Two consumer studies published within four days of each other describe the same phenomenon from two sides of the Atlantic, and both buried the most important finding under a hopeful headline.

On August 14, Deloitte published “The Human and the Agent: The state of Agentic Commerce in Europe,” a survey of 13,500 consumers across 15 countries. Its headline number: 56% of European consumers have already used AI to shop at least once. AI-assisted shopping crossed the 50% adoption mark in roughly 18 months, a transition that took e-commerce about five years and smartphones about ten. Deloitte calls it “the fastest technology adoption European retail has seen.”

Three days earlier, on August 10, performance marketing firm RTB House released “Who’s Buying? Consumer Trust in the Age of Agentic AI,” a survey of 1,840 shoppers across the United States, the United Kingdom, France, and Japan conducted with research firm Cint in June and July 2026. Its headline finding: AI tools have overtaken TikTok, Instagram, Facebook, newspapers, and major media outlets as trusted sources for purchase decisions. In the United States, 44% of consumers now trust AI tools for purchase decisions, second only to friends and family at 59%.

The industry read both reports as a validation of AI shopping. The correct read is more uncomfortable. Consumers have transferred their trust to the fastest, least verifiable information channel in the history of retail, at the exact moment when the data that channel reads is the most manipulated it has ever been.

The Comparison Layer Changed Owners

The most consequential number in the Deloitte study is not the 56% adoption figure. It is this: among shoppers who use AI, 57% use it to compare products, making AI the number one comparison channel in Europe, ahead of marketplaces, retailer websites, and physical stores.

Search still owns research. Social media still owns inspiration. But comparison, the step where consumers decide between functionally similar products with wildly different quality, is now owned by AI assistants.

This matters because comparison is where manipulation concentrates. Amazon’s search results are a blend of organic ranking and sponsored placement. Star ratings aggregate genuine reviews with incentivized, bot-generated, and repurposed ones. Product listings are optimized by professionals whose job is conversion, not accuracy. The comparison layer was never clean, but at least a human doing the comparing could apply skepticism, open a second tab, or notice that 40% of the five-star reviews were written in the same week.

When AI owns comparison, the skepticism layer either exists inside the agent or it does not exist at all. Today, in almost every shipping product, it does not exist. The agent reads the manipulated inputs and produces a confident comparison. The consumer, who now trusts AI more than newspapers, acts on it.

The Trust Inversion Nobody Priced In

The RTB House data describes what happens when an entire generation of consumers gets burned by the old channels. Influencer feeds turned into undisclosed advertising. TikTok reviews turned into performance. Facebook groups turned into affiliate farms. Trust in those channels collapsed, and AI tools inherited the displaced trust.

The survey’s trust hierarchy for US consumers: friends and family at 59%, AI tools at 44%, then influencers, newspapers, major media outlets, TikTok, and Instagram trailing behind. Among the AI tools themselves, Google AI Overviews and ChatGPT lead at 43% each in the US, followed by Claude at 23% and Grok at 21%.

Here is the structural problem the survey does not address: trust in a friend is direct trust, earned through a track record of that person’s judgment. Trust in an AI tool is derivative trust. The AI did not use the product. It aggregated and summarized sources, including product pages, sponsored reviews, star ratings, and the accumulated debris of a marketplace economy that the FTC has spent the past year litigating against. Consumers are not really trusting the AI. They are trusting the AI’s ability to filter the inputs. And the AI, in most cases, has no filtering layer at all.

The RTB House finding that should worry every retailer and regulator: 68% of surveyed shoppers used at least one AI platform to shop in the past three months, and 62% used AI specifically to compare prices, brands, models, and reviews. Review data is now a direct input into the most trusted decision layer in commerce. Fake reviews are no longer a reputational nuisance at the bottom of a product page. They are upstream inputs to the system consumers trust more than newspapers.

The Consideration Engine: AI Makes Buying Slower, Not Faster

The single most counterintuitive finding in the RTB House study, and the one most of the coverage missed: AI does not compress the path to purchase. It stretches it.

42% of US consumers and 33% of non-US consumers say AI tools increase the time it takes them to reach a final purchase decision, because AI surfaces more brands, products, and alternatives to consider. The effect is strongest among Gen Z at 48%, but it holds across every generation measured. With US shoppers typically visiting a website four to six times before completing a purchase, AI is widening that consideration window rather than closing it.

For a decade, e-commerce optimization was built on the opposite assumption: reduce friction, collapse the funnel, get to checkout before doubt sets in. RTB House’s VP of Product Marketing and Analytics Jaysen Gillespie drew the conclusion bluntly: “For more than a decade, e-commerce growth was driven by buying attention on social feeds and optimizing for instant checkouts. This research proves that era is coming to an end.”

An extended consideration window is, in principle, good for consumers. More time to compare means more opportunity for quality to win over marketing spend. But it changes what the bottleneck is. When the funnel was fast, the scarce resource was attention. Now the scarce resource is verification. Shoppers are spending more time than ever considering options that were ranked by an AI that read data it cannot verify.

The $250 Datum: How Consumers Price the Trust Gap

Buried in the RTB House data is the most economically revealing number in the whole study: 42% of US millennials said they would let an AI agent buy items on their behalf within a $250 budget, but only if the purchase could be returned within seven days. Without that safeguard, the number drops to roughly a third.

Read that carefully. The single most requested safeguard globally, per the study, is human approval before checkout. Among all generations, 35% want a human to review the transaction before an agent completes it, rising to 44% among baby boomers.

Consumers are behaving like rational risk pricers. They do not fully trust the agent’s judgment, so they hedge: a capped budget, a return window, a human checkpoint. The seven-day return guarantee functions as a manual, after-the-fact trust layer. It converts “I trust the agent” into “the worst case costs me a trip to the post office.”

