AI referral traffic is spreading across four assistants. Ask what all four can read.

TL;DR

ReFiBuy's Q2 2026 AI1000 counts ChatGPT as the largest AI traffic source at 722 of the 1,000 largest US online retailers, down from 844 in Q1, with Gemini rising from 16 to 32, Perplexity from 8 to 21 and Claude from 1 to 15. Those are counts of retailers, not traffic shares, so they say the mix moved, not by how much. In the same July, Adobe Analytics, as reported by Digital Commerce 360, measured AI-referred visits converting 60% better and generating 53% more revenue per visit, while 39% of retail homepages in its cohort could not be read by a large language model. The 4GPTs read: a retailer cannot pick its assistant, so it should optimize for the one thing all four of them read, the catalog.

An assistant you have never integrated with sent your store more visitors last quarter than the one you built a product feed for. That is the agentic commerce 2026 story inside the two datasets Digital Commerce 360 put side by side on 3 September, and neither dataset says which assistant to build for. This piece reads both with a critical eye, then says what 4GPTs concludes.

What moved in one quarter, and what the count actually counts

ReFiBuy's AI1000, co-developed with Digital Commerce 360, scores the thousand largest US online retailers on four signals: bot friendliness, AI source traffic, diversity of AI sources and 90-day momentum. One of its Q2 2026 tables counts, for each retailer, which assistant is its largest source of AI-referred traffic.

"In Q1, 844 retailers had ChatGPT as No. 1," Scot Wingo, ReFiBuy's founder and chief executive, told Digital Commerce 360. "That decreased to 722." Gemini doubled from 16 to 32. Perplexity went from 8 to 21. Claude went from 1 to 15, which Wingo called "a really big move".

Read the caveat before the numbers. ReFiBuy prints it on the page: counts of retailers, not traffic shares. Each retailer casts one vote, for whichever assistant sent it the most visits, and a retailer where Gemini passed ChatGPT by a dozen sessions flips the count exactly as much as one where Gemini took over. ChatGPT is still first at 722 of 1,000, which is seven in ten. Some arithmetic of ours, not ReFiBuy's: the four listed assistants add up to 869 retailers in Q1 and 790 in Q2, and the page does not say what leads at the other 131 and 210, or whether they had measurable AI traffic at all.

One more thing to hold in mind. ReFiBuy produced the data and the scoring, coined the term Agentic Commerce Optimization, and sells that optimization; how it measures AI-referred traffic is not disclosed on the page or in the write-ups. That does not make the counts wrong. It says what to lean on. Counts survive an undisclosed method better than projections do, which is the rule this site applied when it read the count, not the ranking in the same AI1000 release.

The 4GPTs read: the count supports one sentence, and one is enough. In ninety days the assistant that mattered most changed at more than a hundred large retailers, the gains split three ways, and not one of those retailers had a vote.

Which AI assistants send retailers the most traffic, and does it matter?

ChatGPT, by a distance, and then a spread that did not exist in Q1. That is the direction. The scale cannot be read from this table, and the projections built on it deserve the same caution as the ranking did. Wingo's line that Claude will "be in the hunt" with Gemini and ChatGPT "if that continues over eight quarters" is a projection over eight quarters of a series that is two quarters old.

What makes the Claude line plausible is not the count but what shipped around it. Anthropic put product cards into Claude, per Wingo, and on 2 September released prebuilt shopper-agent and merchant-agent designs, with Shopify, Visa and Mastercard named as early adopters. Each assistant now runs its own product-card program and its own way of reading a store. A retailer that built for one of them in January was building for a shelf that had moved by June.

Does it matter which one leads? For the assistant vendors, entirely. For a retailer, less than it looks, because the retailer cannot change it. The mix is decided by which app a shopper opens, and shoppers decide that on a different clock from any roadmap.

What the agentic commerce 2026 traffic is worth, measured in transactions

The reason to care about a minority traffic source is what it does when it arrives. Adobe Analytics, as reported by Digital Commerce 360 on 19 August, put July 2026 AI-referral traffic to US retail sites up 62% year over year, and up 1,219% since October 2024.

Digital Commerce 360's own proportion sentence belongs next to those three: the volume of AI referrals "has not yet overtaken referrals from traditional search engines". Adobe's AI-referred shoppers also spent 59% more time on site, bounced 33% less and added to cart 28% more.

Two things make this the number worth building on. It is measured on transactions, from more than a trillion visits to US retail sites, rather than asked in a survey. And its cohort is stated: more than 200 of the Top 2000 retailers ran Adobe's analytics in 2025, so this is large, enterprise, instrumented retail, and it says nothing directly about a store too small to run Adobe. Read it as a floor for the retailers it covers, not as a fact about all of retail.

The gap is on the retailer's side, and it is counted

The same Adobe release carries the number that turns this from a traffic story into a legibility story. Adobe built a checker that reads a page the way a large language model does and reports what the model can and cannot read. In April, about 25% of retail homepage content in its sample was not optimized for large language models. By July, on a broader set of US retail sites, visibility was 61%, which means 39% of those homepages were not machine-readable. The sample changed between the two readings, so that is two snapshots, not a trend line.

Adobe's gloss on its own chart is the useful part. For apparel, electronics and cosmetics, "consistent and structured product content is helping drive AI visibility". For the rest, teams need to "ensure content can be easily parsed by machines". And the cost, in Adobe's words:

Put the two datasets together. The traffic converts better and pays more per visit. The assistants sending it changed places in a quarter. And four in ten retail homepages in Adobe's cohort cannot be read by the thing sending the traffic. Machine legibility already had a price tag; this quarter it acquired a second buyer, then a third.

The 4GPTs read: optimize for the reader you cannot choose

Four assistants, four product-card programs, one catalog. Building for one assistant means renting a shelf in a shop whose front door moved at a net 122 large retailers last quarter, and the rent is an integration you maintain on that vendor's schedule. Making the catalog legible means every one of the four reads the same price, the same stock and the same size chart on the day it changes, and so does the fifth reader that has not arrived at scale yet.

That fifth reader is the one 4GPTs is built for: the shopper's own agent, running on a personal agent host the shopper controls, holding the shopper's requirement and scoring offers against it. It reads evidence, not advertising, and the evidence it reads is exactly what Adobe's checker looks for: structured, current product data a machine can parse without guessing. A feed can be read. A store can be asked. The store that can be asked is the one that turns up in every assistant's answer and on the shortlist of the shopper's agent, without building for any of them one at a time.

The 4GPTs read, stated as a position: the assistant mix is not a retailer's decision and never will be. Legibility is, and it is the only decision that pays out on all four assistants at once. Connect once. Sell to every agentic surface.

What a merchant does on Monday

None of this needs a new integration this week. It needs a reading and a rule.

FAQ