By Tom Morgan · Last updated October 21, 2026 · About 14 minutes to read
AI-personalized shopping is the use of machine-learning models to change what each shopper sees (search results, product recommendations, offers, and sometimes prices) based on behavior and context, instead of showing every visitor the same storefront. The term also covers shoppers who use AI assistants to find and compare products before they ever reach a retailer’s site.
Most coverage of this topic stacks percentages that sound like they confirm each other. Many measure different things. This piece traces the main 2026 numbers to where they came from, summarizes what controlled experiments found, and builds a small model for a question most dashboards never ask: when does a rise in returns cancel out a gain in revenue?
I checked each figure against the original report or paper wherever I could reach one. Where I could only reach a secondary source, or could not verify a claim, the text says so, and a full list sits at the end.
Update, October 21, 2026: The pricing section has been rewritten. State legislatures moved from proposals to enacted law in 2026, so the old “no bright line” advice no longer holds. The break-even model now includes more lift scenarios, a closed-form formula, a spreadsheet version, and a stress test on return costs. The holiday-traffic figures now include Adobe’s September 2026 forecast.
In this article
- Key takeaways
- The adoption numbers: who measured what
- What shoppers actually do
- AI-referred traffic: what Adobe measured
- Does personalization pay? Controlled evidence
- The returns problem
- The break-even model (with a spreadsheet version)
- A maturity self-check
- Pricing: personalization vs. surveillance pricing in 2026
- Agentic commerce forecasts
- Trust, consent and what goes wrong
- A practical roadmap
- FAQ · Glossary · What I could not verify · Sources · Method and update log
Key takeaways
- Several headline adoption stats are not retail-specific. The “88% use AI, 7% fully scaled” pair that Stord’s 2026 report leads with is footnoted in that report to McKinsey’s cross-industry survey.
- “89%” answers a different question. It comes from NVIDIA’s retail and CPG survey and measures respondents who say AI increased revenue, not who use AI.
- The best controlled evidence points to the training target. In a live A/B test at DigitecGalaxus, optimizing a recommender for purchases produced a GMV uplift more than five times larger than optimizing for clicks.
- The click-to-returns link is unproven. I found no study showing that click-trained recommenders raise return rates. It is a plausible, testable hypothesis, and this article shows how to test it.
- A revenue lift survives return-rate drift only up to a point. In the model below, a 5% lift is erased by a rise of about 2 points in return rate, and a 20% lift survives about 7.5 points. Under these assumptions, each 1% of sales lift buys roughly 0.45 points of return-rate headroom.
- Individualized pricing is now regulated, not just debated. Three states had enacted limits or bans by mid-2026, and New York’s broader ban was awaiting signature at the last check. Details are in the pricing section.
The adoption numbers: who measured what
Three sources produce the figures in the high 80s and low 90s, and only two of them did the measuring. They are not measuring the same thing.
| Figure | Source | What it actually measures | Population |
|---|---|---|---|
| 88% | McKinsey (2025 survey, repeated by Stord) | Regular use of AI in at least one business function | All industries, global |
| “Nearly nine in ten” | McKinsey, State of AI, Aug. 25, 2026 | Same question, newer edition | 1,719 respondents, 97 countries, fielded May 4 to June 8, 2026 |
| 89% | NVIDIA, State of AI in Retail and CPG 2026 | Said AI increased annual revenue | Retail and CPG professionals (“hundreds” of responses) |
| 95% | NVIDIA (also repeated by Stord) | Said AI decreased annual costs | Same NVIDIA sample |
| 58% | NVIDIA | “Actively deploying” AI, up from 42% in 2024 | Same NVIDIA sample |
1. Stord’s adoption headline is largely McKinsey’s data. Stord, a fulfillment and commerce-technology company, published its State of AI in E-Commerce report on February 11, 2026. Its page leads with “88% of organizations use AI in at least one core function, up from 78%,” and “only 7% fully scaled.” The report’s footnotes credit those figures, and the experimenting/piloting/scaling split, to McKinsey, not to a Stord survey. They describe organizations in every industry, not e-commerce brands. Other figures on the page trace to a leadership blog (the 92% and 99% investment and framework stats), a vendor blog (the “40% higher revenue” claim), a returns-software blog (the “30% higher customer lifetime value” claim), and NVIDIA (the 95% cost figure).
