Which Agentic Commerce Statistics Show Actual Adoption?

Agentic commerce statistics measure consumer demand, merchant adoption, infrastructure readiness, transaction activity, and forecast scenarios. For merchants appearing in ChatGPT, Gemini, Perplexity, or Copilot, completed orders and production logs show observed behavior; surveys show reported attitudes; forecasts show possible scale. No single figure proves market-wide adoption.

A merchant can enable a product feedA structured export of the catalog (id, title, description, link, image, price, availability) in the Google Merchant Center spec, the format most shopping surfaces ingest.Read more today and still have zero agent-completed orders. That gap matters for US brands and ecommerce operators reading adoption claims about ChatGPT, Gemini, Perplexity, and Copilot.

These systems can increasingly read current merchant feeds, indexed product pages, schema.orgMachine-readable schema.org markup (Product, Offer, AggregateRating) embedded in a page as JSON-LD. It is the canonical way to hand engines unambiguous product facts.Read more facts, and clean catalog records. If a size variant has no identifier or the feed says $89 while the page says $99, that product may fail an eligibility check or create trouble during comparison and checkout, depending on the platform.

This article separates observed behavior from surveys, deployments, transaction records, and forecasts. It also shows what each number can prove about a store.

What do the five evidence classes prove?

Agentic commerce statistics provide evidence about consumer demand, merchant adoption, infrastructure readiness, transaction activity, and future scenarios. Each class answers a different question. A survey can measure willingness, a protocol release can establish technical availability, and an order record can show completed behavior within a defined system.

That distinction prevents a common analytical error. A merchant announcement, a consumer saying they would delegate a purchase, and a completed delegated payment are three separate events with three separate denominators.

The unit changes the conclusion.

Six terms keep the numbers honest

Agentic commerce: software agents participate in product research, comparison, cart creation, checkout, or payment under user authority.

Buying agent: the buyer-side system acting on a shopper’s intent, such as finding a waterproof carry-on under $250 and comparing eligible products.

Agentic Selling: the seller-side discipline through which a brand studies the market, readies its store, and sells to people and buying agents.

Observed metric: a measured event or production deployment, such as 12,000 completed checkout sessions during a stated month.

Survey result: a reported attitude, intention, or claimed behavior from a defined sample using a disclosed question.

Forecast: a modeled future scenario built from assumptions about adoption, spending, coverage, and time.

Verdict: A credible statistic exposes its denominator

A credible evidence ledger preserves the unit, denominator, observation period, geography, and source. “Thousands of merchants” is weak by itself. “5,000 enrolled merchants shipping to the US as of November 2025,” when reported by the program operator, is a scoped program statistic.

Transaction records and production activity show behavior. Surveys show responses. Forecasts show scenarios.

Verdict: Demand percentages need the exact question

Demand evidence moves through distinct stages: awareness, current use, willingness to delegate, trust, purchase intent, and completed delegated purchase. Movement from one stage to the next requires separate measurement.

The appeal is obvious: one large percentage can make the market look settled. The evidence does not support that shortcut because question wording, sample composition, country, and fieldwork date change the meaning of every survey result.

A defensible demand table records the exact question before recording the percentage. This edition excludes survey percentages that cannot be checked against an original questionnaire and methodology file.

This evidence control keeps unsupported data out of the analysis. Adding an unattributed “AI shopping” percentage would make the page look fuller while weakening the conclusion.

A useful survey also separates a recommendation from delegated action. Asking ChatGPT for running-shoe ideas is AI-assisted research. Authorizing a buying agent to choose a size, create a cart, and submit payment is delegated commerce.

Verdict: Merchant enrollment proves participation at a stated level

Merchant adoption has at least five levels: announcement, pilot, protocol support, production availability, and active usage. Count the level stated by the primary source.

Perplexity reported that its Merchant Program had expanded to more than 5,000 merchants in November 2025. Its program announcement establishes catalog participation at that date and within the stated scope of merchants shipping to the US. The announcement does not report orders for every enrolled merchant.

