, and compliance in the no-AI condition was incomplete after a Google interface change. Its effect size should not be generalised to every user or query. It does show causally that an AI-search interface can reduce publisher traffic. [24]
The economic response is already visible. In France, the Alliance of General Information Press asked the competition authority to act over Google's AI-generated summaries, and Reuters reported that regulator Arcom estimated traffic declines of 33% to 38% for member publishers' sites because of the summaries. As publishers increase crawler controls, paywalls, licensing demands and structured commercial feeds, answer engines may gain uneven access to open, licensed, restricted and proprietary sources. This is not a peripheral publisher issue. It is an AEO infrastructure issue because it changes the evidence available to each model. [6] [10]
The operational implication is that a brand should record provenance alongside performance. A coverage gain built on a current first-party specification, independent reporting and verified customer evidence is materially different from a gain built on one stale or commercially interested source. Source type, renewal date, retrieval accessibility and cross-model consistency should become fields in the claim register, rather than after-the-fact explanations for a result.
A useful planning framework is to ask which evidence function a customer question requires: product truth, commercial truth, independent validation, expert interpretation or lived experience. Recent cross-study citation evidence suggests that prompt purpose, category and engine can change which source types are useful. This five-function model is an AEO Updates working hypothesis, not a settled taxonomy of retrieval behaviour and not a universal first-party versus third-party ratio. [20] [21]
The companion principle is Evidence Survivorship: priority brand conclusions should remain supportable when one engine, domain or retrieval pathway changes. Promptwatch observed Reddit's share of citation-bearing ChatGPT Search responses fall from 3.83% to 0.52% in its panel, while a separate Peec AI before-and-after study reported a 50.5% decline in listicle citation share around the GPT-5.6 rollout. Neither study proves permanent platform decline or that retrieval stopped. Together they show why durable representation needs multiple credible evidence pathways, time-aware measurement and caution about concentration. [22] [23]
Claim persistence is a testable operating question: does the properly qualified advantage remain accurate across relevant prompts, models, interfaces, dates and markets? Retain the answer context and source identity; distinguish omission, stale evidence, contradiction and deliberate policy refusal. Disclose whether a finding is primary documentation, company-reported deployment, journalism, survey, experiment, forecast or editorial interpretation. These evidence classes are not scores and should not be pooled. [42] [50]
Earned, Owned and Paid Influence
The old organic-versus-paid search distinction is beginning to reappear in AI systems, but with higher stakes. Earned representation occurs when an answer engine retrieves ordinary credible sources and independently chooses to use them. Owned influence comes from a brand's claim and answer architecture. Paid representation has begun to emerge through disclosed advertising, local lead formats, commerce partnerships and agentic infrastructure. Manipulated representation begins when commercial influence borrows the appearance of independent evidence.
A controlled audit of the early U.S. ChatGPT advertising environment observed more than 3,000 advertisements from 186 advertisers across 91 accounts and 335 realistic prompts. OpenAI's documentation indicates that current-conversation context and intent can help select an advertisement, suggesting a paid-intent unit broader than exact-match keywords. The environment was an early rollout and may change. Organic representation in an answer and paid access around it should therefore be measured as distinct systems. [25] [26]
Google's Preferred Sources feature adds another influence layer: users can explicitly identify publishers they want to see more often, including in source presentation connected to AI Overviews and AI Mode. Google has not disclosed a conventional ranking weight or established that the preference materially changes generated-answer composition. The feature is best treated as an explicit human trust signal whose practical effect remains an empirical question. [30]
The recent dispute over machine-targeted advertising makes the distinction operational. Perplexity blocked sponsored material inserted into bot-facing markdown versions of a publisher’s pages from influencing its index, arguing that the practice could amount to cloaking even where the material was labelled. That single platform decision does not settle an industry-wide rule, but it establishes that platforms may judge disclosure, source parity and commercial context before a common standard exists. [15]
An AEO evidence-integrity standard should require four things: the factual core made available to AI should be available for human inspection; commercial material should retain machine-readable disclosure through retrieval and citation; priority claims should be independently verifiable; and platforms should explain how commercial material is treated in retrieval, ranking and recommendation. Paid consideration cannot replace the evidence required to sustain belief and preference.
