Every year around late September, the retail industry’s three big measuring sticks come out: Adobe’s transaction-level ecommerce forecast, Salesforce’s Shopping Index predictions, and the Mastercard Economics Institute’s SpendingPulse read. This year, for the first time, all three are describing the same underlying phenomenon. Holiday 2026 is the first agentic holiday season: the first peak shopping period where AI sits between the shopper and the purchase not as a novelty but as infrastructure.

Adobe projects a record $275.1 billion in US online sales between November 1 and December 31, up 6.7% year over year, with ecommerce taking $1 out of every $4 spent on retail goods this season. Salesforce predicts that 20% of all holiday ecommerce traffic will originate from AI chat agents. Mastercard’s Economics Institute, forecasting 5.5% US retail growth, the strongest since 2022, went as far as titling its report “AI assists and last-minute lists.”

Three independent data operations, three different methodologies, one shared conclusion: the agent has entered the cart. What none of the three forecasts examines is the layer underneath all of it, the evidence those AI recommendations are built on. That omission matters more this year than in any year prior, because 2026 is also the first holiday season where the recommendation itself has largely replaced the human’s own vetting behavior. The numbers explain why.

What the forecasts actually say

Start with Adobe, because its Digital Insights forecast is built on direct measurement of consumer transactions rather than surveys. The headline figures:

  • $275.1 billion in US online sales, November 1 through December 31, up 6.7% year over year, with a “Black Friday’s worth of spending” pulled into October as early discounts proliferate
  • AI-driven traffic to retail sites grew 127% year over year in August 2026, and Adobe projects it to grow 130% year over year across the holidays and 141% on Thanksgiving Day specifically
  • 69% of shoppers say they are less likely to return an item they bought with the help of an AI assistant
  • BNPL spend is forecast to hit a record $21.3 billion, up 6.6%, more than an additional Cyber Monday’s worth of installment spending versus 2025
  • Discounts build earlier and deeper than ever: 14% off in early November, 21% just before Thanksgiving, 30% at Cyber Week

Salesforce’s Shopping Index data adds the agent-side detail. Over the past twelve months, 21% of shoppers engaged a third-party AI chat assistant, 17% used a social media AI assistant, and 12% used a brand-owned one. Consumer reliance on AI assistants as the first stop in the shopping journey grew 200% in a single year. Traditional search engine and marketplace usage each fell 15% year over year while “new channels” such as AI assistants and social surged 38%. And the trust numbers are the striking part: 50% of shoppers now report using an AI assistant somewhere in their buying journey, up 67% year over year, and 74% say they trust the product recommendations they receive from AI chat.

Mastercard rounds out the picture from the payments side. Its economists segmented “AI power users,” defined as consumers with paid AI subscriptions, and found they spread holiday spending across a broader set of merchants, including smaller and boutique retailers, and spent a larger share of their holiday dollars before Thanksgiving than non-AI users. Meanwhile, in the category to watch, electronics, prices are moving the wrong way: the PCE price index for video, audio, photo, and information-processing equipment is up 12.2% year over year, a sharp reversal for a category where prices have historically declined, driven partly by AI-infrastructure demand for memory chips. Electronics spending is already up 10.7% in 2026 to date.

Put together, the forecasts describe a holiday shopper who starts with an agent, trusts the agent, buys earlier and across a longer tail of merchants, and, in the highest-priced, most manipulated category on the list, is spending more money per item than last year.

The statistic that should stop everyone: 69% fewer returns, 74% trust

Buried in Adobe’s data is the number that changes the economics of product deception. When 69% of shoppers say they are less likely to return an AI-assisted purchase, the correction mechanism that has always quietly disciplined online retail, the returned box, the refund request, the one-star follow-up review, is being switched off precisely where the purchase decision was delegated.

Salesforce’s trust data explains the other half of the loop. When shoppers interact with a brand-owned AI assistant, 41% say it makes them “much more confident” in a purchase, to the point where it replaces the need to read reviews entirely. Another 36% report being “somewhat more confident.” Only 5% distrust the answers outright.

Read those two findings as a system. The shopper stops reading reviews because the agent read them. The shopper stops returning items because the agent’s confidence transferred to the purchase. What remains as the sole quality gate between a product and $275 billion is the corpus the agent consulted: the star ratings, the review text, the seller history, the price stability data. If that corpus is clean, this is the most efficient holiday season ever run. If it is contaminated, there is no feedback signal that will catch it before January, and per Adobe’s own numbers, not much of one after.

This is not a hypothetical worry. The FTC’s Rule on the Use of Consumer Reviews and Testimonials, finalized in August 2024, banned fake and AI-fabricated reviews with civil penalty authority precisely because, in the Commission’s words, “fake, false, or otherwise deceptive reviews and testimonials have polluted the marketplace.” The rule exists because the pollution exists. Enforcement actions move at the speed of government; holiday shopping does not.

When the reader is a machine, manipulation scales

The economics of a fake review have historically been unimpressive per unit: one paid five-star post nudges one browsing human, maybe. The agentic season changes the return on investment, in three specific ways visible in this year’s forecasts.

One: leverage per review. When 20% of holiday traffic flows through agents, and those agents summarize review corpora for dozens or hundreds of shoppers each, a single batch of fabricated reviews no longer influences one purchaser at a time. It contaminates a summary that gets multiplied. The manipulator’s addressable audience per fake review went from one reader to every downstream user of any agent that ingested the corpus. Fake review operations become wholesale instead of retail.