What is original to Stord is its consumer sentiment survey, covered below. Stord also sells fulfillment, order-management, and returns software, the category its report says brands need. That does not make the report wrong, but it is useful context for how the findings are framed.
2. “89% of retailers use AI” is a misreading. Search 2026 coverage, and you will find “89% of retailers have adopted AI” credited to McKinsey or Stord. It is NVIDIA’s figure, and it answers a different question: did AI raise revenue? NVIDIA’s own adoption measure is the stricter 58% “actively deploying.” Pages that present 89% as an adoption rate next to 7% as a “scaled” rate are mixing two surveys with different populations and questions.
3. The McKinsey picture is more interesting than “88% vs. 7%.” The August 2026 edition reports that 44% of respondents say AI is scaling across their enterprise (up from 38% a year earlier), that 37% attribute at least some EBIT impact to AI (essentially unchanged), and that about 6% qualify as “high performers” (5% or more of EBIT attributed to AI). Larger organizations lead: 54% of those with $1B+ in revenue report enterprise scaling, versus about one-third of smaller ones. Usage is broad, scaling is growing, and measurable financial impact is still concentrated.
What shoppers actually do (and why surveys disagree)
Stord’s consumer survey reports that 51% of U.S. consumers used AI for online shopping in 2025, up from 38% in 2024 (Stord attributes the 2024 figure to Adobe), and that 17% use it regularly. By generation, 37% of Gen Z and 29% of Millennials actively use AI to shop, versus 12% of Gen X and 5% of Baby Boomers. Stord’s page does not state the sample size or fieldwork dates, so treat these as one vendor’s survey, not a census.
A different survey, the Commerce and Future Commerce New Modes report, asked which tool people prefer when starting product research: 33% of Gen Z and 26% of Millennials chose AI platforms over search engines, versus 13% of Gen X and 3% of Baby Boomers.
The levels differ because “has used,” “actively uses,” and “prefers for research” are three different behaviors. The shape is the same in both: a steep generational gradient. If you plan around one number, record which of the three you mean.
AI-referred traffic: what Adobe measured
Adobe’s holiday 2025 analysis of more than a trillion U.S. retail visits found that generative-AI referrals to retail sites grew 693% year over year in November and December 2025, and that AI-referred visits converted 31% better than other traffic (54% better on Thanksgiving, 38% on Black Friday). Marketing Dive’s summary of the same report adds that revenue per visit from AI traffic was 32% higher, a reversal from 2024, when non-AI traffic was worth 51% more.
The 2026 update. Adobe’s September 2026 holiday forecast projects U.S. online spending of $275.1 billion for November and December (up 6.7%) and expects AI-referred traffic to retail sites to grow 130% year over year. Adobe reported that AI-driven traffic grew 127% year over year in August 2026. Adobe defines this traffic as clicks on links from AI chat and browser services, and the release gives no share of total traffic.
Two cautions:
- This measures traffic arriving from AI assistants. It is evidence about a new acquisition channel, not about on-site recommendation engines.
- It is a modeled forecast from a single analytics vendor, and the channel started from a small base, so large percentage growth can still be a modest share of visits.
Does personalization pay? What controlled evidence shows
The figure repeated most often is that AI-personalization leaders earn “up to 40% higher revenue.” In Stord’s footnotes, it traces to a vendor blog, not a controlled study. Treat it as a marketing claim.
The strongest controlled evidence I found is narrower and more useful. Researchers at DigitecGalaxus, a Swiss e-commerce company with a catalog of more than 25 million products, ran a live A/B/C/D test on its “similar items” recommender starting April 30, 2025, for twelve days, with about 1.8 million block impressions per group. The same ranking model was trained on three different targets: click, add-to-cart or purchase (“order-submit”).
- The model trained on purchase delivered a GMV uplift more than five times larger than the model trained on clicks.
- The click-trained model did not even achieve its own goal, because the baseline (a behavioral-similarity ordering) already captured click patterns.
- The purchase-trained model accepted a small, significant drop in CTR and still won on GMV.
- Product discovery, measured as the share of clicks showing a shopper something new, stayed comparable across variants.