The new advertising-and-trust chapter extends this distinction. Sponsored exposure, organic representation and adviser recommendation should be recorded separately, including whether people recognise the commercial relationship and understand the use of connected-app information. Task-specific experimental findings do not establish a universal penalty for advertising or a measured effect on long-term loyalty. [49] [56] [57] [58]
Advertising, Personalisation and Adviser Trust
Search Engine Land reported on 16 October that Google was beginning to test ads in AI Mode for users with Personal Intelligence enabled. Google's help guidance says ads are not based directly on Gmail or Google Photos, while summaries, excerpts and inferences from connected sources may inform advertising. Advertisers do not receive access to those repositories. A beginning test and a data-use policy are not evidence of universal availability or a measured trust effect. [49] [56]
The relevant question is not simply whether people notice an ad. It is whether they understand the difference between paid placement, independent advice and an agent instructed to promote a sponsor, and whether they understand how personal context is used. Keep ad recognition, understood independence, perceived integrity, recommendation adherence and commercial outcomes separate. Survey concern and stated openness are not observed buying, abandonment or loyalty. [49]
Two preregistered experiments reported by Salvi, Cuevas and Horta Ribeiro had 2,012 final participants. In the controlled eBook task, adjusted sponsored-book selection was 61.2% under sponsor-promoting agent instructions versus 22.4% under search placement, in arms of 404 and 402 participants. That is a task-specific selection effect, not a live retail conversion rate. A warning-plus-label bundle improved bias detection, but its adjusted reduction in sponsored selection was not statistically significant. [57] [49]
Counterevidence matters. NIM's tablet-shopping experiment summary, covering 555 participants in Germany, Austria and Switzerland, reports no direct disclosure effect on recommendation adherence; the perceived-integrity effect depended on interactivity. Its summary does not publish cell sizes or inferential statistics. These findings do not support a universal trust penalty, universal disclosure remedy or a long-term loyalty forecast. [58] [49]
The Measurement Gap
The existing generation of AEO measurement is useful but incomplete. Citation frequency, share of voice and sentiment remain valuable directional indicators. HubSpot's AEO Grader demonstrates that presence quality and brand characterisation matter, not only simple inclusion. Yet a citation score cannot reveal whether the model surfaced the right claim, relied on current evidence, applied the right qualification or created commercial value. [1] [2]
AI visibility also has no common denominator. One platform may divide brand appearances by all tested prompts, another may calculate share only within brand-containing answers, while another weights results by a modelled prompt index. Those numbers are not interchangeable. A measurement claim should therefore declare the prompt universe, denominator, engine and model, geography and user state, observation method, cadence, coding rules, weighting and known exclusions. [18]
The observation method matters as much as the formula. Controlled synthetic browser studies can repeat a declared prompt set under stable conditions. Naturalistic user research observes how people actually use systems. Owned behavioural data measures sessions, events and commercial outcomes on properties the organisation controls. Modelled indices estimate a wider opportunity universe. Each can be useful, but each answers a different question. Public cloud-browser documentation shows what controlled observation can technically support; it does not reveal the undisclosed stack of every measurement provider. [19]
Google Search Console may also expose fragments of AI Mode follow-up demand as new query strings without identifying their generative origin. One 16-month analysis found 1,127 candidate conversational queries representing about 20,300 impressions, but the classifier was author-designed and conversational syntax alone cannot prove AI Mode origin. These fragments are a partial observed-demand signal, not a complete AI-search panel. [31]
A mature measurement framework needs seven lenses. Technical availability asks whether priority claims are present in the bot-visible page state. Coverage asks whether those claims have a current owned answer path and proof. Integrity asks whether the evidence is verified, qualified, current and source-attributed. Retrieval asks whether supporting sources are surfaced in priority prompt domains. Representation asks whether the model states the claim correctly and competitively. Decision impact asks whether the brand enters the consideration set, wins the recommendation and affects downstream choice. Commercial impact asks whether better representation changes qualified traffic, conversion, revenue per visit or pipeline.
