eg

AI ecommerce statistics 2026: traffic, conversion, and agentic commerce

AI-referred traffic to US retail sites grew 693% year over year during the 2025 holiday season, according to Adobe Analytics. It now converts above other channels in Adobe's US panel, though first-party academic data covering an earlier period disagrees. This page tracks both, with sources, samples and periods on every figure.

+693%
Year-over-year growth in AI-referred traffic to US retail sites, 2025 holiday season
Measures
AI-referred
Geography
US
Period
1 Nov to 31 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel
+31%
AI referrals converted 31% more than other traffic sources in Adobe's US panel, 2025 holiday season
Measures
AI-referred
Geography
US
Period
1 Nov to 31 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel, blended non-AI comparator
$67B
Global Cyber Week 2025 sales that Salesforce reports as influenced by AI and agents, 20% of all purchases
Measures
AI-influenced
Geography
Global
Period
25 Nov to 1 Dec 2025
Sample
1.5 billion+ shoppers
Method
Shopping Index, retailer-side AI
38%
of US consumers report having used generative AI for online shopping
Measures
Self-reported behavior
Geography
US
Period
Survey, 2025
Sample
n = 5,000
Method
Consumer survey
00 · Definitions

What each statistic measures

The figures on this page come from different measurement systems that are often quoted as if they were one. Every card, chart and caption below uses these terms, and no two of them are added together.

Measurement terms used on this page
TermMeansOn this page
AI-referredA human clicked a link from an AI surface to a retail site.Adobe's traffic growth figures. The Kaiser and Schulze organic LLM channel.
AI-influencedAI touched the purchase journey at some point, as measured by the vendor's own model.Salesforce's $67 billion and 20% of purchases.
AI-attributedRevenue assigned to AI by a last-touch or modeled attribution rule.Used on this page only where the source uses it. No current figure qualifies.
Agent-executedAn agent completed the action or the transaction itself.Salesforce's 70% growth in agent-driven actions such as address updates and returns.
AI crawler activityAutomated requests from bots, not human sessions.Never mixed with referral traffic on this page.
Retailer-owned assistant usageUse of a retailer's own assistant, such as Amazon Rufus or Walmart Sparky. A separate channel.No verified usage figure is currently on this page.

Consumer survey terms

Consumer survey terms used on this page
TermMeansOn this page
Stated preferenceWants an AI shopping assistant.Not summed with any other row.
IntentionPlans to use one.Adobe: 52% plan to use generative AI for shopping.
Self-reported behaviorSays they have used AI for shopping.Adobe: 38% have used generative AI for online shopping.
Agentic behaviorAllowed an agent to complete a purchase.No verified consumer figure is currently on this page.
01 · AI-referred traffic
Data through December 2025

AI-referred traffic growth

AI-referred traffic to US retail sites grew 693.4% year over year across November and December 2025, according to Adobe Analytics, after a 4,700% year-over-year rise in July 2025. Adobe measures this as shoppers clicking a link from a generative AI chat service or browser, so it is human referral traffic, not crawler activity.

+4,700%
Year-over-year growth in generative AI traffic to US retail sites, July 2025
Measures
AI-referred
Geography
US
Period
July 2025 vs July 2024
Sample
1 trillion+ visits
Method
Analytics panel
+693.4%
Year-over-year growth in AI-referred traffic to US retail sites, 2025 holiday season
November 2025 +769%, December 2025 +673%.
Measures
AI-referred
Geography
US
Period
1 Nov to 31 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel
+1,200%
Year-over-year growth in generative AI traffic to US retail sites, October 2025
Measures
AI-referred
Geography
US
Period
October 2025 vs October 2024
Sample
1 trillion+ visits
Method
Analytics panel
+1,300%
Year-over-year growth in generative AI traffic to US retail sites, 2024 holiday season
Cyber Monday 2024 alone was +1,950%.
Measures
AI-referred
Geography
US
Period
1 Nov to 31 Dec 2024
Sample
1 trillion+ visits
Method
Analytics panel
+539%
Year-over-year growth in AI-referred traffic to US travel sites, 2025 holiday season. Financial services +266%, tech and software +120%, media and entertainment +92%
Measures
AI-referred
Geography
US
Period
1 Nov to 31 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel
26%
of AI-referred traffic to US retail sites arrived on mobile in July 2025, up from 18% in January 2025
Measures
AI-referred
Geography
US
Period
July 2025
Sample
1 trillion+ visits
Method
Analytics panel
0%1,000%2,000%3,000%4,000%5,000%Nov to Dec 2024: +1,300% year over year1,300%Nov to Dec 2024Jul 2025: +4,700% year over year4,700%Jul 2025Oct 2025: +1,200% year over year1,200%Oct 2025Nov 2025: +769% year over year769%Nov 2025Dec 2025: +673% year over year673%Dec 2025
Source: Adobe Analytics, 1 trillion+ visits to US retail sites. Every bar is a year-over-year comparison. Adobe also published +1,100% for January 2025 and +3,100% for April 2025, but those two points are indexed to July 2024, not to the same month a year earlier, because the base was too small before then; they are omitted here so that every bar shares one baseline. Nov to Dec 2024 from the August 2025 release; July 2025 from the August 2025 release; October 2025 from the November 2025 release; November and December 2025 from the January 2026 release.

The growth rates fall as the base grows. Adobe's own sequence runs from +4,700% in July 2025 to +1,200% in October and +693% across the holiday season, which is what a channel compounding off a small base looks like once the prior-year comparison stops being near zero. Adobe does not publish AI-referred traffic as a share of total retail visits, and this page does not estimate one.

02 · Conversion
Data through December 2025

Does AI-referred traffic convert?

Two credible datasets answer differently. Adobe's US retail panel reports AI referrals converting 31% more than other traffic sources during the 2025 holiday season. A working paper by Kaiser and Schulze, using twelve months of first-party data from 973 websites, reports ChatGPT referrals converting below every traditional channel except paid social. Both are shown here with their definitions, and neither is preferred.