This is an unstable equilibrium. Return windows work as a trust substitute only while deception is rare enough that returns remain exceptional. If manipulated listings push agents toward bad recommendations at scale, the hedge stops being cheap: for the consumer, in shipping and restocking; for the retailer, in reverse logistics. The industry is treating reversibility as a substitute for verification. It is actually a loan against unverified recommendations, and someone always pays it back.

The Embedded Agent Problem: Amazon’s 40% Wedge

The third data point this month comes from Amazon’s Q2 2026 earnings call, reported by Retail Dive on August 3. More than 350 million shoppers used Alexa for Shopping, the agentic assistant that replaced Rufus in May, over the past twelve months. Active users nearly doubled year over year in the second quarter, and interactions grew fivefold.

Then the number that should reframe every conversation about agentic commerce: US customers who use Alexa for Shopping spend 40% more per order on average than shoppers who do not use the assistant. Shoppers who try Alexa+ join Amazon Prime at nearly 25% higher rates. CEO Andy Jassy’s summary: “We find that everywhere Alexa goes, it drives momentum for the business.”

Amazon’s agent makes Amazon more money per order. That is not a scandal; it is a business model, and it is working spectacularly. But it does mean the fastest-growing shopping agent in the world is an embedded agent with a structural commercial interest in the outcomes it recommends. The consumer’s trust hierarchy now places AI tools above newspapers and media outlets, and the leading shopping agent by volume belongs to the marketplace it recommends from.

Combined with Deloitte’s framing that retailers “no longer serve one customer, but two: the human, and the agent advising them,” the strategic race is obvious. Every marketplace is now incentivized to build the agent that advises its own shelf. Independent judgment is not a feature anyone’s roadmap prioritizes, because independence does not increase average order value by 40%.

What Deloitte’s Own Data Says the Concern Is

Deloitte’s survey does register consumer anxiety, it just locates it in the wrong layer. Data privacy and security is the leading concern across all European shoppers, running from 32% among Gen Z to 42% among Gen X, alongside worries about bias, manipulation, and the loss of human touch. Morris Boermann, the report’s business lead, concluded that “trust is earned through experience, not messaging.”

Privacy is a real concern. But it is a concern about what the agent does with the consumer’s data. The systemic risk nobody surveyed for is what the marketplace does to the agent’s data. The manipulation flows the other direction: fake reviews, inflated reference prices, and sponsored rankings are the inputs the agent consumes, summarizes, and confidently serves to a consumer who trusts it more than the press.

Boermann is right that trust is earned through experience. The experience of being recommended a product whose 4.8 stars were manufactured last month is precisely how the 44% trust figure will erode. The only question is whether the industry fixes the input layer before or after that erosion begins.

The Verification Layer Has to Be Independent, and It Has to Be Queryable

If AI is now the top comparison channel, then comparison quality depends on the quality of what the agent reads. Three things follow.

First, review authenticity has to be assessed before the agent consumes the data, not after the consumer complains. Fake review detection that lives downstream of the recommendation is decoration. GoBuy’s Smart Score works upstream: review data is filtered for authenticity first, then scored 0-100 on the quality of verified reviews, not the quantity of all reviews. A product with 12,000 reviews and a manipulated profile scores what it deserves, not what its volume implies.

Second, trust data has to be independent of the marketplace and the agent. An embedded agent optimizing for a 40% higher order value cannot also be the neutral scorer. A marketplace selling sponsored placement cannot be the neutral scorer. The scorer must have no commercial stake in the recommendation. This is the same structural logic that made credit rating agencies and independent labs necessary in every other mature market, applied to the comparison layer of commerce.

Third, the trust layer has to be queryable by agents in the protocol agents already speak. This is why GoBuy delivers Smart Scores, review authenticity analysis, and a curated top-7 ranking per category through an MCP server at gobuy.ai/api/mcp. Any agent that can call a tool can ask, before recommending a product, whether the review data behind it survives filtering. The 90-day sustained score requirement behind the GoBuy Verified badge also closes the flash-manipulation loophole: a burst of fake five-star reviews cannot carry a badge that takes three months of verified quality to earn.

Deloitte’s closing warning to retailers is that those who wait “risk being filtered out before a customer ever sees them.” The same sentence applies with more force to products. In a world where the agent pre-filters the shelf, the products that survive will be the ones whose trust data is verifiable, not merely voluminous.

The Uncomfortable Summary

Three findings, three sources, one conclusion:

  • 57% of AI-using European shoppers rely on AI as their number one comparison channel (Deloitte, 13,500 consumers, 15 countries).
  • 44% of US consumers trust AI tools for purchase decisions, ahead of influencers, newspapers, and every social platform measured (RTB House, 1,840 respondents, four countries).
  • The largest deployed shopping agent belongs to the marketplace it recommends from, and its users spend 40% more per order (Amazon Q2 2026 earnings).

Trust in AI shopping is not a bubble because the technology fails. It is a bubble risk because the inputs are manipulated, the embedded agents have commercial interests, and the consumer’s hedge, a return window, only works while bad recommendations stay rare. The fix is not less AI in shopping. It is independent, agent-queryable verification attached to every comparison an agent makes.

If you are building a shopping agent, do not ship it on marketplace data alone. Connect it to an independent trust layer first. The MCP endpoint is at gobuy.ai/api/mcp, and integration docs are at gobuy.ai/agent-docs. Your users already trust you more than the media. Start acting like it.

Sources: Deloitte, “The Human and the Agent: The state of Agentic Commerce in Europe”; RTB House, “Who’s Buying? Consumer Trust in the Age of Agentic AI” via Retail Dive and New Digital Age; Retail Dive, “Amazon customers are embracing Alexa for Shopping”.