The authors list their own limits: one company, a layout shared across its shops, about two weeks of data, no seasonality. They do not claim the specific numbers generalize. Their point is that the training target is a first-order lever, and that the correlation between CTR and revenue is weaker than practitioners assume.
Look at what the target list contains: click, add-to-cart, purchase. Nothing in it describes what happens after delivery.
The objective-function problem: returns
If a model is rewarded for purchases and never told whether the purchase was kept, it has no reason to care. That is the setup for a margin leak. It is worth being precise about what is known and what is not.
What the evidence supports
- Returns are large. The National Retail Federation and Happy Returns estimate $849.9 billion in U.S. returns for 2025, about 15.8% of total sales and 19.3% of online sales, with 9% of returns judged fraudulent (roughly $76 billion).
- Expectation mismatch drives returns. A review of fashion returns reports that, in earlier work by Hong and Pavlou (2014), product-fit uncertainty had a stronger effect on returns than quality uncertainty. A two-phase survey study of online fashion shoppers (609 responses in phase 1) found post-purchase satisfaction was the only significant predictor of return intention. That is correlational survey evidence.
- Fit-aware modeling reduces returns. Zalando’s SizeFlags used large-scale A/B testing across 14 countries to reduce size-related returns. A MIT Sloan write-up describes a model that added product images to return prediction, improving predictions by 13.5% and, applied to online offerings, profits by more than 8%.
What is not established. I found no published controlled study showing that click-trained recommenders raise a retailer’s return rate. The claim that “maximizing add-to-carts quietly pushes returns up” is a hypothesis. It is plausible given the evidence above, and cheap to test.
How to test it. Hold out a slice of traffic on a ranker variant that includes post-purchase signals (kept versus returned, or net margin after returns), and compare net margin per session across variants. Wait at least one full return window plus a buffer before reading the result, or the returns have not arrived yet.
The math: when does a return-rate increase cancel a revenue lift?
This model is a planning tool, not a case study. It uses one hypothetical mid-market store and one explicit set of assumptions so you can swap in your own.
Setup
- 500,000 monthly sessions, 3.1% conversion, $87 average order value: $1,348,500 gross monthly sales (15,500 orders).
- Return rate r = 19.3%, the NRF online figure.
- Gross margin on kept sales m = 45%. Cost of handling a return c = 25% of the returned order’s value (shipping, handling, markdowns). Both are illustrative.
- Returned goods go back into stock, and c is assumed to cover everything lost on them, including write-downs. If a meaningful share of returns can’t be resold, raise c (see the stress test below).
Formula
Monthly contribution = Sales × [ (1 − r) × m − r × c ]
= Sales × [ m − r × (m + c) ]Baseline contribution: $424,643 per month (31.49 cents per sales dollar).
The break-even return rate for a sales lift L has a closed form:
r_breakeven = ( m − k₀ / (1 + L) ) / (m + c) where k₀ = m − r × (m + c)Results
How far can the return rate rise before the net gain is gone?
| Sales lift | Break-even return rate | Headroom vs. 19.3% | Net change if returns rise 2 points |
|---|---|---|---|
| +2.5% | 20.4% | +1.1 pts | −$8,735 per month |
| +5% | 21.4% | +2.1 pts | +$1,409 per month |
| +10% | 23.4% | +4.1 pts | +$21,697 per month |
| +15% | 25.2% | +5.9 pts | +$41,986 per month |
| +20% | 26.8% | +7.5 pts | +$62,274 per month |
| +30% | 29.7% | +10.4 pts | +$102,850 per month |
Headroom at a glance (each █ is about one point of return rate):
| Sales lift | Return-rate headroom |
|---|---|
| +2.5% | █ |
| +5% | ██ |
| +10% | ████ |
| +15% | ██████ |
| +20% | ████████ |
| +30% | ██████████ |
Sensitivity. Across gross margins of 30% to 60% and return-handling costs of 15% to 40%, the break-even headroom for a 10% lift ranges from about +2.1 to +5.5 points.
Stress test: unsellable returns. If returns can’t be resold and the full cost of goods is lost (c = 55% with m = 45%), a 10% lift has only about +2.3 points of headroom. Categories with high write-off rates, such as fashion and consumer electronics, sit closer to this end.