Dataset A · Adobe Analytics, US retail, holiday 2025

+31%
AI referrals converted 31% more than other traffic sources, nearly doubling year over year
Measures
AI-referred
Geography
US retail
Period
1 Nov to 31 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel, blended non-AI comparator
+16%
AI-referred visitors were 16% more likely to convert than non-AI sources in October 2025, the second consecutive positive month after +5% in September
Measures
AI-referred
Geography
US retail
Period
October 2025
Sample
1 trillion+ visits
Method
Analytics panel, blended non-AI comparator
−23%
AI-referred traffic was 23% less likely to convert than non-AI traffic in July 2025, improved from −49% in January and −38% in April
Measures
AI-referred
Geography
US retail
Period
July 2025
Sample
1 trillion+ visits
Method
Analytics panel, blended non-AI comparator
+254%
AI-driven revenue per visit, 2025 holiday season year to date, against −27% revenue per visit in July 2025
Measures
AI-referred
Geography
US retail
Period
Holiday 2025 YTD
Sample
1 trillion+ visits
Method
Analytics panel

Comparator: a blended non-AI baseline that Adobe defines as paid search, affiliates and partners, email, organic search, and social media. Period: 1 November to 31 December 2025. Panel: 1 trillion+ visits to US retail sites.

Dataset B · Kaiser and Schulze, 973 sites, 12 months

973
websites in the Kaiser and Schulze dataset, with $20 billion in combined revenue over twelve months
Measures
AI-referred
Geography
Not stated in abstract
Period
12 months to 2025
Sample
973 sites
Method
First-party analytics, regression
50,000+
ChatGPT referral transactions analyzed against 164 million transactions from traditional channels
Measures
AI-referred
Geography
Not stated in abstract
Period
12 months to 2025
Sample
973 sites
Method
First-party analytics, regression
Below all but paid social
Organic ChatGPT referrals show conversion rates and revenue per session above paid social but below every other traditional channel in the study
Measures
AI-referred
Geography
Not stated in abstract
Period
12 months to 2025
Sample
973 sites
Method
First-party analytics, regression
Lower
session duration and page views for ChatGPT referrals than traditional channels, with a more favorable bounce rate
Measures
AI-referred
Geography
Not stated in abstract
Period
12 months to 2025
Sample
973 sites
Method
First-party analytics, regression

Comparator: each traditional channel individually. Period: twelve months of first-party data, in a paper dated 8 October 2025. Dataset: 973 websites, $20 billion combined revenue, 50,000+ ChatGPT referral transactions against 164 million from traditional channels. SSRN working paper, not published in a journal. Geography of the sites is not stated in the abstract.

-60%-40%-20%0%+20%+40%parity with non-AI trafficJan 2025: -49% vs non-AI traffic-49%Jan 2025Apr 2025: -38% vs non-AI traffic-38%Apr 2025Jul 2025: -23% vs non-AI traffic-23%Jul 2025Aug 2025: -9% vs non-AI traffic-9%Aug 2025Sep 2025: +5% vs non-AI traffic+5%Sep 2025Oct 2025: +16% vs non-AI traffic+16%Oct 2025Holiday 2025: +31% vs non-AI traffic+31%Holiday 2025
Source: Adobe Analytics, US retail, 1 trillion+ visits. Percentage difference in conversion rate for AI-referred visits against Adobe's blended non-AI baseline (paid search, affiliates and partners, email, organic search, social media). January, April and July 2025 from the August 2025 release; August, September and October 2025 from the November 2025 release; November to December 2025 from the January 2026 release. September 2025 was the first month Adobe reported AI traffic converting above non-AI.

Channel conversion ranking, Kaiser and Schulze working paper

  1. 01
    Every other traditional channel in the study
    Higher conversion rate and revenue per session than organic ChatGPT referrals. The abstract does not name the individual channels or their values.
  2. 02
    Organic ChatGPT referral (oLLM)
    Stronger in complex product categories. Conversion rising over the twelve months, average order value declining.
  3. 03
    Paid social
    The only channel below organic ChatGPT referrals on conversion rate and revenue per session.

Ordering as reported in the abstract of SSRN 5585812, posted 18 October 2025, last revised 27 July 2026. Working paper, not published in a journal. Numeric values are omitted pending confirmation from the full text; an estimated paid search value that an earlier version of this page placed in a chart has been removed. Sample: 973 websites, 50,000+ ChatGPT referral transactions, 164 million traditional-channel transactions, twelve months. Kaiser and Schulze, SSRN working paper 5585812, Posted 18 October 2025, last revised 27 July 2026

Engagement of AI-referred visitors, both datasets

Engagement of AI-referred visitors compared with the comparator in each dataset
MetricAdobe, July 2025Adobe, holiday 2025Kaiser and Schulze, 12 months
Time on site+32% longer+45% longerLower session duration
Pages per visit+10% more+13% moreFewer page views
Bounce rate−27% lower−33% less likely to bounceMore favorable

Adobe columns: AI-referred visits against Adobe's blended non-AI baseline, US retail, 1 trillion+ visits; July 2025 from the August 2025 release, holiday 2025 (1 Nov to 31 Dec) from the January 2026 release. Kaiser and Schulze column: organic ChatGPT referrals against traditional channels across 973 sites; the abstract gives direction only, so no values are shown. The datasets agree that AI referrals bounce less and disagree on whether they stay longer. Adobe Analytics, 21 August 2025 · Adobe Analytics, 12 January 2026 · Kaiser and Schulze, SSRN working paper 5585812, Posted 18 October 2025, last revised 27 July 2026

Why the two results can both be right

  1. 01 · Different comparators

    Adobe measures AI referrals against a blended baseline that contains paid social. Kaiser and Schulze compare ChatGPT referrals with each channel individually. Beating a blend that includes paid social is not the same test as beating organic search, and the two cannot be read as one.

  2. 02 · Different periods

    Adobe's own earlier series agrees with the paper. AI traffic in the Adobe panel was 49% less likely to convert in January 2025, 38% less in April and 23% less in July, then crossed to +5% in September, +16% in October and +31% across the holiday season. The paper's twelve-month window largely predates the crossover.

  3. 03 · Different data types

    A vendor analytics panel weighted toward large US enterprise retailers, against first-party transaction logs from 973 sites of mixed size.

  4. 04 · Different populations and definitions

    Adobe is US retail only. The geography of the academic sample is not stated in the abstract and is pending confirmation from the full text; if it is European or mixed, the comparison with a US panel changes materially. Adobe counts AI referrals broadly across chat services and AI browsers, while the paper isolates organic ChatGPT referrals.

  5. 05 · Selection effects

    Peak-season AI referrals arrive with gift intent and a deadline. Off-peak AI referrals skew toward research in complex categories, which is where the paper finds LLM outcomes strongest. The two datasets may be sampling different moments in the same funnel.

