Customers Research With AI. They Buy Somewhere Else

Opening: a feature died, and the data already explained it

In early March, OpenAI quietly shelved Instant Checkout — the feature it launched in September 2025 to let people buy products without leaving ChatGPT. Coverage from TechCrunch and a Forrester post-mortem pinned the cause on a behavioral pattern that had become uncomfortable to ignore: ChatGPT users were researching products inside the chatbot, but very few of them were closing the loop on a purchase there. Roughly a dozen merchants ever went live before the program was wound down.

The shutdown is not a story about OpenAI's roadmap. It is a story about how shoppers behave when an AI sits between them and a checkout page. And it landed on top of a survey wave that had already telegraphed the answer. Salsify's 2026 Consumer Research, fielded in October 2025 across 2,712 shoppers in the United States, Canada, and the United Kingdom, found that 22% of shoppers use AI tools to research products — but only 14% trust AI recommendations enough to rely on them regularly. The eight-point gap between research and trust is the entire commercial story.

For sellers — especially cross-border operators trying to figure out where to put a marketing dollar — that gap reframes the question. AI is not the destination. It is increasingly the front door. The job is not to sell inside the chatbot. The job is to be the place the shopper lands when they leave it.

The behavioral evidence

1, AI traffic is exploding. AI conversion is not where the traffic is.

Adobe Analytics, which tracks more than a trillion visits across U.S. retail sites, reported that traffic to retail sites from generative AI tools rose 693.4% during the November–December 2025 holiday season versus the prior year, and 670% on Cyber Monday alone. Adobe also flagged that "the base of users remains modest" — a phrase doing a lot of work. A 700% increase off a small base still produces a small absolute number. The growth curve is real; the absolute share of revenue from AI referrals is still single digits.

What sellers should notice is the qualitative signal underneath the headline number. Adobe's data shows AI-referred shoppers spend more time on-site, view more pages, and convert at a higher rate than non-AI traffic. That is not a story about AI as a checkout channel. It is a story about AI as a high-intent qualifier — the customers AI sends are pre-educated, narrowed-down, and closer to a decision when they arrive.

2, Most shoppers don't trust AI to make the call alone

The trust ceiling is the number sellers should not look away from. Salsify's panel found only 14% of shoppers trust AI recommendations enough to act on them without verifying elsewhere. A separate IBM Institute for Business Value/NRF study released in January reported 45% of consumers turn to AI somewhere in the buying journey — but the breakdown is telling: 41% use AI to research products, 33% to interpret reviews, and 31% to hunt for deals. Each of those is a research behavior. None of them is a purchase behavior.

Forrester's Consumer Benchmark Survey adds the long view: 24% of U.S. online adults have used ChatGPT at all. That is the addressable ceiling for any "shop in ChatGPT" play right now. It is not zero — it is meaningful — but it is not "the future of commerce" either. It is a research surface that a quarter of online adults occasionally use.

3, Stores are quietly winning back the certainty trade

The most under-covered finding in the Salsify report is also the most behaviorally important. Daily online shopping frequency dropped from 21% to 9% year-over-year — a 57% collapse in the share of shoppers who buy something online every day. Sixty percent of respondents said they discover new products in physical stores, edging out online marketplaces (57%) and social platforms (52%).

That is not a return to 2014. It is a redistribution of trust. When the online research experience is full of AI summaries, sponsored placements, and reviews of uncertain provenance, the physical store becomes the cheapest way to verify that a product is real, fits, and looks the way the listing promised. The same survey found 45% of Gen Z and 43% of millennial shoppers abandon a purchase when product details don't match what they were shown — a defection rate that does not exist offline because the verification happened before the wallet opened.

4, The fake-review tax is making AI research more useful, not less

A January 2026 Omnisend report found that 84% of consumers still trust online reviews and 33% trust them more than they did two years ago — but 82% encountered at least one fake review in the past 12 months. The FTC followed up on its August 2024 Consumer Review Rule by issuing warning letters to ten companies on December 22, 2025, with violations carrying civil penalties of up to $53,088 per instance.

The fake-review problem is the unspoken reason AI research adoption has scaled even when AI purchase trust hasn't. An LLM that summarizes 200 reviews and surfaces patterns can flag inconsistencies that a tired shopper scrolling on a Tuesday night cannot. AI is becoming the review-cleaning layer above the actual review layer — a behavior closer to "decision support" than "decision delegation."

