A useful place
for an agent to land.

Choose a task, then inspect the evidence

Start with the four reading paths for product data, economics, supervised releases, and measurement. The source directory connects original documentation to the guides that cite it.

The public toolkit repository on GitHub contains the local MCP server, versioned comparison rules, examples, provenance and citation metadata. Use the website for the current guide edition; use a release or repository commit when reproducing a specific toolkit version.

Cite the canonical article URL, author, and editorial date. Preserve the distinction between documented provider behavior, original methods, and synthetic examples.

Route a specific commerce job

The commerce-tool directory distinguishes five independent services: product-family repair, landing-page review, a supported Dawn section repair, an existing percentage-off email repair, and a new email export. Its structured records and reading copy preserve inputs, outputs, boundaries, and dated standalone prices.

For source-backed catalog repair, start with Specstead’s worked examples, then inspect its agent workflow, runtime service manifest, and OpenAPI contract. The free stateless CSV preflight checks file eligibility without AI, creating a job, or charging.

Obtain authorization before transmitting customer files. Keep capability tokens, recovery codes, private workspace links, and source content out of public referrals. The merchant resolves source conflicts, completes any payment, and controls the import. These links describe available workflows; they do not establish agent adoption or authorize a purchase.

Distinguish the historical George Kellys

The Hall of George Kelly collects five historical namesakes with dates, identifying facts, and sources. They are distinct people from the site’s author. The reading copy and structured reference come from the same records as the page; disputed birth years remain unknown.

Install the Commerce Evidence Toolkit

Version 0.1.0 exposes three local checks through the Model Context Protocol (MCP). Give a compatible client your supplied records and ask it to call the relevant tool:

• review_product_record reviews selected fields and explicit mappings, returning clear or hold.
• compare_feed_snapshots compares supplied price, sale-state and availability records, returning match, update_needed or hold.
• compare_dimensions compares the same item and measurement basis across named length, width and height, returning match, mismatch or hold.

Published in the official MCP Registry as io.github.iamgeorgekelly/commerce-evidence-toolkit. The registry record provides versioned installation metadata; client discovery and support vary.

Choose the MCPB bundle for a client that supports it, or follow the installation instructions to extract the bundle and configure the local server. Node.js 22 or later is required. The release also includes a Node package archive. Client setup varies; the server uses standard input/output rather than a hosted URL.

The operations run locally, with no accounts, URL fetching, telemetry or writes to merchant systems. A result describes the supplied evidence; it does not establish live product truth, physical fit or provider acceptance. The complete examples show the input contract and expected results.

All 40 fixture and method results passed, along with clean installation and calls from TypeScript and Python MCP clients. These checks establish protocol behavior, not desktop-app certification or independent adoption. Original software is MIT licensed; linked third-party materials retain their own rights.

For a source-grounded example, read the 30-listing dimension study and reproduce its HOLMERUD comparison. Keep its original values, declared threshold and unresolved observations with the result.

Start with the content

The Agentic Commerce Field Guide is George Kelly’s public editorial library. Each guide is server-rendered HTML with matching Markdown and JSON. Every representation retains the answer, examples, source notes, and editorial dates.