No source has yet measured both populations on the same definition over the same window. The direction of travel in the US panel data is upward and crossed zero in September 2025. Whether that holds outside peak season and outside the United States is unresolved, and this page will not resolve it by averaging the two.

03 · Consumers
Data through December 2025

Consumer AI shopping: behavior and intention

38% of US consumers report having used generative AI for online shopping and a further 52% say they plan to, in Adobe's 2025 survey of 5,000 US consumers. The first number is self-reported behavior and the second is intention. They are shown separately below and never added together.

0%25%50%75%100%Have used generative AI for online shoppingSelf-reported behaviorHave used generative AI for online shopping: 38% (Self-reported behavior)38%Plan to use it this yearIntentionPlan to use it this year: 52% (Intention)52%
Source: Adobe consumer survey, n = 5,000 US consumers, published 21 August 2025. Behavior (have used) and intention (plan to) are separate responses to separate questions and are not additive. No stated-preference figure is charted, because the only preference figures retrievable from a primary source on this page come from surveys with different samples.

Adobe survey, n = 5,000 US consumers, published August 2025

38%
of US consumers report having used generative AI for online shopping
Measures
Self-reported behavior
Geography
US
Period
Survey, 2025
Sample
n = 5,000
Method
Consumer survey
52%
plan to use generative AI for online shopping this year
Measures
Intention
Geography
US
Period
Survey, 2025
Sample
n = 5,000
Method
Consumer survey
85%
of those who have used AI for shopping say it improved their shopping experience
Measures
Self-reported behavior
Geography
US
Period
Survey, 2025
Sample
n = 5,000
Method
Consumer survey
73%
cite AI as their primary source of product research
Measures
Self-reported behavior
Geography
US
Period
Survey, 2025
Sample
n = 5,000
Method
Consumer survey
83%
say they are more likely to use AI for larger or more complex purchases
Measures
Stated preference
Geography
US
Period
Survey, 2025
Sample
n = 5,000
Method
Consumer survey

Adobe Holiday 2025 Consumer Survey, n = 1,000+ US respondents, published January 2026

81%
of consumers using AI assistants for online shopping report an improved shopping experience
Measures
Self-reported behavior
Geography
US
Period
Holiday 2025 survey
Sample
n = 1,000+
Method
Consumer survey
47%
report trust in AI-generated results
Measures
Stated preference
Geography
US
Period
Holiday 2025 survey
Sample
n = 1,000+
Method
Consumer survey
64% / 55%
of shoppers using AI are satisfied with the links they receive; 55% say they actively click on them
Measures
Self-reported behavior
Geography
US
Period
Holiday 2025 survey
Sample
n = 1,000+
Method
Consumer survey
65% / 68%
more confident in their purchase after AI help; 68% less likely to return the product
Measures
Self-reported behavior
Geography
US
Period
Holiday 2025 survey
Sample
n = 1,000+
Method
Consumer survey

The two Adobe surveys have different sample sizes and fielding dates. The 85% improved-experience figure (n = 5,000) and the 81% figure (n = 1,000+) are not a trend line.

Dated baseline: Capgemini Research Institute, fielded October to November 2024

In Capgemini's survey of 12,000 consumers aged 18 and over in 12 countries across North America, Europe and Asia-Pacific, 58% said they had replaced traditional search engines with generative AI tools for product and service recommendations, up from 25% in 2023, and 71% wanted generative AI integrated into their shopping experiences. These are self-reported behavior and stated preference respectively, fielded in late 2024, and multi-country, so they are not comparable with Adobe's US-only figures. They are kept here as a dated baseline, not a current reading. Capgemini Research Institute, 9 January 2025

04 · AI-influenced sales, retailer-side AI and agents
Data through December 2025

AI-influenced sales: Salesforce Shopping Index

AI and agents drove $67 billion in global Cyber Week 2025 sales, influencing 20% of all purchases through personalized product recommendations and conversational customer service, according to Salesforce. This is retailer-owned AI measured on Salesforce-powered storefronts. It is not LLM referral traffic and it is kept apart from the Adobe figures above for that reason.

Read this group on its own terms. Salesforce's wording is "AI and agents drove $67 billion in sales, influencing 20% of all purchases through personalized product recommendations and conversational customer service." A purchase counts if a recommendation engine or a service agent touched the journey. None of these figures measure a shopper clicking from ChatGPT, Gemini or Perplexity to a store.
$67B / 20%
AI and agents drove $67 billion in global Cyber Week sales, "influencing 20% of all purchases through personalized product recommendations and conversational customer service"
Measures
AI-influenced
Geography
Global
Period
25 Nov to 1 Dec 2025
Sample
1.5 billion+ shoppers
Method
Shopping Index, retailer-side AI
$336.6B
global online sales during Cyber Week 2025, up 7% year over year. US: $79.6 billion, up 5%
Measures
Company disclosure
Geography
Global and US
Period
25 Nov to 1 Dec 2025
Sample
1.5 billion+ shoppers
Method
Shopping Index
32%
faster sales growth for retailers using Agentforce 360 than peers without dedicated AI agents. Salesforce names Pandora, Shark Ninja and Funko
Compares Salesforce's own customer base with non-users. Not a random sample of retailers.
Measures
AI-influenced
Geography
Global
Period
25 Nov to 1 Dec 2025
Sample
Salesforce customers vs non-users
Method
Vendor customer comparison
+55%
week-over-week growth in AI-guided customer service conversations
Measures
Agent-executed
Geography
Global
Period
25 Nov to 1 Dec 2025
Sample
Agentforce customers
Method
Shopping Index
+70%
growth in agent-driven actions such as address updates and return initiations
Measures
Agent-executed
Geography
Global
Period
25 Nov to 1 Dec 2025
Sample
Agentforce customers
Method
Shopping Index
61 million
orders powered on Agentforce Commerce during Cyber Week 2025
Measures
Company disclosure
Geography
Global
Period
25 Nov to 1 Dec 2025
Sample
Agentforce Commerce customers
Method
Shopping Index
$0B$20B$40B$60B$80BCyber Week 2023: $51 billion AI-influenced global sales$51B2023Cyber Week 2024: $60 billion AI-influenced global sales$60B2024Cyber Week 2025: $67 billion AI-influenced global sales$67B2025
Source: Salesforce Shopping Index, global Cyber Week, billions of US dollars. AI-influenced spend as Salesforce defines it: purchases influenced by personalized product recommendations and conversational customer service. Not LLM referral traffic. 2023 and 2024 from Salesforce releases distributed by Business Wire; 2025 from the Salesforce press release of 5 December 2025. Salesforce's measurement has expanded as its own agent products launched, so part of the year-to-year change reflects product coverage.