OpenAI’s current product-feed specification documents fields and eligibility rules for product ingestion. Because specifications change, this article does not freeze an attribute count. A feed that passes the applicable checks establishes technical readiness for that profile and version.

Google announced the Universal Commerce Protocol, or UCP, with Shopify and named partners in its NRF remarks. That source supports an announcement and its stated partner scope. Merchant-level activity requires separate usage records.

Access labels belong beside every deployment claim:

A logo slide cannot supply the missing denominator.

Verdict: Transaction activity is strongest and scarcest

A completed order records behavior. Merchant-side checkout completion, delegated-payment success, repeat use, cancellations, and refunds reveal more than impressions or feature availability.

The denominator is decisive. A 92% payment success rate means little if it excludes authentication failures before the payment attempt. Likewise, 10,000 completed orders need a time window, geography, merchant count, and refund policy.

No public protocol repository can fill this table for the whole industry. Repositories document capabilities and releases. Industry totals require transaction records from a defined population.

Verdict: Forecasts describe possible scale

McKinsey modeled a scenario in which the US business-to-consumer retail market could see up to $1 trillion in orchestrated revenue by 2030. Its original analysis defines the scenario and its boundaries. “Could,” “up to,” “US B2C,” “orchestrated,” and “2030” are essential qualifiers.

Orchestrated revenue may include spending influenced or coordinated by agents. The figure does not automatically equal autonomous checkout volume, payment value processed by one protocol, or incremental retail revenue.

Forecasts belong in the scenario column of the ledger.

Buyer-side agents create seller-side work

Buying agents express and execute shopper intent. A brand agent covers the seller side of agentic commerce, where the store answers with current product facts and carries the order onto the brand’s commerce rails.

Buyer-side adoption estimates say little about whether a specific brand can answer, “Is the navy size 10 in stock, and can it arrive in Austin by Friday?” That readiness requires a SKU-level test.

Five measurable stages connect a question to an order

A useful measurement model follows the transaction path: question, retrieval, comparison, cart, and merchant checkout. Each stage creates a separate event that can be tested.

1. The question creates a retrieval task

“Find a carry-on under $250 that fits Delta limits” creates a request for dimensions, current price, stock, and suitability. Retrieval-augmented generation, or RAG, means an AI system can fetch external records or pages while producing an answer. Some commerce surfaces use retrieval because price and stock can change between model-training cycles.

Measure the query set, retrieval success, market, and test date. A test with 100 fixed prompts remains comparable over time only when the prompts and engine settings remain fixed.

2. Feeds and indexed pages supply product facts

Commerce surfaces can draw from combinations of merchant feeds, indexed pages, structured data, and platform catalog records. Access, freshness, and supported fields vary by platform and account, so each test should name the surface and date.

On the open web, schema.org Product, Offer, and AggregateRating markup can make facts explicit to machines. Core schema.org does not impose one universal validity profile. For Google product rich-result eligibility, price is required within an Offer, while properties such as priceCurrency and availability follow Google’s current required or recommended rules for the applicable experience. If the page says “In stock” and the feed says “Out of stock,” record a source divergence.

3. The engine compares eligible products

Eligibility can place a product into a candidate set. Selection depends on the engine, query context, available data, and proprietary systems.

OpenAI says organic shopping results can consider relevance, availability, price, quality, and whether the merchant is the maker or primary seller. Its shopping documentation states those factors, while their exact weights remain undisclosed.

4. Protocols can carry cart and authorization data

Commerce protocols describe different parts of product exchange, cart creation, checkout, and payment authorization. Their supported stages, access rules, and maturity vary by specification version and implementation.

The Model Context Protocol, or MCP, can expose resources and callable tools such as product search or inventory checks. The MCP specification does not itself define a universal commerce checkout or payment-authorization flow. The agentic commerce protocol map explains the current roles and gaps.

