Decentralised News · Agentic Finance · Decision research
By Heath Muchena · Source review 2026-11-14 · Instrument version 1.0.0
What Matters
AI influence, referral traffic, shopping intentions and completed purchases are different measurements. A retailer deciding where to invest needs the event, population, period and method behind each number. The eight figures below support useful decisions, but none establishes a universal share of purchases completed autonomously by agents.
A board presentation can place a large AI-influenced sales estimate beside a fast-growing referral channel and make agent checkout look inevitable. The arithmetic may be correct while the inference is unsupported. One measure describes money associated with AI-assisted activity; another counts visits; a third asks consumers what they intend to do.
This guide reads eight published figures as decision evidence. It also supplies an original DN calculator for the two numerical mistakes that travel easily between articles: treating percentage growth as market share, and treating percentage-point change as percentage growth. It does not estimate the size of an autonomous-shopping market.
DN Evidence Block
Sources reviewed: 2026-11-14 · Evidence type: Primary publisher releases and survey methodology; original DN interpretation method and local arithmetic tool.
Scope: Selected 2025 holiday observations and 2026 consumer survey intentions. US and global populations are identified separately.
- The Salesforce and Adobe releases describe different events and measurement systems; their headline values are not interchangeable market-share estimates.
- PwC identifies its 2026 survey as 4,093 US respondents interviewed online from 15–26 June 2026. NRF describes its IBM research as covering 18,000 global consumers.
- The eight-number selection and claim-boundary method are DN editorial analysis. Calculator scenarios are labelled illustrations, not additional market observations.
Not established by this work: DN did not collect shopper transactions, audit the underlying vendor datasets, test checkout completion or measure causal sales lift. Survey intentions are not completed holiday outcomes.
Original DN instrument · version 1.0.0
DN Shopping Statistic Interpreter
Translate a growth percentage into an index, or compare two percentage shares when their definitions are comparable. The tool never infers autonomous-checkout share from referral, influence or intention data.
Enter your assumptions or load a labelled example. Required blank numbers stay missing; unknown evidence remains unknown. Use nonpersonal labels only. This instrument makes no network requests and uses no persistent browser storage.
Use your own inputs
No current result. Complete inputs or load an illustrative example.
Local arithmetic only. It does not validate sampling, link survey respondents to purchases, estimate causal lift, discover a missing denominator or verify autonomous execution. The growth example uses a synthetic index of 100; the observed baseline traffic count is not supplied.
1. Salesforce’s 20% describes attributed influence
Salesforce’s January 2026 holiday release associates 20% of global holiday retail sales with AI and agents, including recommendations and engagement. Its methodology uses aggregated shopper activity and extrapolation. That is broader than an agent receiving purchasing authority and completing checkout without further human action.
For a decision memo, preserve the source’s event definition. A discovery investment can be justified by evidence that AI participates in research. A checkout integration needs additional evidence about authorized orders, successful completion, customer support and the economics of that integration. The influence estimate alone does not supply those observations.
Record the claim in a sentence that survives being copied out of context: publisher, period, market, measurement and limitation. If the available methodology does not expose the exact attribution rule, mark that rule unresolved. Do not replace an unavailable rule with a plausible description of how you think the vendor counted it.
2. The $262 billion figure is not a second independent confirmation
The same Salesforce release reports $262 billion associated with AI and agents during November–December 2025. It is another expression of the reported influence estimate, not an independent study proving autonomous purchases or incremental revenue.
Separate three questions: how much activity was associated with AI, how much would have happened anyway, and how much was completed through delegated execution. A revenue total can answer the first under the publisher’s method without answering the other two.
When preparing a business case, give correlated figures one evidence-family ID. A percentage and the money derived from the same measurement should not become two votes in favour of a forecast. Keep the source’s rounding and broader market extrapolation visible rather than reverse-engineering an apparently precise denominator from rounded headline values.
3. Adobe’s 693.4% is referral-traffic growth
Adobe’s January 2026 release reports 693.4% year-on-year growth in shoppers clicking from generative AI tools to US retail sites during the 2025 holiday period. It explicitly notes a modest user base. This measures growth in referrals, not the share of retail sales or agent-completed orders.
A 693.4% increase means the new value is 7.934 times the old one. On an illustrative index beginning at 100, the new index is 793.4. It does not mean AI accounts for 693.4% of anything, and it does not reveal the channel’s share without the total traffic denominator.
For channel planning, ask for absolute visits, comparable conversion events, acquisition costs and the portion of the audience you can actually reach. A small channel can grow quickly and still be commercially useful; it does not need an inflated market-share story. The calculator below preserves the missing share as unknown.
4. NRF’s 41% is a reported research activity
NRF’s January 2026 summary of IBM consumer research says 41% use AI assistants to research products, from research covering 18,000 global consumers. Researching a product is an activity, not evidence that the same person delegated payment.
A respondent may research, read reviews and seek deals in the same journey. Do not add activity percentages unless the questionnaire establishes mutually exclusive groups. Similarly, a global survey should not silently become a South African, US or customer-specific adoption estimate.
This evidence can inform what information a product page should expose: specifications, compatibility, total cost and limitations. Deciding whether to automate purchasing requires a different observation. Keep the proposed intervention beside the measured activity so a research statistic is not doing the work of a transaction study.
5. PwC’s 29% versus 22% is a change in stated intentions
PwC’s 2026 Holiday Outlook says 29% plan to use AI somewhere in holiday shopping, compared with 22% in the prior-year survey. These are intentions. The difference is seven percentage points; the relative increase is approximately 31.8%.
Those two ways of describing change answer different questions. Percentage points describe the distance between shares. Relative change describes that distance as a fraction of the earlier share. Both require comparable definitions, even though the arithmetic can be performed on any two numbers.