Series sources: Salesforce, via Business Wire, 28 November 2023 · Salesforce, via Business Wire, 4 December 2024 · Salesforce, 5 December 2025

Salesforce forecast $73 billion and 22% for Cyber Week 2025 before the event and reported $67 billion and 20% after it. The forecast is not a result and is cited here only so the two are not confused. Salesforce, 2025

05 · Peak season
Data through December 2025

Black Friday and Cyber Week 2025: the first AI-native peak season

AI referrals converted 38% better than non-AI traffic sources on Black Friday 2025 and 54% better on Thanksgiving, in Adobe's US retail panel. Cyber Monday AI traffic was up 670% year over year. These are AI-referred figures from Adobe. Salesforce's AI-influenced Cyber Week totals are in the section above and measure something different.

+38%
AI conversions vs non-AI traffic sources on Black Friday 2025
Measures
AI-referred
Geography
US retail
Period
Black Friday, 28 Nov 2025
Sample
1 trillion+ visits
Method
Analytics panel, blended non-AI comparator
+54%
AI conversions vs non-AI traffic sources on Thanksgiving 2025
Measures
AI-referred
Geography
US retail
Period
Thanksgiving, 27 Nov 2025
Sample
1 trillion+ visits
Method
Analytics panel, blended non-AI comparator
+670%
AI traffic to US retail sites on Cyber Monday 2025, measured by shoppers clicking a link
Measures
AI-referred
Geography
US retail
Period
Cyber Monday, 1 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel
+760%
AI traffic to US retail sites, season to date at Cyber Monday
Measures
AI-referred
Geography
US retail
Period
1 Nov to 1 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel
$257.8B
total US online holiday spend, 2025
Measures
Company disclosure
Geography
US
Period
1 Nov to 31 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel
5 categories
where AI chat services were used most on Cyber Monday 2025: video games, appliances, electronics, toys, personal care
Measures
AI-referred
Geography
US retail
Period
Cyber Monday, 1 Dec 2025
Sample
1 trillion+ visits
Method
Analytics panel

The peak-day conversion gaps are wider than the season average of +31%, which is consistent with the selection effect described in the conversion section: shoppers arriving from an AI service on Black Friday carry gift intent and a deadline. Adobe reports the season's total US online spend at $257.8 billion but does not publish AI-referred traffic as a share of it, and no share figure is estimated here.

06 · Protocols
Data through September 2026

Agentic commerce protocols: ACP, UCP, and MCP

OpenAI withdrew Instant Checkout as a native in-chat checkout on 24 March 2026 and refocused the Agentic Commerce Protocol on product discovery, with purchases completing on the merchant's own store. Google's Universal Commerce Protocol is an open specification whose checkout on Google surfaces is in early access for select merchants. MCP is a model-to-tool integration standard and defines no cart, checkout or payment primitives.

Where each protocol operates: ACP, UCP, AP2, A2A and MCP against discovery, cart, checkout, payment and post-purchase
Standard and layerSponsor or governanceDiscovery and catalogCartCheckoutPayment credentialsPost-purchaseStatus, September 2026
ACP
Commerce
Stripe and OpenAIYes. Product feeds and promotionsNoWithdrawn March 2026Shared Payment Token (launch design)Merchant-handledDiscovery layer live in ChatGPT. Native in-chat checkout withdrawn; purchases complete on the merchant's own store
UCP
Commerce
Google, co-developed with Shopify, Etsy, Wayfair, Target and WalmartYesYes, optional capabilityYesVia AP2 and regulated payment providersYesSpecification open. Checkout on Google surfaces in early access for select merchants, products eligible in the US, Canada and Australia
AP2
Payments
FIDO Alliance, donated by Google in April 2026NoNoNoYes. Signed intent, cart and payment mandatesNoSpecification open. Announced by Google 16 September 2025 with 60+ collaborating organizations; donated to the FIDO Alliance 28 April 2026. No transaction volume disclosed
A2A
Agent-to-agent
Linux Foundation, donated by Googlen/an/an/an/an/aVersion 1.0 released as the first stable specification. 150+ supporting organizations as of 9 April 2026. Not a commerce protocol
MCP
Integration
Agentic AI Foundation, under the Linux FoundationTool and data access onlyNot in scopeNot in scopeNot in scopeTool access onlyStable. 10,000 active servers and 97 million+ monthly SDK downloads at donation, December 2025

Sources: ACP from OpenAI (24 March 2026) and Stripe (29 September 2025). UCP from Google's developer announcement (11 January 2026), its capability update (19 March 2026) and Merchant Center Help (retrieved 8 September 2026). AP2 from Google Cloud (16 September 2025) and Google (28 April 2026). A2A from the Linux Foundation (9 April 2026). MCP from Anthropic, the MCP project and the Linux Foundation (9 December 2025). MCP is included to show what it does not cover.

3× lower
conversion rate for selection sold directly inside ChatGPT than for products requiring a click out to the merchant, per Walmart EVP Daniel Danker
Grade C. Etsy told the same outlet it did not see large volume from Instant Checkout but does see valuable referral traffic.
Measures
Company disclosure
Geography
US
Period
Reported March 2026
Sample
Walmart, one retailer
Method
Executive disclosure via trade press
7
retailers OpenAI names as integrated into ACP for discovery: Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, Wayfair
Measures
Protocol status
Geography
US
Period
24 March 2026
Sample
Named by OpenAI
Method
Company announcement
10,000
active MCP servers and 97 million+ monthly SDK downloads at the time MCP was donated to the Agentic AI Foundation
Integration activity, not commerce transactions. MCP defines no cart, checkout or payment primitives.
Measures
Protocol status
Geography
Global
Period
9 December 2025
Sample
Self-reported
Method
Company announcement
Select merchants
Google Merchant Center: "The checkout feature enabled by UCP is available for select merchants at this time." Eligibility limited to products in the United States, Canada and Australia
Measures
Protocol status
Geography
US, Canada, Australia
Period
Retrieved 8 September 2026
Sample
Not applicable
Method
Product documentation