5. Merchant records establish the completed order

In one common reference flow, the merchant or its commerce provider returns the authoritative cart, handles an approved payment method, and records the order. Implementations vary, including who creates the cart and how payment authority is represented.

The practical measures are checkout creation, valid-cart rate, payment success, cancellation, fulfillment, and refund. This closes the earlier denominator problem: protocol support can establish readiness, while merchant order records establish completed purchases.

Why does this matter when an agent does the buying?

ChatGPT, Gemini, Perplexity, Copilot, and AI Overviews can increasingly use structured merchant feeds, indexed product pages, and schema.org facts, with availability and mechanics varying by surface. Each engine controls its own retrieval, eligibility, and selection decisions.

A market forecast says little about whether an engine can read a brand’s current price, stock, size, shipping promise, return policy, and checkout path. For a merchant, the immediate question is whether a navy jacket in size 10 appears as the same variant across the page, feed, and checkout response.

That is why generative engine optimization for ecommerce starts with readable, consistent product facts. An article can earn a citation from prose. A product with current commercial facts is easier to compare and carry into checkout.

Seven questions expose a weak statistic

A statistic passes only when another analyst can reconstruct what the number means. Reproducing the result may require private records, but reproducing the definition should be possible.

Merchants can measure readiness before public totals exist

Field completeness: measure the percentage of SKUs with the properties required by each target feed or consumer profile. Segment the full catalog because perfect hero products can hide a weak long tail.

Surface divergence: compare the product page, feed, and commerce endpoint. Count every price, stock, and variant mismatch by SKU.

Propagation latency: time the interval from a stock or price change in the commerce system to its appearance on the page, feed, and sampled engine answers.

Answer accuracy: run a fixed prompt set by engine and compare stated price, availability, and specifications with the live source. Preserve screenshots, timestamps, surface, and market.

Checkout coverage: measure the percentage of active SKUs eligible for the relevant agent checkout implementation, then record checkout-session completion and payment success with their denominators.

Answer or citation share is an outcome proxy inside a fixed prompt set. It does not establish revenue or a position outside the measured prompts.

The verified timeline records availability

Infrastructure availability and usage are separate records. This compact timeline includes milestones supported by the cited primary pages and avoids unsupported release dates.

A launch proves that infrastructure existed at its stated scope. Usage requires another row backed by activity data.

Five actions turn market claims into store measurements

1. Label every number as transaction, deployment, readiness, telemetry, survey, or forecast evidence.
2. Preserve denominators, measurement dates, geography, exclusions, and exact question wording.
3. Prioritize transaction records and SKU-level readiness measures when allocating implementation work.
4. Audit the agent view from outside the store, then compare current price, stock, variants, shipping, returns, and checkout across each relevant surface.
5. Repeat the same fixed tests after every catalog or protocol change, so movement reflects the store and the documented method.

Veliu’s brand agent studies the market, readies the store, and sells to people and to the AI shopping agentsSoftware agents that search, compare and complete purchases on a buyer's behalf. They read structured catalog data and transact through protocols like ACP and MCP.Read more that arrive. Catalog retrieval, normalization, feeds, and commerce endpoints help it answer with a current price and a valid variant. On the brand’s site, it can adapt the selling experience from permitted signals and approved components, answer shopping agents machine to machine, and return the questions customers asked. Checkout stays on the brand’s rails.

Author: Veliu Editorial Team

Editorial methodology: This evidence ledger prioritizes first-party product documentation, official protocol repositories, standards organizations, and original research methodology. Every numerical claim is classified by evidence type and checked for date, geography, denominator, production scope, and limitation. Vendor statements establish what the vendor documented within the cited scope. The timeline excludes milestones whose exact dates or release status could not be confirmed from the linked primary record.

The next operational step is concrete: export 100 active SKUs and count every price, stock, identifier, and variant mismatch across the page, feed, and checkout response.