Before treating a year-on-year difference as a trend, inspect questionnaire wording, sampling and fieldwork. A changed definition can create an apparent adoption increase. DN has not audited respondent-level comparability here; the calculator therefore asks the reader to declare whether the comparison is supported and withholds the comparison when it is unresolved.
6. PwC’s 64% points to an attribution gap
The PwC survey says 64% of consumers using AI for ideas open a separate website tab to buy instead of following the tool’s recommended link. This is a conditional, self-reported group, not 64% of all shoppers or a measured referral-loss rate.
The decision implication is a question to test: how much discovery occurs outside the referral trail your analytics can observe? A direct visit after an AI recommendation may not carry a useful referring URL. That possibility does not authorize relabelling all direct traffic as AI traffic.
Use a small measurement plan with consented customer feedback, tagged links where available and transaction records for the outcome. Report each signal separately. The result may remain an estimate with an unresolved portion; that is more useful than a complete-looking attribution model built from an unobserved assumption.
7. PwC’s 4% is an expected purchasing channel
PwC reports that 4% expect to purchase gifts through AI tools. It does not say 4% of realized holiday orders were executed autonomously. The survey was conducted before the holiday season, and a purchase through an AI interface may still involve human approval.
Do not divide this percentage by the broader AI-use percentage and call the result a checkout conversion rate. A conversion rate needs a defined eligible population and linked outcomes for that population. Separate survey questions, even within one study, do not automatically produce a transactional funnel.
A retailer can use the finding to design a bounded pilot. Define an authorized checkout attempt, a completed order, an abandonment and a refund before measuring results. Keep assisted and unattended execution separate. The pilot can then answer a question the survey was never designed to answer.
8. Salesforce’s nine-times comparison depends on the comparator
Salesforce reports AI-search referrals converting nine times more often than social-media referrals. This is a comparison between referred audiences. It is not a universal AI conversion rate or proof that moving the same shopper to AI causes a ninefold improvement.
A ratio needs both absolute rates and a matching event definition to become a budget input. Nine times a small base can still be a modest rate. Audience intent, product mix, device, attribution window and selection into each channel can also affect the comparison.
For an investment decision, request a matched analysis or run a properly designed experiment with a predeclared outcome. Until then, treat the ratio as a channel signal with a specific comparator. Keep the distinction between observed association and incremental effect in the sentence that a colleague or an AI assistant will quote.
Methodology, limitations and corrections
DN reviewed the linked primary releases and methodology statements on 14 November 2026. The selection deliberately contains different evidence types. It is not a meta-analysis, a comprehensive market estimate or a ranking of research providers.
| Claim field | What to retain | What fails without it |
|---|---|---|
| Event | Research, referral, authorized attempt or completed order | A discovery statistic becomes a checkout claim |
| Denominator | Visits, respondents, orders or revenue, with geography | Different populations are treated as one funnel |
| Period | Observation or fieldwork dates, plus publication date | A forecast becomes a reported outcome |
| Method | Observed, self-reported, attributed or modelled | Association becomes causation |
| Evidence family | Original dataset and dependent restatements | The same measurement is counted twice |
For growth g, the index is 100×(1+g/100). For comparable shares p and c, the percentage-point difference is c−p; relative change is 100×(c−p)/p when p is nonzero. Missing values stay missing. A source-defined percentage and its derived dollar amount are one evidence family. None of these arithmetic operations upgrades intentions into observations.
The supplied article list also proposed a general guide to verifying AI-generated numbers. Its source, date, denominator, method and reproducibility checks are included here rather than published as a second generic article. To extend this research, collect a consented, linked attempt-to-completion dataset with an explicit human-approval field and a dated correction log.
For corrections, identify the claim, source and date when contacting Decentralised News. A new record-creation timestamp does not refresh its underlying evidence.
Change log: version 1.0.0, 2026-11-14. Initial documentation review, original instrument and labelled examples.
Frequently asked questions
Does AI-influenced spending mean an agent paid?
No. Influence can include recommendations, research or engagement. Payment execution needs a separate definition and observation.
Is 693.4% growth the same as a 693.4% share?
No. It describes change from a baseline. A baseline index of 100 becomes 793.4; the total-market denominator is still missing.
Can 29% and 4% be divided into a checkout conversion rate?
Not without a linked, eligible population and measured outcomes. Separate survey responses do not automatically define a funnel.
Why are there two different expressions of share change?
Percentage points describe the distance between two shares. Relative growth divides that distance by the earlier share. The latter is undefined when the earlier share is zero.
Does the calculator certify that two sources are comparable?
No. Comparability is a user-declared input that needs support from the source methodology. The tool preserves that limit in its result.
Primary sources
- Salesforce: 2025 Holiday Shopping Data. 2025 holiday AI-influence estimate, associated sales value, referral comparison and vendor methodology. Reviewed 2026-11-14.
- Adobe: 2025 holiday shopping season results. Generative-AI referral-traffic growth, event definition, period and modest-base qualification. Reviewed 2026-11-14.
- NRF and IBM: Own the agentic commerce experience. Consumer-reported research activity and global study scope. Reviewed 2026-11-14.
- PwC: Holiday Outlook 2026. Holiday intentions, separate-tab behavior, anticipated purchasing channel and survey methodology. Reviewed 2026-11-14.
Turn the finding into a decision
A merchant or research team can ask the DN Research Desk to scope a claim register for one named investment decision. Bring the proposed claim, original sources and the budget choice it is meant to support. Access, feasibility and a written quote must be agreed before work begins; no sales-lift or market-share result is promised.
Related DN research
Disclosure: Analysis and instrument results are editorial. Where labelled affiliate links appear, DN may receive compensation when you sign up or buy through them. Commercial relationships do not determine findings. No tested integration or guaranteed outcome is implied.