What happened to in-chat checkout

Stripe and OpenAI launched Instant Checkout in ChatGPT on 29 September 2025, US only, with US-based Etsy businesses at launch and Shopify merchants described as coming soon. The launch introduced the Shared Payment Token, scoped to a specific merchant and cart total. Stripe, 29 September 2025 On 24 March 2026 OpenAI wrote: "We've found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we're allowing merchants to use their own checkout experiences while we focus our efforts on product discovery." OpenAI, 24 March 2026

The best available evidence on why comes from Walmart. Its EVP Daniel Danker told Wired, as reported by Modern Retail, that conversion rates were three times lower for selection sold directly inside the chatbot than for products requiring a click out to the merchant. Etsy told the same outlet it did not see large volume from Instant Checkout but does see valuable referral traffic. This is a named-executive disclosure from one retailer, graded C, and it is the only conversion comparison of in-chat checkout against redirect that has been made public. Modern Retail, 27 March 2026

Reported counts of merchants that went live on Instant Checkout vary across outlets and none has been confirmed by OpenAI or Shopify. [DATA NEEDED: confirmed live-merchant count]

What ACP is now

OpenAI describes ACP as "the connective layer between merchants and users throughout discovery," through which merchants share product feeds and promotions so their catalogs are represented in ChatGPT, with delivery paths that include third-party providers such as Salesforce and Stripe. OpenAI names Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot and Wayfair as integrated for discovery. Shopify merchants' product data is integrated through Shopify Catalog with no additional merchant work, and Shopify's Mani Fazeli is quoted saying buyers "can complete purchases on the merchant's online store through an in-app browser." Walmart has an in-ChatGPT app with account linking, loyalty and Walmart payments. The rollout covers free, Go, Plus and Pro users. OpenAI, 24 March 2026 The agentic commerce explainer covers the model in more depth.

What UCP is, and what is actually available

Google announced UCP at NRF on 11 January 2026 as an open-source standard co-developed with Shopify, Etsy, Wayfair, Target and Walmart, endorsed by more than 20 partners including Adyen, American Express, Best Buy, Flipkart, Macy's Inc, Mastercard, Stripe, The Home Depot, Visa and Zalando, and compatible with the Agent Payments Protocol, Agent2Agent and the Model Context Protocol as transports. Google Developers Blog, 11 January 2026 On 19 March 2026 Google added optional Cart and Catalog capabilities, the latter giving agents real-time pricing and inventory from retailer catalogs, and highlighted Identity Linking for loyalty benefits. Google, 19 March 2026

The specification and the checkout feature are different things. The specification is published and open. Google Merchant Center Help states that "the checkout feature enabled by UCP is available for select merchants at this time," that eligibility is limited to products in the United States, Canada and Australia, and that merchants express interest through a form. Retailers remain merchant of record. Google Merchant Center Help, Retrieved 8 September 2026

Where MCP fits

On 9 December 2025 Anthropic donated the Model Context Protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. MCP joined goose and AGENTS.md as founding projects, and its maintainers retain autonomy over technical direction. At donation Anthropic reported 97 million+ monthly SDK downloads and 10,000 active servers. Anthropic, 9 December 2025 MCP is not a commerce protocol. It defines how a model reaches tools and data, which is why UCP lists it as one transport among several, and it does not define carts, checkout or payment credentials. The MCP server directory lists the ecommerce servers that exist on it.

The arc across both sponsors is the same: a native in-chat checkout was tried, it underperformed on the one public comparison, and both OpenAI and Google have converged on merchant-owned checkout with a protocol-mediated discovery and catalog layer in front of it.

07 · Limitations
Data through September 2026

The measurement blind spot

Every AI traffic figure on this page comes from a tag-based analytics tool or a first-party analytics dataset, and those tools are built to exclude automated requests. That design choice, together with attribution rules and platform log access, shapes what the numbers can and cannot show.

  1. 01 · Tag-based analytics reports human sessions by design

    Google Analytics 4 excludes traffic from known bots and spiders automatically, using Google research and the Interactive Advertising Bureau's International Spiders and Bots List, and the exclusion cannot be disabled or its volume seen. Any figure from a tag-based tool is therefore a human-session figure by construction. The requests that AI systems make to fetch product pages on a shopper's behalf, or to crawl a catalog, live only in server logs. Google Analytics Help, Retrieved 8 September 2026

  2. 02 · User-agent strings are claims, not identity

    Google publishes a verification procedure because, in its words, site owners may be "concerned that spammers or other troublemakers are accessing your site while claiming to be from Google." The procedure is a reverse DNS lookup on the requesting IP, or a match against Google's published IP range files. Google Search Central, Retrieved 8 September 2026 OpenAI publishes separate IP address lists for OAI-SearchBot, GPTBot and ChatGPT-User for the same reason. OpenAI, Retrieved 8 September 2026 Allowlists and blocklists built on user-agent strings alone work from unverified identity data.

  3. 03 · Attribution rules credit the last recorded touch

    Google Analytics 4 attribution models assign credit to touchpoints by rule or by a data-driven algorithm, and every model excludes direct visits from receiving credit unless the whole path is direct. Google Analytics Help, Retrieved 8 September 2026 A shopper who researches in an LLM without clicking through, then buys via a branded search or by typing the URL, leaves no AI touchpoint to credit. The purchase is attributed to search, or to whatever preceded the direct visit.

  4. 04 · Some platforms expose no server logs at all

    The only unambiguous record of automated requests is the server or edge log, and not every hosted platform gives merchants one. Our server log monitoring guide documents what each major platform exposes; Shopify's online store, for example, offers no web server or edge log access, so merchants there cannot see crawler activity directly.

Ecommerce Guide reading: taken together, these mean most figures on this page probably understate AI influence rather than overstate it, because the research step that happens inside an assistant and never produces a click is invisible to every dataset cited here. The only measurement that captures automated activity with verified identity is server-side, matched against published IP ranges and confirmed with reverse DNS, which are the two methods Google documents.

08 · Implications
Data through September 2026

What these statistics mean for ecommerce teams

Split by how much evidence sits behind each practice. The first group is measurement and data hygiene that the sources on this page directly support. The second group is work whose effect nobody has yet measured in public.

Established practice

  • Analytics and attribution setup

    Know which attribution model your reporting uses and what it does with direct visits, so an AI-assisted purchase that returns as direct or branded search is at least a known gap rather than an invisible one.

  • Identify AI referrals in reporting

    Adobe measures AI referrals as shoppers clicking a link from a generative AI chat service or browser. The same referrer-based definition can be applied in any analytics tool, and it is the only definition under which the Adobe and Kaiser and Schulze figures are comparable with your own.

