Commerce Graph Method v1.2

We start from the demand side. For a category we build a prompt frame of 300–600 buying-intent prompts covering product discovery, comparison, budget bands, use case, and retailer-led phrasing. Prompts are run across the assistants that matter for the category and the cited sources are recorded per answer.

The cited URLs, not the prompts, are the unit of analysis. We rank cited domains and pages by citation frequency, then crawl the pages that carry the most citation weight — typically roundups, reviews, and category hubs on editorial publishers. Every crawled page is retained with its fetch timestamp so a result can be reproduced against the version we saw.

Publisher pages repeat the same destination many times: an inline mention, a product card, a sticky price widget, and a footer roundup can all point at the same merchant. Counting raw anchors overstates a merchant's share, so we deduplicate at two levels.

First, identical resolved destinations within a single page collapse to one link, with a count of placements retained separately. Second, we normalise destinations before comparing them — tracking parameters, affiliate identifiers, session tokens, and locale variants are stripped so amazon.com/dp/X with four different tags resolves to one merchant destination. Counts on our pages are deduplicated link counts unless a figure explicitly says placements.

Affiliate links rarely point at the merchant directly. They pass through networks and wrappers, often two to four hops. We follow each chain to its terminal URL, with a hop ceiling, a timeout, and no JavaScript execution.

Across published benchmarks we resolve 93–97% of affiliate-wrapped links to a terminal merchant. The remainder fail for identifiable reasons: geo-gated redirects, wrappers requiring a live session, expired campaign IDs, and bot-blocked networks. Unresolved links are reported as unresolved. They are never redistributed across merchants pro rata, because the failures are not random — they skew toward networks with aggressive bot defences.

Routing is a property of a page at a moment in time. Every crawl carries a fetch date, and a benchmark reports the window in which its pages were fetched. Published benchmarks use a rolling 30-day window; pages older than 90 days at time of crawl are flagged as stale and excluded from headline figures.

Because assistants cite pages that are themselves updated on the publisher's own schedule, we re-crawl the highest-weight pages monthly and treat a changed merchant mix as a finding rather than an error.

Each resolved destination is assigned to exactly one class: brand-owned DTC, marketplace, authorised specialty retailer, department store, mass or club retailer, reseller or grey-market, or non-commerce. Classification runs off the registrable domain against a maintained merchant registry, not off page text.

Brand-owned means the brand controls the checkout. A brand's storefront inside a marketplace is classified as marketplace, because the demand lands in the marketplace's cart and the brand does not own the transaction or the data. Ambiguous domains are reviewed by hand and the decision is recorded in the registry so it stays consistent across categories.

Not every link on a page earns equal traffic, and not every cited page earns equal citation volume. We weight in two stages. Citation weight distributes category demand across cited pages in proportion to observed citation frequency. Placement weight then distributes a page's clicks across its links using position, prominence, and link type — an above-the-fold primary buy button carries far more weight than a footer link.

Placement weights are calibrated against published affiliate click-distribution research and, when a client shares it, their own outbound referral data. Weighting affects share-of-clicks figures. It never affects the link and merchant counts, which are direct observations.

Value at risk is modeled, and we always publish it as a range rather than a point estimate. The chain is: category demand volume, times weighted share of clicks landing on each merchant class, times a conversion rate band for that class, times a category average order value band.

Three assumptions drive almost all of the spread. The conversion band by merchant class matters most — marketplace conversion runs multiples of DTC conversion, so small changes to that ratio move the total sharply. Average order value band is second, particularly where a category spans entry and premium price points. Click-through from the assistant answer to the cited page is third and the least observable from outside. We report the range endpoints, state the assumption set behind each, and label every such figure as modeled on the page it appears on.

This is a routing measurement, not a sales measurement. It cannot see actual orders — we observe where a click lands, not whether it converted or what it was worth.

It cannot establish incrementality. A shift in routing is not proof that revenue moved, and none of these figures should be read as a causal claim about demand that would not have existed otherwise.

It cannot see margin. Wholesale terms, retailer funding, and channel-level profitability are yours, not ours, and they can invert the ranking of a channel entirely.

It cannot see identity. There is no person, household, or cross-device journey in this data — only pages, links, and destinations. Any brand-specific work that needs orders, margin, or identity uses your first-party data alongside this, and we label which is which.