Machine Positioning adds an explicit strategy-to-observation comparison inside that framework. The organisation defines the truthful intended position first, then tests observed AI descriptions, brand-attribute associations, consideration-set entry and recommendation separately. Any gap is a diagnostic result. It is not evidence that one source or channel caused the answer. [34]
The measurement unit should also remain platform-specific. The Bali audit found low agreement across the four tested answer engines, and reported stale closed-venue recommendations more often than likely fabricated business mentions. A single aggregate score can therefore hide a brand that is eligible in one environment, absent in another, or represented with out-of-date facts. Repeated, source-aware multi-model testing is more demanding than a dashboard snapshot, but it is closer to the system being measured. [12]
Cross-platform reporting also needs two complementary views. A platform-neutral view asks whether the brand's facts, positioning and evidence survive across engines. A market-weighted view asks where a gain or loss may affect the largest probable standalone-assistant audience. The first is an evidence-quality diagnostic. The second is an exposure-planning estimate. Neither should hide the engine-level scorecard. [32]
AEO Updates' current planning allocation is ChatGPT 60%, Gemini 25%, Claude 7%, Copilot 6% and Perplexity 2%. The allocation is a dated editorial judgement informed by traffic, primary-choice, B2B and population evidence with incompatible denominators. It is not a measured universal market share and does not cover embedded AI experiences. Audience, task, buying stage and distribution can justify documented overrides. [32]
October update: a prompt panel must also disclose the decision stage, prompt families and near-neighbour wording, geography and language, account state, model, interface and observation dates. A larger prompt corpus is not a census of people, search occasions or transactions. Public answers, authenticated account responses and commercial handoffs require different observation frames. [39] [40] [50]
Retained baseline: the standalone-assistant allocation below is the August 2026 planning judgement, not an October update or universal market-share estimate. Show platform-neutral robustness alongside any documented audience weighting. No reviewed October evidence supports silently reallocating the weights or adding embedded Google audiences to standalone use. [40]
Representative Prompts and Human-Machine Gap Research
A prompt panel is a declared model of a decision space, not the market itself. Select relevant audiences, decision stages, question families and near-neighbour wording; state the language, geography, interface, account state and observation period. Retain repeat observations rather than treating one answer or a very large vendor corpus as a population census. [39] [40]
A properly designed gap between human and machine evidence is a useful research opportunity. The machine side can record retrieval, mentions, citations, correct qualifications, association and recommendations. The human side separately measures awareness, understanding, consideration, trust, willingness and choice. Match the audience, need, market and time period before interpreting disagreement. Do not directly subtract unlike percentages or present the matrix as a validated score. [41]
The gap creates specific questions: is a credible advantage absent from AI answers; is a prominent recommendation irrelevant to customers; is the offer misunderstood; or is a correct recommendation blocked by an appropriate permission boundary? The answers direct evidence, positioning, experience and access work. They do not by themselves establish that a visibility intervention caused a commercial outcome. [41] [50] [42]
Next, connect the research to observations or controlled tests of human outcomes, permitted handoffs and completed business. Use longitudinal or experimental designs where causality is the question. Keep missing data visible. The report's 117-article archive review and 26 dated candidate articles are an editorial evidence search, not a market sample, popularity ranking or quantitative trend frequency. [51]
What AEO Is Not
AEO is not a replacement for SEO. Technical accessibility, semantic structure, authority, internal linking and credible content all remain essential. AEO adds new work on top of that foundation: intermediate-query mapping, claim-level evidence, entity consistency, answer architecture, model-specific representation testing and commercial outcome measurement.
AEO is also not a license to manufacture evidence. The growing use of covert brand mentions and apparently independent recommendations in high-citation communities such as Reddit illustrates the difference between optimisation and manipulation. Making truthful information easier to retrieve and verify is legitimate optimisation. Fabricating information to look independent, or presenting a different promotional factual environment to machines than people, is an attack on the evidence environment that answer engines need to function.
The 18-Month Outlook
Five developments will define the next phase of AEO. First, more search interfaces will become task interfaces, reducing the distance between question, recommendation and action. Second, the evidence supply chain will become more commercial as publishers restrict, license or structure their information for AI use. Third, AI shopping and local commerce will make revenue attribution more important than top-line citation counts. Fourth, paid AI influence will mature alongside earned representation, creating a stronger need for disclosure, integrity standards and model-aware media strategy. Fifth, bot-visible auditing will become a precondition for any claim, content or commerce strategy that depends on retrieval.
The category will not be won by the platform with the most attractive dashboard. It will be won by brands and partners that can make the underlying information environment more useful, more verifiable and more commercially accountable. The operational advantage will belong to organisations that can connect prompt domains, claims, evidence, answer paths, retrieval tests and business outcomes into one managed system.
