Agentic commerce statistics: a study of AI citations.

What the sample includes

We select citation rows labelled Tru Commerce and non-branded. The selected rows represent ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode, and the US, India and UK. Repeated citation observations are retained. Questions are counted by exact prompt text, only where that text has a citation in the export.

This is a fixed, selected prompt panel, not a census of shoppers or all AI answers. The presence of five engines and three regions does not prove every prompt ran successfully in every combination.

Correction, 23 September 2026. An earlier version confused a reported response count with unique questions, included branded citations in its headline totals, and used a two-brand fragment total. This version consistently uses the non-branded Tru Commerce subset. Vendor-mention and causal claims have been removed because this export cannot establish them.

Which domains appear across the most prompts?

Breadth counts distinct prompt texts associated with a domain. Citation count measures recorded references, including repeats. These are the ten domains with the greatest breadth in this sample, not a ranking of trust, traffic or commercial recommendations.

Inspect source coverage separately from whether an answer recommends a vendor. A source reference does not establish purchase intent. Our measurement methodology explains the distinction between observed coverage and business outcomes.

What can URL text fragments tell us?

The subset includes 2,704 nonempty decoded URL text-fragment observations. Their median whitespace-delimited length is 13 words and their median character length is 91 Unicode code points (Tru Commerce export, September 2026).

A text-fragment directive can contain start and end selectors. It is not necessarily a contiguous quotation or the complete text used by a model. These measurements describe the exported fragment strings, not the average length of every AI citation. Rows without fragments are excluded from these medians.

Method, reproducibility and limits

The downloadable aggregate includes every domain count, the brand and query-type filters, and a SHA-256 fingerprint of the source archive. It contains no raw prompts or answer text. The raw archive is not distributed with this page, so the aggregate alone cannot independently reconstruct the original extraction.

The snapshot date comes from the export handoff. Rows have no individual timestamps or response identifiers. We cannot reconstruct total answers, uncited answers, weekly changes, or extraction completeness from this file. This page is a dated snapshot, not an automatically refreshed weekly report.

Domain breadth does not prove trust. The sample cannot establish that shorter pages cause more citations, that one regional strategy outperforms another, or that a citation produces a visit or sale. For interpretation, see the agentic commerce value chain, the commerce glossary, and our case studies with separate measurement windows.

For your own store, define coverage and outcome metrics before making changes. Our commerce integrations and implementation guides cover the next steps.