The AI Shopping Agent Readiness Index — How Much of the Digital Shelf Can an Agent Actually Read?

Introduction

Three buyers.

Heads of e-commerce and digital shelf leads at brands and retailers who are being asked, internally, whether their catalogue is "ready for AI shopping" and have no evidence either way. They need a scoring method they can run against their own listings and a benchmark to sit against.

Founders and heads of growth at direct-to-consumer and marketplace-native brands, who need to know whether the listings carrying their revenue are legible to a machine that is increasingly the first reader.

Data and platform leads who will have to fix whatever this report finds — the people who own feeds, PDP templates, structured markup and identifier hygiene.

On inquiry themes: through September, a recurring pattern of inbound questions clustered around three things — whether product listings are structured well enough for AI assistants to use, how to audit a catalogue's identifier and attribute coverage at scale, and how to monitor whether a retailer's own pages and their structured markup actually agree. This report answers all three with measured evidence instead of opinion.

What this report measures

This report scores a defined sample of public product listings against a published rubric of what an agent demonstrably needs in order to answer a shopping question and complete a purchase: unambiguous product identity, structured attributes with units, a resolvable price with currency, location-aware availability, legible shipping and returns terms, a stable purchase URL, and agreement between what the page shows a human and what its structured markup tells a machine.

The output is a readiness score per listing, aggregated by platform, category and country, plus the distribution of specific failure modes.

What this report does and does not measure

It does measure: data legibility. Whether the fields an agent needs are present, machine-readable, internally consistent and stable across observations. All of it from what is publicly observable on a listing.

It does not measure: market share, brand share, sales, conversion, traffic, or revenue of any kind. It does not measure agent behaviour — we do not run agents against listings and we make no claim about which listings an agent would choose. It does not measure the quality of the product, the competitiveness of the price, or the reputation of the seller. It does not measure data a retailer supplies privately through a merchant feed, an API partnership or any other non-public channel.

A listing that scores poorly here may sell extremely well. The two are different questions.

The readiness rubric

The rubric is Actowiz Solutions' own. It is published in full so readers can dispute it, re-weight it, or run it themselves.

Each dimension is scored 0 or the full weight — pass or fail, no partial credit — to keep the index reproducible. A listing's total readiness score is the sum of weights passed, out of 100. Scores are then banded: Agent-ready (85–100), Partially readable (60–84), Weakly readable (35–59), Not agent-readable (0–34).

Dimension 9 is deliberately weighted as heavily as identifier presence. A listing whose markup disagrees with its own page is not partially readable — it is actively misleading to a machine reader, and we treat it as such.

Collection design

Sample at a glance

These figures are the planned scope of the design below, not an achieved count. A records-per-edition figure is deliberately absent, and would have to be invented: the per-cell target is defined statistically rather than as a fixed number, and the number of platforms is not fixed in the design, so sample size is an outcome of collection rather than an input to it. The sample size is therefore reported as a result — the distinct listings tracked total listings in sample, the observations behind them total observations, and the realised cell sizes with their margin of error cell sample sizes and margin of error. Realised volume runs under any full-grid expectation, because cells that fall below the minimum are reported as insufficient sample rather than combined upward, and sponsored placements are excluded from the frame entirely.

Sources. Major general marketplaces, category-specialist marketplaces and quick-commerce platforms across three countries: India, the United States and the United Kingdom. Quick-commerce is included specifically because its availability data is location-bound and short-lived, which stresses dimensions 6 and 9 harder than any other channel.

Category sampling. A fixed set of categories chosen for contrast in attribute complexity: consumer electronics (high attribute count, strong GTIN culture), grocery and packaged food (high GTIN coverage, weak attribute structure), apparel (weak GTIN culture, variant-heavy), beauty and personal care, and home and kitchen. Each category has a published mandatory attribute set, defined before collection begins, used for dimension 4.

Sampling method. Within each platform × country × category cell, listings are drawn from category browse and search result pages, stratified to include both high-visibility and deep-catalogue listings, so the index is not merely a measure of how well the best-merchandised listings are maintained. Sponsored placements are excluded. The per-cell target is set so that a difference of a few percentage points between platforms is distinguishable from sampling noise at the cell level; the realised cell sizes and the resulting margin are published in the appendix. Cells that fall below the minimum are reported as insufficient sample rather than being combined upward.

Cadence and window. Each listing in the sample is observed on multiple occasions across the collection window. Repeat observation is required — not optional — because dimensions 1, 5, 6, 8 and 9 are all stability or consistency tests and cannot be scored from a single visit. The collection window and the as-of date are stated on the cover.