Why it's happening

The temptation is to call this a transitional phase — shoppers will eventually trust agents enough to delegate the buy. That framing flatters the technology and underestimates the psychology.

Buying is a commitment device. Research is reversible — an open tab, a saved comparison, a screenshot. A purchase is not. The cognitive load of "if this is wrong, I have to return it, and I have to do it on my phone, and I have to remember which carrier picks it up" is the load shoppers are protecting themselves against. AI is brilliant at compressing the research load. It does nothing to reduce the post-purchase load. So shoppers happily delegate the upstream work and reserve the downstream commitment for the channel where they already have a stored payment, a known returns policy, and a customer service surface they recognize.

There is also a generational seam worth naming. Capgemini's 2026 consumer research found 71% of consumers want generative AI integrated into shopping experiences and 76% want clear rules for when an AI assistant acts on their behalf (Capgemini sells AI consulting services — treat the directional appetite as real and the magnitude as upper-bound). The desire is enthusiastic. The conditions are strict. "Help me compare" tracks. "Buy without asking" does not.

Layer in the fake-review erosion and the macro caution Salsify documents — 39% of shoppers comparing prices more carefully, 38% reducing spending in specific categories, 37% trading down to lower-priced alternatives — and the behavior pattern resolves: shoppers are doing more pre-purchase work, not less, because the cost of getting it wrong has gone up.

What it means for sellers

1. Optimize for AI citation, not AI checkout. The unit of work that pays off in 2026 is being the product an LLM names when a shopper asks "which one should I get." That requires structured product data, machine-readable specifications, durable spec sheets, and the kind of FAQ content that an LLM can lift wholesale. It does not require a new checkout integration. Salsify's panel reported 31% of shoppers are convinced to buy when AI delivers detailed product descriptions and clear specifications — the conversion lever is information density, not interactivity.

2. Treat AI traffic as your highest-quality cohort and instrument it accordingly. Adobe's holiday data shows AI-referred visitors browse longer and convert better. Most analytics setups still bucket them as "referral/other." Tag them. Build a separate landing experience for them — one that doesn't repeat what the LLM already told them. Lead with proof points the chatbot can't render: video, third-party test results, a real return policy.

3. Close the post-purchase loop the chatbot can't see. The reason shoppers leave the AI to buy is friction they don't trust the agent to handle: returns, sizing, fit, replacement parts. A 2-line returns policy on the PDP, an owned WhatsApp/SMS support line, and a visible warranty number do more for AI-referred conversion than any prompt-engineering exercise.

4. Audit your review surface this month. Eighty-two percent fake-review encounter rate is not background noise — it is the macro condition the FTC is now actively litigating. Sellers running review-incentive programs, gated negative reviews, or undisclosed seeding should assume enforcement risk has moved from theoretical to operational.

5. Don't chase agentic commerce destinations. Chase being the destination. The OpenAI Instant Checkout shutdown is a warning shot for any seller building bespoke integrations into a single AI platform. The 2026 model that worked was the model where ChatGPT, Perplexity, or Gemini sent a qualified visitor to a Shopify, Amazon, or TikTok Shop product page. That model rewards owned-storefront discipline, not platform bets.

What to watch next

The clearest leading indicator that this gap is closing — that shoppers are starting to trust AI for the buy, not just the search — will not come from press releases. It will come from one number: the share of agentic-commerce sessions that complete a purchase without the user navigating to the merchant. Walmart's announced Sparky-into-ChatGPT integration and Amazon's expanded Rufus Scheduled Actions in April are the first tests at meaningful scale. If conversion rates inside those flows surface in earnings calls, the gap is narrowing.

The clearest indicator that the gap is widening is the opposite: rising AI-referred traffic to retail sites with stagnant or falling AI-completed purchases. Adobe's next quarterly retail traffic report and the Salsify mid-year refresh are the two reads worth watching.

For sellers, the actionable question this week is narrower. Pull the last 90 days of traffic by referrer. Find the AI-referred bucket. Look at its conversion rate, time-on-page, and product-detail engagement separately from your search and social cohorts. If the AI-referred cohort is converting better than average — which most operators will find — that is the cohort whose landing experience deserves the next sprint of attention. The shoppers researching with AI are already telling you they are ready to buy. They just want to do it on a page that looks like yours.