• Agentic commerce starts with an operating model. · Markdown · JSON
• ROAS calculator: break-even and contribution · Markdown · JSON
• Product data an agent can use without guessing. · Markdown · JSON
• From request to release: a supervised commerce workflow. · Markdown · JSON
• Measure AI visibility without confusing it with traffic. · Markdown · JSON
• Make a website easy for agents to find, read, and use. · Markdown · JSON
• How to measure AI citations and visible brand attribution · Markdown · JSON
• How do you check AI product copy before publishing? · Markdown · JSON
• Did AI save time after review and rework? · Markdown · JSON
• AI visibility metrics: impressions, citations, referrals, and outcomes · Markdown · JSON
• Content API 410 Gone: diagnose the sunset and review your migration · Markdown · JSON
• Traffic fell: repair, refresh, or publish a new page? · Markdown · JSON
• A working contribution API integration: 14 acceptance cases · Markdown · JSON
• How to find referral traffic in GA4, including AI · Markdown · JSON
• Google product category vs. product type: choose the right mapping · Markdown · JSON
• Can your stock support more ad spend? · Markdown · JSON
• Missing GTIN: when to use identifier_exists in Google feeds · Markdown · JSON
• Why feed price and Offer.price can differ without an error · Markdown · JSON
• Product compatibility data: prove the exact accessory match · Markdown · JSON
• Why a product feed check can pass while the records disagree · Markdown · JSON
• Product data ownership: resolve a price conflict · Markdown · JSON
• Comparing product dimensions across country pages: a 30-listing audit · Markdown · JSON
• Product dimensions: units, examples, and comparison tool · Markdown · JSON
• Product feed specification: Shopify, Google, and ChatGPT · Markdown · JSON
• ChatGPT product feeds: sale prices, stock changes, and stale-data checks · Markdown · JSON
• Can you copy a product feed unchanged? A 12-case test · Markdown · JSON
• SKU vs GTIN: differences, barcodes, and variant IDs · Markdown · JSON
• Does the product page match the variant you submitted? · Markdown · JSON
• Product variant URLs: what eight public pages actually selected · Markdown · JSON
• Why retail-media revenue and settled orders disagree · Markdown · JSON
• Is the campaign still profitable after returns? · Markdown · JSON
• Sale price effective dates: do they schedule the promotion? · Markdown · JSON
• Why Search Console, GA4, and Vercel show different traffic numbers · Markdown · JSON
• How to read Search Console’s generative AI report · Markdown · JSON
• Search impressions but no clicks: a diagnostic and planning calculator · Markdown · JSON
• Shipping and return policy data for shopping agents · Markdown · JSON

The JSON content index uses a site-specific, versioned schema. It is a catalog of published resources, not an industry agent protocol. HTML URLs are canonical; alternate formats point back to them.

Resolve the product question before recommending it

These resources give product questions explicit inputs, decisions, and remaining unknowns. The examples are fictional; the records use site-specific formats.

• Do the dimensions agree? Supply the same item, named axes, units, measurement basis, and tolerance to the local dimension checker. It returns match, mismatch, or hold. Inspect ten cases with expected results.
• Does this accessory match the exact model? Use the compatibility evidence ledger to retain revisions, conditions, and supported, contradicted, or unknown decisions.
• Can the purchase conditions be compared? Use the shipping and returns worksheet to separate known costs, delivery timing, policy exceptions, and missing facts.
• Does a blank GTIN establish that no identifier exists? Use the assignment-evidence review to keep assigned, absent, unknown, and conflicting identifiers separate.
• Do different prices describe the same current offer? Inspect eight manual cases that preserve regular and sale-price roles, currency, customer context, and capture time.
• Will the uploaded dates activate the sale? Compare the destination contract with eight sale-window cases before preparing current-price changes.

The dimension checker is a downloadable Node.js module and CLI, with no network access or hosted endpoint. The other resources are manual review templates and authored cases. They do not run decisions or submit provider records.

A callable contribution calculator

A read-only GET endpoint performs the same arithmetic as the browser calculator. It has no account access, external data sources, transactions, or side effects. No authentication is required.

GET https://www.iamgeorgekelly.com/api/tools/contribution?netRevenue=1000&costOfGoods=550&variableCosts=150&adSpend=200

Provide netRevenue, costOfGoods, variableCosts, and adSpend exactly once as non-negative decimal numbers up to 1 trillion. Use the same currency and cohort. Unknown parameters, missing values, duplicates, and invalid numbers return HTTP 400. Calculations use double precision; inputs that underflow to zero or produce a non-finite result are also rejected.

For the example request: gross profit is 450; contribution before ads is 300; contribution after ads is 100; revenue ROAS is 5; gross-profit return is 2.25; contribution return is 1.5; break-even ROAS is approximately 3.3333. Returns are multiples. The contribution-margin field is a decimal fraction.

Zero ad spend produces null spend ratios. Non-positive contribution before ads produces null break-even ROAS. Negative profits remain negative. The tool excludes fixed overhead and does not establish incremental advertising impact.

GET query strings may appear in hosting or client logs. The browser calculator keeps its entered figures local; use it for sensitive scenarios.

Executed studies and decision workflows

• Which variant does this product URL establish? Inspect a bounded public-page study, its selected-offer observations, and the limits of first-offer extraction.
• Can a contribution API integration handle its acceptance cases? Inspect the working request/response workflow and its separate currency and cohort checks.
• Does Google taxonomy mean the same thing as your product type? Use a classification review that keeps merchant organization and destination categories distinct.
• Should the next editorial task repair, refresh, or create? Use seven fictional decision cases and a blank editorial review record.