  • Server-log and crawler monitoring

    Automated requests from AI systems do not appear in tag-based analytics. Server-side logging with verification against published IP ranges and reverse DNS is the established way to see them.

  • Product feed completeness and accuracy

    OpenAI describes ACP as the layer through which merchants share product feeds and promotions so their catalogs are represented in ChatGPT, and Shopify Catalog does this for Shopify merchants with no additional work. The product feed management tools directory covers the software category.

  • Product and Offer structured data, catalog APIs and Merchant Center feeds

    UCP's catalog capability is designed to give agents real-time variants, inventory and pricing, and Google's UCP onboarding runs through Merchant Center product data.

  • Inventory and price accuracy

    Google's UCP documentation makes the merchant of record responsible for real-time inventory checks and requires an out-of-stock error during checkout when an item is unavailable, which halts that transaction. Google for Developers, Retrieved 8 September 2026 The UCP catalog specification states that catalog responses are not transactional commitments and that checkout is authoritative, so a catalog that disagrees with checkout fails at the point of sale. Universal Commerce Protocol project, Retrieved 8 September 2026

Experimental

  • Content and citation visibility tactics

    No source on this page demonstrates that any content change causes higher visibility inside an assistant. Tools that monitor brand mentions across assistants exist and are listed in the AI visibility tools directory; what they measure is presence, not cause.

  • Whether to test ACP, UCP or MCP integrations

    As of September 2026: ACP discovery integration has a low-cost path for Shopify merchants through Shopify Catalog with no per-merchant work, and named large retailers have integrated directly. UCP checkout is early access for select merchants with product eligibility limited to the US, Canada and Australia. MCP is not a commerce decision; it is an integration standard that UCP can use as a transport. Anything stronger than that is not supported by the sources on this page.

  • Consent and delegated purchasing design

    AP2 models both human-present and human-not-present purchases through signed intent and cart mandates. How shoppers will actually delegate, and at what limits, has no behavioral data behind it yet.

  • Fraud and identity controls for agent traffic

    The verification methods above establish whether a request is from a named crawler. They do not yet establish which human, if any, an agent acts for, and no source here reports fraud rates for agent-initiated orders.

09 · At a glance
Data through September 2026

Every headline figure with its measurement, sample and grade

Only figures retrieved from a primary document appear here. The grade describes the type of evidence, not the accuracy of the number; an A-grade panel can still be unrepresentative of the market.

AI ecommerce statistics at a glance, with measurement, geography, period, sample, source and evidence grade
StatisticGradeWhat it measuresGeographyPeriodSample or datasetSource
+693% AI-referred retail traffic, year over yearAAI-referred: humans clicking from an AI surface to a US retail siteUS1 Nov to 31 Dec 20251 trillion+ visits, analytics panelAdobe Analytics, 12 January 2026
+4,700% generative AI retail traffic, year over yearAAI-referredUSJuly 20251 trillion+ visits, analytics panelAdobe Analytics, 21 August 2025
+31% conversion for AI referrals vs other traffic sourcesAAI-referred, against Adobe's blended non-AI baselineUS1 Nov to 31 Dec 20251 trillion+ visits, analytics panelAdobe Analytics, 12 January 2026
+38% AI conversions vs non-AI on Black Friday; +54% on ThanksgivingAAI-referred, against Adobe's blended non-AI baselineUS27 and 28 Nov 20251 trillion+ visits, analytics panelAdobe Analytics, 12 January 2026
AI referrals 23% less likely to convert than non-AIAAI-referred, against Adobe's blended non-AI baselineUSJuly 20251 trillion+ visits, analytics panelAdobe Analytics, 21 August 2025
AI referrals 16% more likely to convert than non-AIAAI-referred, against Adobe's blended non-AI baselineUSOctober 20251 trillion+ visits, analytics panelAdobe Analytics, 10 November 2025
38% have used generative AI for online shopping; 52% plan toBSelf-reported behavior; intentionUS2025n = 5,000, surveyAdobe Analytics, 21 August 2025
83% more likely to use AI for larger or more complex purchasesBStated preferenceUS2025n = 5,000, surveyAdobe Analytics, 21 August 2025
$67B, 20% of purchases, influenced by AI and agentsAAI-influenced: personalized recommendations and conversational serviceGlobalCyber Week, 25 Nov to 1 Dec 20251.5 billion+ shoppers, Shopping IndexSalesforce, 5 December 2025
AI-influenced Cyber Week sales: $51B (2023), $60B (2024), $67B (2025)AAI-influenced, Salesforce definitionGlobalCyber Week 2023 to 20251.5 billion+ shoppers, Shopping IndexSalesforce, 5 December 2025
32% faster sales growth for Agentforce 360 retailers vs peersCAI-influenced, vendor customers vs non-usersGlobalCyber Week 2025Salesforce customer baseSalesforce, 5 December 2025
+70% agent-driven actions; +55% AI-guided service conversationsAAgent-executedGlobalCyber Week 2025, week over weekAgentforce customersSalesforce, 5 December 2025
ChatGPT referrals convert above paid social, below all other traditional channelsAAI-referred, organic ChatGPT only, against each named channelNot stated in abstract12 months to 2025973 sites; 50,000+ ChatGPT transactions; 164M traditionalKaiser and Schulze, SSRN working paper 5585812, Posted 18 October 2025, last revised 27 July 2026
In-chat checkout converted 3× lower than click-out to the merchantCCompany disclosure, WalmartUSReported March 2026One retailerModern Retail, 27 March 2026
10,000 active MCP servers; 97M+ monthly SDK downloadsCProtocol adoption, integration activity onlyGlobal9 Dec 2025Self-reportedAnthropic, 9 December 2025
UCP checkout available to select merchants; US, Canada, AustraliaCProtocol statusUS, CA, AURetrieved 8 Sep 2026Not applicableGoogle Merchant Center Help, Retrieved 8 September 2026
10 · Methodology
Data through September 2026

How this page is researched and updated

Every statistic on this page links to the organization that produced the data, not to a report about it. Where a figure originates in a vendor's own release, earnings material, or a research paper, we link to that document. Where we have been unable to retrieve the underlying document, the figure has been removed rather than retained on trust. Each statistic carries its geography, the period measured, the sample or dataset size where disclosed, and the method used. Where two credible sources disagree, both are shown with their definitions, and we explain why they can differ rather than averaging them.