What this says. The smaller the lift, the less return-rate drift it can absorb. A 5% lift is almost entirely consumed by a 2-point rise, and a 2.5% lift is wiped out by it. An earlier version of this article claimed that trimming returns by two points could outweigh the headline revenue number. That holds for small lifts and fails for large ones, so the honest claim is conditional.
Break-even on platform cost. At these assumptions, each extra dollar of sales contributes about 31.5 cents. A hypothetical all-in monthly cost of $10,000 needs a 2.4% sales lift to break even, $25,000 needs 5.9%, and $50,000 needs 11.8%. These costs are placeholders, not vendor quotes.
Run it in your own spreadsheet
Put your inputs in cells and paste the formulas. They work in Excel and Google Sheets.
| Cell | Label | Example value |
|---|---|---|
| B1 | Monthly sales | 1348500 |
| B2 | Return rate (r) | 0.193 |
| B3 | Gross margin on kept sales (m) | 0.45 |
| B4 | Return handling cost (c) | 0.25 |
| B5 | Expected sales lift (L) | 0.10 |
| B6 | Baseline contribution | =B1*(B3-B2*(B3+B4)) |
| B7 | Break-even return rate | =(B3-(B3-B2*(B3+B4))/(1+B5))/(B3+B4) |
| B8 | Headroom (points) | =(B7-B2)*100 |
| B9 | Contribution at a new return rate in B10 | =B1*(1+B5)*(B3-B10*(B3+B4)) |
To compare scenarios, change B5 and B10. A negative result in B9 minus B6 means the lift is more than offset by returns.
A maturity self-check (editorial framework)
The five stages below are my own framing, not a published standard, and the thresholds are judgment calls. Their use is as a diagnostic.
| Stage | What it looks like |
|---|---|
| Rules-based | Segments by demographics or geography |
| Behavioral | ML recommendations from browsing and purchase history |
| Predictive | Anticipates the next need, not just the last purchase |
| Real-time individual | Layout, search and offers adapt per session |
| Governed | Consent, explanation and review built in at every stage |
Five questions to place yourself:
- Can website, app, in-store and support history merge into one customer record quickly?
- Does the recommendation model’s training target include anything beyond clicks or purchases, such as returns or net margin?
- Can a customer see in plain language why an item was recommended?
- If prices ever vary by individual behavior, does anyone review that for fairness and legal risk?
- Could an AI shopping agent read your catalog and return policy without a human interpreting it?
Three or more “no” answers suggest you are earlier on the curve than your vendor dashboard implies.
Pricing: personalization versus surveillance pricing in 2026
This is the part of the original article that aged fastest. In 2025, the advice was to get legal review because there was no bright line. In 2026, state legislatures drew several lines. This is a summary for orientation, not legal advice, and the details vary by state, product category, and enforcement.
What changed in 2026
- New York (disclosure). The Algorithmic Pricing Disclosure Act took effect in November 2025. It requires businesses that use personalized algorithmic pricing to display a label reading “This price was set by an algorithm using your personal data,” with civil penalties of up to $1,000 per violation. The National Retail Federation’s First Amendment challenge failed at an early stage.
- New York (ban, pending). In June 2026, the legislature passed the One Fair Price Act, which would bar algorithmic pricing that relies on personal data while preserving legitimate discounts, coupons, loyalty programs, and senior pricing. It applies to businesses that solicit New York consumers, including out-of-state and online sellers. As of the sources I found, it awaited the governor’s signature, and the effective date would be 180 days after signing. Check its current status before relying on it.
- Maryland. The Protection From Predatory Pricing Act, signed in April 2026, was described as the first U.S. law to prohibit surveillance pricing rather than just require disclosure. It targets certain food retailers and third-party delivery services. One 2026 guide reports an effective date of October 1, 2026, so confirm that against the statute.
- Connecticut. Lawmakers approved limits on surveillance pricing in May 2026, and later 2026 legal summaries list Connecticut among the states that have acted. Check the statute’s scope and obligations.
- New Jersey. On July 23, 2026, New Jersey became the third state to ban surveillance pricing for groceries, with penalties of up to $50,000 per violation, treble damages, and a private right of action.