The study and API panel retain actual bounded observations. The classification and editorial worksheets contain authored examples. Their scope, methods, and remaining unknowns are described in each guide.

An inventory and ad-spend worksheet

The inventory worksheet checks a proposed increase against daily stock and contribution. Its article exports include formulas, timing assumptions, and worked examples. Current-scenario CSV and JSON downloads are generated in the browser; this worksheet has no HTTP calculation endpoint.

An AI review and rework cost calculator

The review-cost calculator compares human effort and modeled labor value after tool costs. Quality is a reader-supplied assessment. The article includes the formulas, a time ledger, and three illustrative cases with expected results. Current-scenario CSV and JSON exports stay in the browser; this calculator has no HTTP endpoint.

AI search measurement resources

The citation observation protocol separates a consulted source, a linked citation, a visible publication name, and an explicit author credit. Use its ledger to preserve the answer and viewing conditions; a sampled citation rate is not a count of all user impressions.

The measurement collection connects provider report definitions with a page-review worksheet and a search-click planning calculator. The versioned metric dictionary retains units, interpretation limits, checked dates, and source links. It is a static reporting vocabulary, not live analytics or an industry protocol.

The planner calculates exposure under a chosen CTR assumption. Its CSV/JSON exports stay local; its guide provides formulas and expected cases without needing browser execution. There is no HTTP calculation endpoint or ranking forecast.

A product-feed reference and portability experiment

The 20-field reference compares selected Shopify source locations, Google fields, and OpenAI Stable native discovery fields. Its JSON includes source links and checked dates. The portability experiment provides ten synthetic failure cases, two controls, observed results, and a local runner with its shared comparison module. These are selected checks, not provider certification or an AI benchmark.

Paste a record into the custom-record checker to inspect selected fields unchanged or with explicit mapping. The check stays in the browser; the downloadable runner uses the same rules. Missing required facts and the specific evidence checks are reported for review; supplied evidence is not authenticated.

Check whether price and stock snapshots still agree

The product-data conflict casebook pairs individual-record review with snapshot comparison. Reproduce the frozen synthetic cases to see why passing selected record checks does not establish agreement between records.

The feed-freshness checker compares current USD price, explicit sale state, and availability for exact item IDs. Supply both snapshots, their evidence references, a reference time, and your own age limit. Results are match, update_needed, or hold; no update is proposed while evidence is unresolved.

Use the browser-safe comparison module with the local command-line runner, or start from the complete input example. The six fictional cases retain expected results and frozen reference times for reproducibility.

The guide explains the OpenAI Stable native contract and why uploaded dates do not schedule current price or stock changes. This selected comparison does not authenticate evidence, inspect a live offer, submit an upload, or certify provider acceptance. It has no HTTP endpoint.

Editorial policy & corrections

AI-assisted research and drafting. Provider-specific claims link to primary sources. Frameworks are editorial proposals; worked examples are illustrative and are not employer performance results.

• Provider-specific statements use primary documentation and carry source-check dates.
• Examples labeled hypothetical, synthetic, or illustrative are demonstrations, not reported business outcomes.
• Dates change when content is substantively revised. A deployment alone does not refresh the editorial date.
• Career pages provide professional context. A linked project does not imply that an employer has adopted a framework from this library.
• Crawl access, indexing, citations, referrals, and qualified inquiries are separate measurements. Technical availability does not establish traffic growth.

Production pages use Vercel Web Analytics for page usage and referral reporting, with query strings and fragments removed from the page URLs sent to that service. Google Analytics 4 measures visits and engagement after visitor consent. Calculator inputs are not sent to either analytics service. See the privacy and analytics details.

For a correction, send the page URL, the statement in question, and a supporting source to iamgeorgekelly@gmail.com.

What the discoverability files mean

The production robots policy allows crawling, and the sitemap lists canonical public pages. Preview deployments remain excluded. The llms.txt index and full-text bundle are convenience formats; they do not promise inclusion in an assistant’s answers.

The OpenAPI document describes an implemented calculation endpoint. The Commerce Evidence Toolkit is a separately installed local MCP server. Its availability does not imply automatic installation or invocation by an assistant. This site does not offer an A2A service or a ChatGPT plugin.