Published April 19, 2026. Last reviewed September 8, 2026, when every source below was opened and each figure checked against it. Any calculation performed by us is labeled "Ecommerce Guide calculation" or "Ecommerce Guide reading" in place. Maintained by the Ecommerce Guide editorial team; contact details are in the site footer.

Evidence grades

Grades describe the type of evidence, not the accuracy of the number. An A-grade panel can still be unrepresentative, and a C-grade company disclosure can be exactly right about the one company it describes.

A
Large first-party behavioral or transaction dataset: the Adobe Analytics panel, the Salesforce Shopping Index, the 973-site academic dataset.
B
Credible survey with disclosed methodology, sample and fielding period.
C
Company disclosure, earnings material, product documentation, or a limited vendor dataset (fewer than about 200 sites or brands).
D
Secondary report, unclear methodology, or a directional estimate or forecast.
Sources cited on this page (28), with method, sample and what each cannot tell you
  • Adobe: Cyber Monday hits recordAdobe, 2 December 2025Grade A
    Method: Adobe Analytics panel of over 1 trillion visits to US retail sites. AI traffic is measured by shoppers clicking a link from a generative AI chat service or browser. Geography: United States. Sample: 1 trillion+ visits.
    Limitations: Single-day and month-to-date figures. Adobe notes the base of AI users remains modest.
  • Method: Adobe Analytics panel of over 1 trillion visits to US retail sites, 100 million SKUs, 18 product categories. Companion consumer survey. Geography: United States. Sample: 1 trillion+ visits (behavioral); n = 5,000 US consumers (survey).
    Limitations: Weighted toward large US retailers that run Adobe Analytics. Growth rates for January and April 2025 are indexed to July 2024, not year over year, because the base was too small before then. The non-AI comparator is a blend of paid search, affiliates and partners, email, organic search and social media, so a lift against it is not a lift against any single channel.
  • Method: Adobe Analytics panel of over 1 trillion visits to US retail sites, 100 million SKUs, 18 product categories. Geography: United States. Sample: 1 trillion+ visits.
    Limitations: Same panel and same blended non-AI comparator as the August 2025 release. Monthly figures, so single-month conversion differences can move with promotional calendars.
  • AI-driven traffic surges across industriesAdobe Analytics, 12 January 2026Grade A
    Method: Adobe Analytics panel of over 1 trillion visits to US retail sites. Companion Holiday 2025 Consumer Survey. Geography: United States. Sample: 1 trillion+ visits (behavioral); n = 1,000+ US respondents (survey).
    Limitations: Covers the 2025 holiday season (1 November to 31 December), a peak period with gift intent and deadlines. The companion survey is smaller than the August 2025 survey (1,000+ against 5,000), so the two survey series are not directly comparable. Same blended non-AI comparator as earlier releases.
  • Method: Company announcement. Geography: Global specification. Sample: Not applicable.
    Limitations: SDK download and server counts are self-reported at the time of donation and count integration activity, not commerce transactions.
  • Method: Consumer survey for the report "What Matters to Today's Consumer 2025", fielded October and November 2024. Geography: 12 countries across North America, Europe and Asia-Pacific. Sample: n = 12,000 consumers aged 18 and over.
    Limitations: Fielded in late 2024, so it is a dated baseline rather than a current reading. Multi-country, so not comparable with Adobe's US-only surveys. Whether a 2026 edition exists has not been confirmed.
  • Method: Company announcement. Geography: Global specification. Sample: Not applicable.
    Limitations: A governance announcement with no adoption figures.
  • Method: Company announcement. Geography: Global specification. Sample: Not applicable.
    Limitations: Announces optional capabilities. Says nothing about how many merchants have adopted them.
  • Google Analytics 4: known bot traffic is excluded automaticallyGoogle Analytics Help, Retrieved 8 September 2026Grade C
    Method: Product documentation. Geography: Not applicable. Sample: Not applicable.
    Limitations: Describes Google Analytics 4 only.
  • Google Analytics 4: attribution models and lookback windowsGoogle Analytics Help, Retrieved 8 September 2026Grade C
    Method: Product documentation. Geography: Not applicable. Sample: Not applicable.
    Limitations: Describes Google Analytics 4 only.
  • Method: Company announcement. Geography: Global specification. Sample: Not applicable.
    Limitations: A launch announcement with partner names and no adoption figures.
  • Under the hood: Universal Commerce Protocol (UCP)Google Developers Blog, 11 January 2026Grade C
    Method: Company announcement and technical overview. Geography: Global specification. Sample: Not applicable.
    Limitations: Describes the specification and its partners. Says nothing about live transaction volume.
  • Universal Commerce Protocol on Google: frequently asked questionsGoogle for Developers, Retrieved 8 September 2026Grade C
    Method: Product documentation. Geography: Not applicable. Sample: Not applicable.
    Limitations: Documentation for Google's implementation of UCP checkout only.
  • Universal Commerce Protocol checkout: availability and eligibilityGoogle Merchant Center Help, Retrieved 8 September 2026Grade C
    Method: Product documentation. Geography: United States, Canada and Australia. Sample: Not applicable.
    Limitations: Documentation is updated without a changelog, so availability wording may change after the retrieval date.
  • Verify requests from Google crawlers and fetchersGoogle Search Central, Retrieved 8 September 2026Grade C
    Method: Product documentation. Geography: Not applicable. Sample: Not applicable.
    Limitations: Describes Google's crawlers only.
  • ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels?Kaiser and Schulze, SSRN working paper 5585812, Posted 18 October 2025, last revised 27 July 2026Grade A
    Method: Twelve months of first-party analytics data from 973 websites with $20 billion combined revenue. Over 50,000 ChatGPT referral transactions compared with 164 million transactions from traditional channels, using regression models that account for data sparsity. Geography: Not stated in the abstract; pending confirmation from the full text. Sample: 973 websites; 50,000+ ChatGPT referral transactions; 164 million traditional-channel transactions.
    Limitations: An SSRN working paper that has not been published in a journal. Isolates organic ChatGPT referrals only. The twelve-month window ends before the late 2025 crossover in Adobe's US series. Geography of the 973 sites is not in the abstract.
  • Method: Foundation press release. Geography: Global specification. Sample: Not applicable.
    Limitations: Organization counts are self-reported by the foundation.
  • Method: Foundation press release. Geography: Global. Sample: Not applicable.
    Limitations: Governance announcement only.
  • MCP joins the Agentic AI FoundationModel Context Protocol project, 9 December 2025Grade C
    Method: Project announcement. Geography: Global specification. Sample: Not applicable.
    Limitations: Project's own statement of its governance.
  • What went wrong with ChatGPT's Instant CheckoutModern Retail, 27 March 2026Grade C
    Method: Reporting of comments by Walmart EVP Daniel Danker to Wired, and an Etsy statement. Geography: United States. Sample: Two named companies.
    Limitations: A named-executive disclosure reported by a trade publication. Walmart did not publish the underlying data. Two retailers, not a market measurement.
  • Powering product discovery in ChatGPTOpenAI, 24 March 2026Grade C
    Method: Company announcement. Geography: Global rollout to free, Go, Plus and Pro users. Sample: Not applicable.
    Limitations: Describes OpenAI's own product direction. Names retailers integrated for discovery but gives no usage or conversion figures.
  • OpenAI crawlers and botsOpenAI, Retrieved 8 September 2026Grade C
    Method: Product documentation. Geography: Not applicable. Sample: Not applicable.
    Limitations: Describes OpenAI's crawlers only.
  • Method: Salesforce Shopping Index, aggregated commerce data from over 1.5 billion shoppers on Salesforce-powered storefronts, with AI and agent activity measured through Agentforce and Commerce Cloud. Geography: Global, with a US split. Sample: 1.5 billion+ shoppers.
    Limitations: AI-influenced spend is defined by Salesforce as purchases influenced by personalized product recommendations and conversational customer service. It is retailer-owned AI, not LLM referral traffic, and the two cannot be compared. The Agentforce growth comparison is between Salesforce's own customers and non-users, not a random sample of retailers.
  • Method: Forecast, not a measured result. Geography: Global. Sample: Not applicable.
    Limitations: A forecast. Cited on this page only to show that the forecast ($73 billion, 22%) exceeded the reported result ($67 billion, 20%).
  • Salesforce Cyber Week 2024 resultsSalesforce, via Business Wire, 4 December 2024Grade A
    Method: Salesforce Shopping Index, Cyber Week 2024. Geography: Global. Sample: 1.5 billion+ shoppers.
    Limitations: Same AI-influenced definition as the 2025 release. Salesforce's measurement of AI influence has expanded as its own agent products launched, so year-to-year comparison partly reflects product coverage.
  • Salesforce Cyber Week 2023 resultsSalesforce, via Business Wire, 28 November 2023Grade A
    Method: Salesforce Shopping Index, Cyber Week 2023. Geography: Global. Sample: 1.5 billion+ shoppers.
    Limitations: The 2023 figure predates Salesforce's Agentforce products and measures AI-influenced spend through predictive and generative recommendations only.
  • Method: Company announcement. Geography: United States at launch. Sample: Not applicable.
    Limitations: A launch announcement. It contains no merchant fee figure and no volume figures. The product it describes was withdrawn as a native in-chat checkout in March 2026.
  • Universal Commerce Protocol specification: catalog capability (draft)Universal Commerce Protocol project, Retrieved 8 September 2026Grade C
    Method: Protocol specification, draft. Geography: Not applicable. Sample: Not applicable.
    Limitations: A draft specification that can change between retrievals.