- California. The attorney general opened an investigative sweep in early 2026 under the existing privacy law (CCPA), and a bill to ban the practice (AB 2564) was introduced. California’s separate AB 325 targets shared pricing algorithms used to coordinate prices among competitors.
- Colorado. The governor vetoed HB 1210, a broader bill on algorithmic individualized price- and wage-setting.
- Federal. The FTC ordered eight companies to explain how they use consumer data to set targeted prices in July 2024 (a study, with no allegation of wrongdoing), and staff published preliminary findings in January 2025. In March 2026, the House Oversight Committee opened an investigation into AI-driven surveillance pricing, and further congressional inquiries followed. Federal bills had been proposed but not enacted as of the sources I found.
- Volume. More than 40 bills on personalized algorithmic pricing were introduced in at least 24 states in 2026, according to a legislative tracker.
A practical distinction (mine, not a legal standard): pricing that changes with transparent, non-personal factors such as inventory level, shipping distance, or time of day carries far less exposure than pricing that charges more to people inferred to be willing to pay it. Coupons, loyalty discounts and promotions are generally treated differently from individualized price-setting, but the drafting varies, so confirm how each statute defines them.
What to do now:
- Inventory every place a price, fee, discount, or shipping cost changes by customer. Include vendor tools that do this for you.
- Separate cost-based or market-based dynamic pricing from pricing that uses personal data, and document the inputs.
- Check each state where you sell. The groceries-and-delivery scope in some laws does not cover general retail, but New York’s pending bill is broader.
- Get legal review before individualizing prices. I also could not verify the commonly quoted share of retailers using AI dynamic pricing, so it is left out.
Agentic commerce: why the forecasts differ by 35x
Forecasts of how much commerce AI agents will handle by 2030 look contradictory until you read the definitions.
| Forecast | Scope | Definition of “agentic” |
|---|---|---|
| Bain | $300B to $500B, U.S., 15% to 25% of e-commerce | Purchases initiated, influenced or completed by third-party or retailer-hosted agents, excluding journeys that only use AI-assisted search or discovery |
| Morgan Stanley | $190B to $385B, U.S., 10% to 20% | Narrower; requires autonomous action by the agent |
| McKinsey | Up to $1T U.S. B2C; $3T to $5T globally | Broad; includes everything an agent orchestrates |
| eMarketer | About $144B | Only checkouts completed inside an AI chat |
Morgan Stanley, McKinsey and eMarketer figures are as compiled by a third-party comparison. Bain’s are from Bain.
The spread comes from definitions, not from disagreement about direction. What every version shares is mechanical: an agent reads structured product data, checks the return policy, and compares prices in seconds. A catalog an agent cannot parse is a catalog it skips. Making product attributes, pricing and return terms machine-readable pays off under any of these forecasts.
Trust, consent and what goes wrong
A January 2026 study in Sustainability surveyed 400 e-commerce customers in Turkey. Its abstract reports that privacy concerns reduce psychological well-being and partly offset the benefit of personalization, and that brand trust weakens the damaging effect of those concerns. It is a single-market survey, so read it as directional.
In Stord’s consumer survey, 30% of respondents say they would never let AI handle shopping or access their payment details, 16% are very comfortable with it, and 21% are open to it if they can review transactions first. That reads as a design problem more than an attitude to argue people out of.
Customers also treat a retailer’s AI as the retailer speaking. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline responsible after its support chatbot gave a customer wrong information about bereavement fares. A hallucinated discount is a promise the business may have to honor.
Common symptoms and fixes (editorial diagnostics, not research findings):
| Symptom | Likely cause | Fix |
|---|---|---|
| Rising opt-outs or “not relevant” flags | Consent treated as a one-time checkbox | Persistent, granular preference center |
| “Creepy” complaints | No visible reasoning for a recommendation | Plain-language “because you bought X” |
| Obviously wrong recommendations | Stale or poorly weighted signals | Recency decay; ingest return-reason data |
| One bad recommendation, outsized backlash | No quick correction path | “Not helpful” control with immediate effect |
A practical roadmap
- Define the objective before any vendor call. Use net margin per cohort over a window that includes your return period, not click-through rate.