Two: zero human skepticism in the loop. A human skimming reviews applies cheap heuristics: too generic, too clustered, all five stars, weird grammar. A shopper who has stopped reading reviews entirely, per Salesforce’s 41%, applies nothing. And the agent itself is optimizing for relevance and consistency of the corpus, not for authenticity of it. A well-formed fake review is, from a language model’s perspective, indistinguishable from, and often cleaner than, a genuine annoyed human one.

Three: the long tail is where the armor is thinnest. Mastercard found AI power users spreading spend to smaller and boutique merchants. That is genuinely good news for honest small sellers, and it is equally good news for dishonest ones. Long-tail listings are exactly where review histories are short, where a few dozen fabricated posts can flip a rating decisively, and where neither platform moderation nor FTC enforcement has the cycles to look. The K-shaped dynamics Salesforce documents cut the same way: budget-stressed shoppers using AI defensively to compare prices are the least likely to double-check, and premium shoppers using AI for curation are delegating the most judgment.

Salesforce’s own advice to merchants makes the exposure explicit: “If your product data isn’t integrated into the feeds powering these agents, your inventory won’t be part of the consideration set.” True, and incomplete. Being in the consideration set means your data, and everyone else’s, is being consumed uncritically at machine scale. Feed quality is now a merchandising concern. Evidence quality is a trust concern. The industry is investing heavily in the first and barely in the second.

The forecast nobody published: agent-influenced revenue, by evidence quality

Here is the analysis gap in all three reports. Adobe can measure that AI traffic is up 130% but not what share of AI-influenced purchases rested on manipulated listings. Salesforce predicts 20% of traffic from agents, including, in its own words, “competitor scrapers fueling algorithmic price-matching,” but publishes no estimate of how much of that 20% will act on contaminated review data. Mastercard sees AI power users spreading to boutique merchants but cannot see whether those merchants’ ratings were earned or assembled.

None of this is a criticism of the forecasts; measuring the evidence layer is not their job. It is an observation that the industry is flying its biggest season ever on instruments that measure demand beautifully and merit not at all. The reasonable estimate, based on what platforms themselves have disclosed about fake review volumes removed, is that the contamination rate is not zero, and every point of it now compounds through agent summaries instead of dissipating through human skimming.

There is a second-order effect worth naming for merchants playing the long game. Salesforce reports that retailers with branded AI shopper agents saw 59% higher holiday sales growth in 2025, +6.2% versus +3.9%. The agent gold rush is real. But an agent that recommends a badly reviewed product burns the very trust that made the agent valuable: 74% of shoppers trust AI recommendations today partly because the technology is new and the failure statistics are not yet famous. The first widely covered scandal of the agentic season, the AI shopping guide that steered thousands of buyers to a manipulated listing, will move that 74% faster than any forecast model. Brands that survive that moment will be the ones that can show their recommendations stood on verified evidence, not just on whatever the corpus said that morning.

What each side should do before Cyber Week

For merchants: treat your review corpus as part of your agent feed. Salesforce is right that unstructured product data means invisibility; contaminated product data means visibility with a time bomb attached. Audit what your own listings’ ratings are built on before an agent audits it for you, badly.

For agent builders: the difference between a toy and infrastructure is that infrastructure survives contact with adversarial data. An agent that recommends from raw marketplace ratings is consuming the single most manipulated consumer dataset in existence. Filter for review authenticity before scoring, weight verified purchase history, prefer stable signals over bursty ones, and log what evidence informed each recommendation so the decision can be reconstructed when a customer asks why. That last part is not optional polish; as payment networks build agent registration and cart-context standards, decision provenance is what makes an agent insurable.

For shoppers: you are the control group. The 5% who distrust everything and the quiet majority who never checked are both exposed to the same corpus. The move that costs nothing is to check what the agent checked: whether the rating survived fake-review filtering, whether the score reflects review quality or review count, whether the seller has history. Ten seconds of independent verification beats a January return window.

The layer the season is missing

GoBuy exists because this exact gap, the distance between what a listing claims and what the evidence supports, widens whenever mediation increases. Fake and incentivized reviews are filtered out before any score is computed, authentic ones are weighted up, and the Smart Score runs 0 to 100 on the quality of the evidence rather than the quantity of stars, which is the one thing a review farm cannot manufacture on demand. Sustained performance is the threshold that matters: the GoBuy Verified badge requires a score above 80 held across 90 days, a window that covers an entire holiday season and cannot be burst-campaigned through. And because a trust layer only works if it reaches the decision point, GoBuy ships where the decisions happen: only the top seven verified products per category, a Chrome extension that puts the trust panel directly on the Amazon page, and an MCP server at gobuy.ai/api/mcp so the agents handling that 20% of holiday traffic can consult filtered evidence as a step before checkout instead of after a dispute.

Holiday 2026 will be recorded as the season the recommendation became the shelf. $275.1 billion will flow through it, one in five ecommerce visits will originate from it, and three-quarters of shoppers will trust it. Whether it deserves that trust will be decided by whatever is underneath it. Check the shelf before you buy, and if you build the agents doing the shelving, wire them to evidence that survives scrutiny at gobuy.ai and gobuy.ai/agent-docs.