October qualification: the commercially useful next step may be reliable preparation and an explicitly authorised handoff, rather than unrestricted autonomy. Financial-service controls, merchant eligibility and buyer approval can coexist with public AI discovery. Schwab's proposed secure connections remain reported plans. None of these announcements supplies a population adoption rate or proves switching, trust or revenue outcomes. [50] [61]
Recommendations for Brand Marketers
1. Appoint one accountable owner. Give one AEO or AI Search Intelligence leader responsibility for observing representation, diagnosing causes, prioritising problems, assigning specialist interventions, measuring change and reporting learning. Preserve broad input and distributed execution across SEO, Content, Brand, PR, Product Marketing, Insights, Social, Product and Customer Success. This is an AEO Updates operating hypothesis; the underlying ownership surveys measure professional perceptions, not which org chart causes better performance. [27]
2. Diagnose the maturity constraint. Identify whether the organisation is still experimenting, actively implementing, measuring systematically, integrating execution or building dedicated capability. Choose one intervention that can plausibly affect the present bottleneck rather than aiming abstractly for the final stage. The maturity curve is a diagnostic synthesis, not a universal benchmark. [33]
3. Define and audit Machine Positioning. Write the intended position in one sentence for a named audience and decision territory. Then compare it with observed AI descriptions, associations, consideration and recommendation across the systems that matter. Treat every gap as a diagnosis, not proof of a causal channel. [34]
4. Diagnose the decision constraint before choosing the tactic. Determine whether the brand lacks eligibility for consideration, inclusion in the relevant prompt domain, preference once included, or evidence of downstream impact.
5. State the weighting purpose. Use platform-neutral results to test cross-engine robustness and documented market weights to estimate probable exposure. Preserve every engine-level result, record the baseline and any audience override, and date the decision. [32]
6. Build a claim register. Catalogue the specific assertions the brand needs an AI system to make, the evidence supporting each one, the required qualification, source provenance, the owner and the renewal date. Treat stale or unprovable claims as a risk, not simply a content gap.
7. Audit the bot-visible representation. Verify that priority claims, evidence, pricing, availability, comparison criteria and structured data are present in a form relevant retrieval systems can encounter, rather than only in a human-triggered interaction.
8. Design query-lattice answer paths. Create and connect direct answer pages for integrations, comparisons, use cases, pricing conditions, methodologies, proof requirements and relevant customer constraints.
9. Measure the decision chain, not just citation. Track eligibility, inclusion, top-recommendation rate, representation quality, freshness and commercial outcome separately across material answer engines.
10. Protect the evidence ecosystem. Invest in original research, credible third-party corroboration and transparent earned coverage. Do not manufacture supposedly independent recommendations, source signals or bot-only promotional claims.
11. Treat paid AI influence as a separate, disclosed layer. Test emerging commerce and local-ad formats only where the commercial context can remain intelligible to both the platform and the user.
12. Add agentic-commerce readiness where it is relevant. Retailers and transaction-led services should assess product data, real-time availability, merchant capability and the path from AI recommendation to completed action.
13. Preserve the SEO foundation. Ensure critical content is accessible to the bots and retrieval systems the brand has chosen to permit, is structured semantically, and is consistently linked to its supporting proof.
14. Govern canonical decision claims. Give material advantages an owner, supporting source, qualifying conditions, market and review date; test claim persistence across declared prompts and interfaces. Treat the seven-action model as an operating checklist, not a validated ranking or sales formula. [42] [43]
15. Test one permissioned buying route. Verify catalogue accuracy, eligibility, customer approval, checkout, payment, fulfilment and recovery separately. Record correct security refusals rather than counting them as poor visibility or broken conversion. [44] [50]
16. Research existing-customer advice. Test whether an assistant fairly recommends staying, negotiating, correcting, downgrading, cancelling or switching. Separate simulated advice from authorised actions and longitudinal retention; document missing denominators. [47] [46]
17. Audit paid-influence comprehension. Test whether the audience understands sponsorship, adviser independence and connected-data use, then measure verification and choice separately. Do not generalise a controlled task effect into a category-wide trust or loyalty prediction. [49]
Machine Positioning diagnostic
- Intended position
- What should the brand truthfully stand for, for whom and in which decision territory?
- Observed representation
- How do named AI systems describe the brand, its attributes and its fit?
- Decision fit
- Does the brand remain relevant as audience, use case, criteria and constraints become specific?
- Recommendation boundary
- Is the brand mentioned, considered or preferred, and what observable evidence surrounds the result?
The response boundary ends at Recommendation. Choice or behaviour requires downstream human or business evidence.