- Build the data foundation before the features. Start with a unified customer profile across web, app, store, and support.
- Test a returns-aware training target on a holdout before rolling out any new recommender.
- Start with the lowest-regret use case, such as search relevance and basic recommendations, measured on net margin.
- Ship consent and explanation with personalization, not after an incident.
- Keep prices out of the personalization scope until legal has mapped the states you sell in (see the pricing section).
- Make the catalog and return policy machine-readable, since that helps under every agentic-commerce forecast.
On company size, McKinsey’s data shows large organizations scaling AI faster than small ones, which supports a simple rule: the smaller you are, the more you should buy proven, narrow tools and the less you should build a custom data platform. That is a judgment, not a benchmark. I would also distrust any fixed budget split between data, models, and consent tooling, since I found no credible published benchmark for one.
FAQ
What is AI-personalized shopping?
Machine-learning systems that tailor search results, recommendations, offers, and sometimes prices to each shopper’s behavior and context. It also covers shoppers using AI assistants to research and compare products.
How many retailers use AI?
It depends on the question. McKinsey’s cross-industry survey reports regular use by nearly nine in ten organizations, NVIDIA’s retail and CPG survey reports 58% actively deploying, and 89% of NVIDIA’s respondents say AI raised revenue. These are different measures.
How many shoppers use AI to shop?
Stord’s survey reports 51% have used AI for online shopping and 17% use it regularly. Another survey found 33% of Gen Z prefer AI platforms for product research. Definitions differ, so compare the question, not just the percentage.
Does AI personalization increase revenue?
Controlled evidence exists but is narrow. In one live test, optimizing a recommender for purchases instead of clicks produced a much larger GMV uplift. The widely quoted “up to 40%” figure traces to a vendor blog, not an experiment.
Does AI personalization increase returns?
Not established. Fit and expectation mismatch are known drivers of returns, and fit-aware models reduce them, but I found no controlled study showing click-optimized recommenders increase returns. Test it on a holdout.
How much can returns rise before a revenue lift disappears?
It depends on the lift, margin, and return-handling cost. In the model above, a 5% lift tolerates about 2.1 points of return-rate increase and a 20% lift tolerates about 7.5 points. Use the spreadsheet formulas to run your own numbers.
Is AI dynamic pricing legal?
It depends on the state, the product, and how prices are set. As of 2026, New York requires disclosure of personalized algorithmic pricing, and Maryland, Connecticut, and New Jersey have enacted laws limiting or banning surveillance pricing, mostly in grocery and food-delivery contexts. New York’s broader ban was awaiting signature at the last check. Get legal review before individualizing prices.
Should retailers prepare for agentic commerce now?
Preparing the catalog (structured attributes, clear return policy, accurate pricing) is low-regret under every forecast, because even the narrowest forecast assumes agents reading product data.
Glossary
- Customer data platform (CDP): unifies customer data into one persistent profile that other systems can read and write.
- Objective function: the target a model is trained to maximize, such as clicks, purchases, or net margin after returns.
- GMV: gross merchandise value, the total value of goods sold before returns and costs.
- Agentic commerce: purchases initiated, influenced, or completed by an AI agent acting for a shopper.
- Surveillance pricing: setting individualized prices using personal data about a customer’s characteristics or behavior.
- Break-even return rate: the return rate at which a sales lift produces the same contribution as the baseline.
What I could not verify
- Stord’s consumer-survey sample size, dates and method, and the single-source generational splits.
- The McKinsey 2025 “7% fully scaled” figure on McKinsey’s own page. I confirmed the 2026 edition and relied on Stord’s attribution for the 2025 numbers.
- Morgan Stanley, McKinsey, and eMarketer forecast details, which came through secondary compilations.
- Moffatt v. Air Canada details beyond the widely reported outcome. Check the tribunal decision before quoting it.
- The 2026 state-law details (effective dates, scope, penalties), which I read in law-firm and trade summaries rather than the statutes themselves, and whether New York’s One Fair Price Act has been signed.
- Any controlled study linking click-optimized recommenders to higher return rates.
- The model’s margin and return-handling illustrative cost inputs.
Sources
Adoption and surveys
- Stord. (Feb. 11, 2026). State of AI in E-Commerce 2026 and press release.