Frequently asked questions

How much ecommerce traffic comes from AI?
No primary source publishes AI-referred traffic as a share of total retail visits. Adobe Analytics reports growth instead: AI-referred traffic to US retail sites rose 693.4% year over year across November and December 2025, after a 4,700% year-over-year rise in July 2025, measured as shoppers clicking a link from a generative AI chat service or browser across more than 1 trillion visits. Adobe describes the base as modest.
Does AI referral traffic convert better than Google search?
The two credible datasets disagree. Adobe's US retail panel reports AI referrals converting 31% more than a blended non-AI baseline in the 2025 holiday season, up from 49% worse in January 2025. Kaiser and Schulze's working paper, using twelve months of first-party data from 973 sites, finds ChatGPT referrals converting below every traditional channel except paid social. Different comparators, periods and populations; neither has been measured against the other.
What share of consumers use AI for shopping?
In Adobe's 2025 survey of 5,000 US consumers, 38% reported having used generative AI for online shopping and a further 52% said they planned to. The first is self-reported behavior and the second is intention, and they are not added together. Among those who had used it, 85% said it improved their shopping experience and 83% said they were more likely to use AI for larger or more complex purchases.
What is agentic commerce?
Commerce in which a software agent acts for a shopper: discovering products, assembling a cart, and in some designs completing payment. In practice, as of September 2026, OpenAI's ChatGPT handles discovery through the Agentic Commerce Protocol and sends purchases to the merchant's own store, having withdrawn its native in-chat checkout in March 2026, while Google's Universal Commerce Protocol offers checkout on Google surfaces to select merchants in the US, Canada and Australia.
How much retail revenue is AI-influenced?
Salesforce reports that AI and agents drove $67 billion in global Cyber Week 2025 sales, influencing 20% of all purchases through personalized product recommendations and conversational customer service, measured across more than 1.5 billion shoppers on Salesforce-powered storefronts. The series runs $51 billion in 2023, $60 billion in 2024 and $67 billion in 2025. This is retailer-owned AI, not LLM referral traffic, and the two are not comparable.
What is the difference between ACP, UCP and MCP?
ACP, from Stripe and OpenAI, is now a discovery and catalog layer for ChatGPT with checkout on the merchant's own store. UCP, from Google with Shopify, Etsy, Wayfair, Target and Walmart, is an open specification covering discovery, an optional cart, checkout and post-purchase, with payments via AP2. MCP, governed by the Agentic AI Foundation under the Linux Foundation, is a model-to-tool integration standard that defines no cart, checkout or payment primitives.
Is ChatGPT still the largest source of LLM referral traffic?
Probably, but no retrievable primary source on this page measures the split between assistants. An earlier version of this page cited a 97% ChatGPT share from a vendor dataset of 94 mid-market brands tracked through 2025; that figure was removed because the report could not be retrieved, the sample was too small to generalize, and a 2025 share should not be presented as the state of a market in which Gemini, Perplexity and Copilot have grown since.
Can retailers accurately attribute AI-assisted purchases?
Not with tag-based analytics alone. Google Analytics 4 excludes known bot traffic automatically, so automated requests from AI systems never appear, and its attribution models give no credit to direct visits, so a shopper who researches in an assistant and returns by typing the URL or searching the brand is credited elsewhere. Server-side logs with crawler identity verified against published IP ranges and reverse DNS are the only record of the automated side.