Retained August 2026 planning baselines
Retained quantitative findings keep their original sources, samples and periods. Adoption, maturity and standalone-assistant planning allocations remain August 2026 baselines. Product descriptions, survey intentions, forecasts and observed outcomes are different evidence classes.
Standalone assistant planning weights
- ChatGPT: 60%
- Gemini: 25%
- Claude: 7%
- Copilot: 6%
- Perplexity: 2%
Maturity planning diagnostic, not one prevalence study
- Aware or responding: ~85%. One vendor survey
- Intentional AEO activity: 55–65%. Working range
- Active implementation: ~40%. Editorial synthesis
- Systematic measurement: ~30–35%. Editorial synthesis
- Integrated operating model: ~20–25%. Anchored interpretation
- Dedicated capability: ~10%. Editorial synthesis
- Mature cross-functional AEO: Under 10%. Working hypothesis
Three studies, three different measures
Ahrefs rank correlations
In a screened 75,000-domain study, YouTube mentions correlated 0.737 with AI visibility. Observational rank correlation is not a channel prescription.
Semrush topic-ownership status
1,094 US ChatGPT categories, January to June 2026. Source-rounded values total 100.1%.
Webflow maturity-group differences
2,000-site, 123-gate cross-sectional study. Group differences are not observed gains as companies move between levels.
Source ledger
Retained quantitative findings keep their original sources, samples and periods. Adoption, maturity and standalone-assistant planning allocations remain August 2026 baselines. Product descriptions, survey intentions, forecasts and observed outcomes are different evidence classes.
- [1] HubSpot, AEO Grader and HubSpot AEO product launch, Spring 2026 Spotlight. https://www.hubspot.com/aeo-grader
- [2] HubSpot, State of AEO in 2026, survey of 1,100+ global marketers and citation data from the Xfunnel LLM database. https://offers.hubspot.com/state-of-aeo
- [3] OpenAI, ChatGPT Search documentation, including query rewriting, source citations, OAI-Searchbot access, location and reservation experiences. https://help.openai.com/articles/9237897-chatgpt-search
- [4] Chapekis et al., Pew Research Center, Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview, August 2026. https://arxiv.org/pdf/2608.04831
- [5] Reuters, Retailers tap AI shopping traffic but fight to keep customer data, August 2026. https://www.reuters.com/business/retail-consumer/retailers-tap-ai-shopping-traffic-fight-keep-customer-data-2026-08-07/
- [6] Reuters, French press body asks competition watchdog to take action over Google AI, August 2026. https://www.reuters.com/world/french-media-asks-french-anti-trust-watchdog-act-googles-ai-2026-08-11/
- [7] SFGATE, Google is quietly testing out new ways to crush the human spirit, August 2026. https://www.sfgate.com/bayarea/article/google-search-changes-22381960.php
- [8] AEO Updates, What a query lattice looks like in practice: a CRM software example. /articles/query-lattice-crm-software-worked-example
- [9] AEO Updates, Stop Retrieving Documents. Start Retrieving Claims. /articles/stop-retrieving-documents-start-retrieving-claims
- [10] AEO Updates, What happens to AEO when the sources AI needs can no longer afford to publish? /articles/when-ai-sources-cannot-afford-to-publish
- [11] Choi et al., Do People Follow AI Advice? Evidence from a Pension Portfolio Choice Experiment, August 2026. https://arxiv.org/html/2608.11371
- [12] Pitenin, Invisible to the Machine: Auditing AI Restaurant, Café, and Bar Recommendation Against a Complete Market Census, August 2026. https://arxiv.org/abs/2608.07069
- [13] Google, Under the Hood: Universal Commerce Protocol, January 2026. https://developers.googleblog.com/under-the-hood-universal-commerce-protocol-ucp/
- [14] Profound, Introducing Agent Analytics, February 2025. https://www.tryprofound.com/blog/introducing-agent-analytics
- [15] AEO Updates, Perplexity says you cannot advertise to the machine. /articles/perplexity-machine-targeted-ads-cloaking-aeo
- [16] AEO Updates, AI search may be moving the website down the funnel, covering Brainlabs data from 54 clients and 140.1 million organic sessions. /articles/ai-search-website-downstream-brainlabs
- [17] AEO Updates, 3.4 billion sessions point the same way: AI may be moving websites downstream, covering First Page Sage data from 430 websites. /articles/ai-search-traffic-purpose-first-page-sage
- [18] AEO Updates, AI visibility has no common denominator. /articles/ai-visibility-no-common-denominator