- McKinsey & Company. (Aug. 25, 2026). The state of AI in 2026: On the road to ROI.
- NVIDIA. (2026). State of AI in Retail and CPG.
- Commerce and Future Commerce. (2025). New Modes report, via CXM.
Traffic and holiday data
- Adobe Digital Insights. (2025). AI-sourced traffic insights. PDF; summary via Marketing Dive.
- Adobe Digital Insights. (Sept. 2026). 2026 U.S. Holiday Forecast. business.adobe.com/resources/holiday-shopping-report. Coverage: Digital Commerce 360, Sept. 29, 2026.
Recommenders and returns
- Weiss, M., Rosenbach, R., & Eggenberger, C. (2025). Clicks Versus Conversion: Choosing a Recommender’s Training Objective in E-Commerce. arXiv:2508.10377.
- National Retail Federation & Happy Returns. (Oct. 15, 2025). 2025 Retail Returns Landscape. nrf.com.
- SizeFlags: Reducing Size and Fit Related Returns in Fashion E-Commerce. ACM KDD 2021. dl.acm.org.
- MIT Sloan. How better predictive models could lead to fewer clothing returns. mitsloan.mit.edu.
- From purchase to return: How personalized E-commerce recommendations shape consumer behavior. ScienceDirect.
- Fits like a glove? Knowledge and use of size finders and high-end fashion retail returns. ScienceDirect.
- Brand Trust in AI-Driven E-Commerce Personalization: The Well-Being–Privacy Trade-Off. Sustainability, 18(2), 2026. Abstract.
Pricing and regulation
- Federal Trade Commission. (July 2024). FTC issues orders to eight companies seeking information on surveillance pricing. ftc.gov.
- Reuters. (July 22, 2025). Delta plans to use AI in ticket pricing, draws fire from U.S. lawmakers. reuters.com.
- King & Spalding. (Mar. 16, 2026). Federal and State Lawmakers Take Aim at Surveillance Pricing Practices. kslaw.com.
- Kelley Drye. (Apr. 15, 2026). “Surveillance Pricing”: Key Concepts, the Current Legal and Legislative Landscape, and Mounting Scrutiny. kelleydrye.com.
- Faegre Drinker. (May 5, 2026). Surveillance Pricing: The Next Frontier of Privacy Litigation. faegredrinker.com.
- Loeb & Loeb. (June 2026). One Fair Price Act: New York’s Developing Approach to Surveillance Pricing. loeb.com.
- Troutman Pepper Locke. (June 2026). New York passes One Fair Price Act. regulatoryoversight.com.
- Consumer Reports. (June 2026). New York legislature passes key bill to curb personalized pricing. advocacy.consumerreports.org.
- LexisNexis. (June 24, 2026). Legislators Ramp Up Efforts to Regulate Surveillance Pricing. lexisnexis.com.
- Mondaq. (Aug. 6, 2026). Surveillance Pricing and Dynamic Pricing: What General Counsels Need to Know. mondaq.com.
Agentic commerce
- Bain & Company. (Dec. 2025). 2030 Forecast: How Agentic AI Will Reshape U.S. Retail. bain.com.
Method and update log
Method. I traced each statistic to the earliest source I could reach, compared definitions and populations before comparing numbers, and recomputed every figure in the break-even model. Items that rest on secondary sources are labeled in the text and listed under “What I could not verify.” The maturity framework and symptom table are editorial and are labeled as such. No vendor named here paid for placement.
Update log
- Oct. 21, 2026: Rewrote the pricing section around 2026 state laws and federal inquiries. Added Adobe’s 2026 holiday forecast. Expanded the break-even model with more lift scenarios, a closed-form formula, a spreadsheet version, and an unsellable-returns stress test. Corrected the $10,000 platform-cost break-even from 2.3% to 2.4% (the exact figure is 2.35%). Fixed the count of sources behind the high-80s adoption figures from four to three. Clarified that the model assumes returned goods are restocked.
- Oct. 18, 2026: Previous version. Figures last checked.
Where a number drives a real business decision, check the linked original before relying on it.
Tom Morgan covers retail AI and e-commerce infrastructure. No sponsorship: no vendor named here paid for placement.
