Agentic commerce statistics: what 1,795 AI shopping answers reveal.

The dataset

This is a running census of how AI shopping answers get built, not a survey of opinions.

Every week we put 1,795 non-branded buyer questions - “what's the best agentic commerce platform”, “how do I sell through ChatGPT”, “which checkout works with AI agents” - to five AI engines: ChatGPT, Gemini, Perplexity, Copilot, and Google AI Mode. We run each across three regions - the United States, India, and the United Kingdom - and record every source the answer cites and every vendor it names. Branded questions are tracked separately and deliberately kept out of these numbers, so nothing here is flattered by people asking about us.

The September 2026 corpus holds 13,246 citations across 2,099 distinct domains. That is the raw material of the agent surface: the specific pages an engine reaches for when it decides what to tell a shopper. The pattern in that material is the whole story.

AI answers cite a small, stable set of sources

The engines pull from a narrow shortlist, and the same domains reappear across unrelated questions.

We rank sources by breadth - the number of distinct questions a domain is cited across - not by raw count. Breadth identifies a source the model trusts by default; volume just means one question ran on many engine-region pairs. On breadth, the leaders are LinkedIn, YouTube, Shopify, OpenAI, and Reddit. Two of the top seven are documentation domains, and the payments layer (Stripe) is cited as a source across 30 different questions on all five engines.

trucommerce.ai sits inside that shortlist - cited 230 times across 20 distinct questions on all five engines - which is why we can report this at all. Being cited, though, is not the same as being recommended, and the gap between the two is the most important number on this page.

Being cited and being recommended are different things

An engine can read your page for facts and still name a competitor as the answer.

When we count vendors named as the answer - distinct from sources cited - the category leaders are Shopify (625 mentions), Stripe (273), and WooCommerce (169). These are the platforms an engine offers up when a shopper asks who to use. A domain can be a heavily-cited source and a rarely-named vendor at the same time; the engine takes the facts from one place and hands the customer to another. Closing that gap is a content-shape problem, and it is exactly what our Citation Rank and Share of Voice products measure.

A citation is 13 words long

The average fragment an engine lifts from a page is one short, self-contained sentence.

Across 5,767 exact quoted spans, the median citation is 13 words, or 67 characters. Only 13.4% contain a number and just 1.5% contain a percentage - but 22.7% use definitional phrasing (“X is”, “means”, “refers to”). The lesson runs against instinct: definitions get retrieved; statistics persuade the human who reads the answer. Pages that win citations open each section with one clean definitional sentence and then support it with a sourced number - in that order, not the reverse.

The engine matters more than the country

Which AI engine a shopper uses reshapes the answer far more than where they live.

ChatGPT and Gemini share only 4 of their top 30 cited domains. The UK and US share 25 of 30. In practice that means a per-engine strategy beats a per-region one: ChatGPT leans on primary, spec-like sources (openai.com, help.openai.com, shopify.com); Perplexity leans on third-party corroboration and listicles; Gemini pulls community and definitional pages. Optimising once and expecting it to carry across all five surfaces is the most common mistake we see.

Shorter pages get cited more widely

Length is inversely correlated with citation breadth across the whole corpus.

Pages cited across 10 or more distinct questions have a median length of 1,202 words; pages cited across fewer run to 1,945. The single most broadly-cited page in the set is 607 words. What correlates with breadth is not word count but primary-source authority and roughly seven clean sections. Padding a page to hit a length target actively costs citations.

What this means for merchants

You can engineer for citations, and the shape that earns them is now measurable.

Write the definition before the statistic. Keep pages tight and structured. Target the engine, not the country. And treat the citation-to-recommendation gap as the metric that matters - being read is table stakes; being named is the win. If you want to see where your own brand sits in this corpus, the free Citation Rank scan runs your domain against the same five engines, and our value-chain research explains why owning the agent loop matters more every quarter. For the underlying terms, the agentic commerce glossary defines every concept referenced here.

Method & sourcing. All figures are first-party, from the Tru Commerce prompt tracker, corpus dated 16 September 2026. Five engines (ChatGPT, Gemini, Perplexity, Copilot, Google AI Mode) × three regions (US, IN, GB), 1,795 non-branded questions, re-run weekly. Source counts are reproducible from the published cited-source dataset; vendor-mention counts are from the same corpus. Numbers update as the tracker re-runs.