- [19] AEO Updates, Behind the AEO dashboard: how AI Search Intelligence is measured. /articles/behind-aeo-dashboard-ai-search-measurement
- [20] AEO Updates, One study says 52% of AI citations come from brand websites. Another says 2%. They may both be right. /articles/ai-citation-benchmarks-evidence-functions
- [21] AEO Updates, YouTube is #1. LinkedIn is rising. Reddit just fell. Stop calling them ‘social’. /articles/youtube-linkedin-reddit-human-evidence-ai-citations
- [22] AEO Updates, Reddit's ChatGPT Citation Share Collapsed. That Is a Warning for AEO. /articles/reddit-chatgpt-citation-share-source-concentration-risk
- [23] AEO Updates, ChatGPT’s listicle citation share fell by half after GPT-5.6. /articles/chatgpt-listicle-citation-share-gpt-5-6
- [24] Wang et al., AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence, preregistered randomised field experiment, August 2026. https://arxiv.org/abs/2608.18352
- [25] AEO Updates, ChatGPT has an ad platform. Why it matters for AEO. /articles/chatgpt-ads-paid-counterpart-to-aeo
- [26] AEO Updates, The keyword may not survive AI advertising as the primary unit of intent. /articles/keyword-not-survive-ai-advertising
- [27] AEO Updates, Who owns AEO? Three surveys give three different answers. /articles/who-owns-aeo-marketing-operating-model
- [28] AEO Updates, Nearly 1 in 11 retrieved pages may already be GEO-optimised, covering the GEO-Flag preprint. /articles/geo-flag-detecting-geo-optimized-web-content
- [29] AEO Updates, ChatGPT cited Reddit in 2 of 100 answers. The clue was in its searches. /articles/chatgpt-reddit-query-fanout-source-retrieval
- [30] AEO Updates, Google just gave publishers a new way into AI answers. /articles/google-preferred-sources-ai-search
- [31] AEO Updates, Search Console is exposing AI Mode follow-ups. Google still will not label them. /articles/google-search-console-ai-mode-query-fragments
- [32] AEO Updates, ChatGPT may represent 60% of standalone AI use. That changes how citation data should be weighted. /articles/chatgpt-ai-assistant-usage-citation-weighting
- [33] AEO Updates, AEO has gone mainstream. Maturity has not. /articles/aeo-adoption-outpaces-maturity
- [34] AEO Updates, What makes a brand show up in AI? Five studies point to its evidence environment. /articles/what-makes-brand-show-up-ai-evidence-environment
- [35] Ahrefs, Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews, December 2025. https://ahrefs.com/blog/ai-brand-visibility-correlations/
- [36] Semrush and Growth Memo, AI visibility is a topic-level game: a study of 50,000 brands in ChatGPT, July 2026. https://www.semrush.com/blog/chatgpt-topic-authority-study/
- [37] Webflow, The AI discovery gap: analysis of 2,000 websites, July 2026. https://webflow.com/blog/aeo-maturity-index
- [38] Yext, AI Citation Behavior Across Models: Evidence from 17.2 Million Citations, April 2026. https://www.yext.com/research/articles/ai-citation-behavior-across-models/
- [39] AEO Updates. The prompt is not the market. Why one question cannot measure AI visibility. August 30, 2026. Source-led analysis; see the article's original references and limitations. /articles/prompt-not-market-ai-visibility-measurement
- [40] AEO Updates. AI search reaches billions. That still does not tell us its market share. August 30, 2026. Source-led analysis; see the article's original references and limitations. /articles/ai-search-reaches-billions-no-single-market-share
- [41] AEO Updates. The AI visibility gold rush is missing the bigger prize. September 28, 2026. Source-led analysis; see the article's original references and limitations. /articles/ai-visibility-human-choice-measurement
- [42] AEO Updates. Nine million AI answers. Seven things brands should do now. October 4, 2026. Source-led analysis; see the article's original references and limitations. /articles/nine-million-ai-answers-seven-aeo-actions
- [43] AEO Updates. What happens to your website when AI builds the interface?. October 9, 2026. Source-led analysis; see the article's original references and limitations. /articles/when-ai-builds-interface-brand-evidence
- [44] AEO Updates. How retailers can sell through AI shopping agents. October 9, 2026. Source-led analysis; see the article's original references and limitations. /articles/how-retailers-sell-through-ai-shopping-agents
- [45] AEO Updates. Google wants the cart. Shopify welcomes the agent. Who gets the customer?. October 15, 2026. Source-led analysis; see the article's original references and limitations. /articles/google-shopify-ai-shopping-who-gets-customer
- [46] AEO Updates. The future of customer loyalty: AI could make every renewal a decision to stay or leave. October 14, 2026. Source-led analysis; see the article's original references and limitations. /articles/ai-consumer-advocacy-proactive-bill-review
- [47] AEO Updates. What would AI advise your existing customer to do?. October 11, 2026. Source-led analysis; see the article's original references and limitations. /articles/what-would-ai-advise-your-existing-customer
- [48] AEO Updates. JPMorgan puts up gates against AI agents. What happens when customers want a better deal?. October 13, 2026. Source-led analysis; see the article's original references and limitations. /articles/jpmorgan-ai-agents-customer-loyalty
- [49] AEO Updates. When AI starts selling, what happens to trust?. October 17, 2026. Source-led analysis; see the article's original references and limitations. /articles/ai-advertising-personalisation-trust
- [50] AEO Updates. AI recommendations are not permission to act. October 17, 2026. Source-led analysis; see the article's original references and limitations. /articles/ai-agent-access-discover-compare-prepare-act
- [51] AEO Updates. Six emerging AEO trends: from visibility to customer loyalty. October 15, 2026. Source-led analysis; see the article's original references and limitations. /articles/six-emerging-aeo-trends-visibility-customer-loyalty
- [52] Google. Holiday shopping updates. Official Universal Commerce Protocol and checkout description; product-status claims retained within the original Google and Shopify analysis. Accessed in the October review. https://blog.google/products-and-platforms/products/shopping/google-shopping-updates-holiday-shopping/
- [53] OpenAI. GPT-6 for everyone, 7 October 2026. Intelligent UI product announcement, not evidence of customer adoption or commercial uplift. https://openai.com/index/gpt-6-for-everyone/
- [54] Shopify. Shop.md shopper-agent setup instructions. Official product documentation; compatible agents, account connection and buyer approval remain distinct. Reviewed 15 October 2026. https://shop.md/
- [55] Rocket Money. What is Rowan? Official product description, September 2026. Proactive finance and recurring-bill assistance with human verification; not independently audited savings or switching evidence. https://www.rocketmoney.com/learn/personal-finance/what-is-rowan
- [56] Google Search Help. Personal Intelligence in Search. Advertising and connected-app FAQs, reviewed 17 October 2026. Distinguishes direct use of Gmail or Photos from summaries, excerpts and inferences. https://support.google.com/websearch/answer/16859283
- [57] Salvi, Cuevas and Horta Ribeiro. Sponsored recommendations in agentic search, 2026. Two preregistered experiments, 2,012 final participants; task-specific selection and detection, not live retail conversion. https://arxiv.org/abs/2604.04263
- [58] Nuremberg Institute for Market Decisions. How sponsored content affects consumer decisions in AI-powered retail. Tablet experiment summary, 555 participants in Germany, Austria and Switzerland; cell sizes and inferential statistics not published in the summary. https://www.nim.org/en/publications/detail/how-sponsored-sontent-affects-consumer-decisions-in-ai-powered-retail
- [59] Cloudflare. Bots: bot types and behaviours. Search, Agent and Training classifications; a bot may have more than one behaviour. Reviewed 17 October 2026. https://developers.cloudflare.com/bots/concepts/bot/
- [60] Visa. Trusted Agent Protocol. Official recognition and message-integrity documentation; not proof of universal customer consent, spending authority, merchant acceptance or completed settlement. Reviewed 17 October 2026. https://developer.visa.com/docs/trusted-agent-protocol
- [61] Moneycontrol, citing Bloomberg. Schwab blocks unauthorised AI agents, 16 October 2026. Reported restrictions and planned secure read/research connections, not a deployed universal account-access policy. https://www.moneycontrol.com/news/business/after-jpmorgan-charles-schwab-blocks-unauthorised-ai-agents-from-accessing-customer-accounts-14054121.html
- [62] Gartner. Consumers want AI shopping help but not AI purchase decisions, 27 May 2026. January survey of 322 US consumers; category-specific stated willingness, not observed delegation. https://www.gartner.com/en/newsroom/press-releases/2026-05-27-gartner-survey-finds-consumers-want-ai-shopping-help-but-not-ai-purchase-decisions