Rendering. Listings are collected as a browser would render them, so that client-side-injected price and availability are captured. Structured markup is captured separately from rendered content, which is what makes dimension 9 measurable at all.

Data schema

One row per listing per observation. Scoring runs over the observation set.

Report sections

Cover and as-of statement — title, collection window, as-of date, sample scale, one-line scope.

Executive summary — the headline: a large share of the sampled digital shelf is not agent-readable, with the dominant failure modes named. One page, no figure.

What this report measures, and what it does not — the does/does-not note, verbatim.

The readiness rubric — the full table, published for dispute.

Figure 1: Readiness bands across the sample — stacked bar. X: country. Y: share of listings (%). Series: the four readiness bands. Stacked because the bands are mutually exclusive and sum to 100% of listings — the shape of the whole is the finding.

Figure 2: Score distribution — histogram. X: readiness score (0–100, 5-point bins). Y: count of listings. A histogram because the question is whether the shelf is bimodal — a well-maintained head and a long unreadable tail — which a mean would hide.

Figure 3: Pass rate by rubric dimension — horizontal bars. X: pass rate (%). Y: the nine dimensions, sorted ascending. Items against one measure, so bars; sorted so the weakest dimension reads first.

Figure 4: Readiness by platform type — grouped bars. X: platform type (marketplace, quick-commerce). Y: pass rate (%). Series: the availability and consistency dimensions only. Isolates the channel effect on the two time-sensitive dimensions.

Figure 5: Identifier resolvability by category — bars. X: category. Y: share of listings with a valid, cross-resolvable external identifier. The clearest single illustration of why identity, not markup, is the first blocker.

Figure 6: Attribute completeness against the mandatory set — box plot. X: category. Y: attributes found as a share of the mandatory set. A box because the spread within a category is the finding, not the median.

Figure 7: Visible-versus-markup price disagreement — histogram. X: signed markup price delta as a share of rendered price. Y: count of listings. Shows both direction and magnitude of disagreement; a pass rate alone loses both.

Figure 8: Readiness over the collection window — line. X: observation date. Y: mean readiness score. Series: country. A level tracked over time, so a line. Establishes the quarterly baseline.

Failure mode profiles — four short written profiles of the most common failure patterns, each with an anonymised example structure and the fix.

What a brand can do about it — a prioritised remediation order derived from dimension weights and observed pass rates.

Methodology and limitations — in full.

Appendix: category mandatory attribute sets, cell sample sizes, margin of error, rubric change log.

Data slots

Methodology and limitations

The rubric is ours. Actowiz Solutions designed the nine dimensions and set the weights. They are not a standard and we do not present them as one. We publish the rubric in full, with the weights and the pass conditions, so that a reader who disagrees can re-weight it and re-derive the index. Any change to the rubric between quarterly editions is recorded in a change log so that period-on-period comparisons are auditable.

The readiness criteria are derived from the task, not from a vendor spec. Agentic commerce protocols are still consolidating. We therefore do not build this index on any single vendor's published requirements, we do not assert a fixed specification or version number, and we do not describe commercial arrangements between companies. Instead, each dimension is derived from what any agent demonstrably needs in order to answer a shopping question and complete a purchase: it must know which product this is (dimensions 1–3), what the product is (4), what it costs (5), whether it can be had here and now (6–7), where to buy it (8), and whether the machine-readable version of the page can be trusted (9). If the protocol landscape settles into something narrower or broader, the rubric will be revised in the open rather than retrofitted.

Public listings only. Everything here is observed from a public product listing as a browser renders it. Many retailers supply richer, cleaner data privately — through merchant feeds, partner APIs or direct integrations — and none of that is visible to us or counted here. A low score means the public listing is hard to read. It does not mean the retailer has no good data anywhere.

Poor score, strong sales. Readiness is not performance. A listing can fail most of this rubric and still be a category bestseller, because human shoppers tolerate ambiguity that machines do not. Readers should treat the index as a measure of exposure to a coming constraint, not as a diagnosis of current commercial health.

Agent behaviour is not measured. We do not run shopping agents against these listings, and we make no claim about which listings an agent would surface, prefer or purchase. Agent selection logic is proprietary, changing and outside the scope of this work.

Pass/fail scoring is deliberately blunt. Binary scoring loses nuance — a listing missing one of eleven mandatory attributes scores the same on dimension 4 as one missing all eleven. We accept that cost in exchange for reproducibility; the underlying per-field counts are in the schema for anyone who wants a graded version.

Sampling. The index describes the sampled cells, not the entire catalogue of any platform. Cells below the minimum sample are reported as insufficient rather than merged. Sponsored placements are excluded, which means the index reflects organic catalogue quality.

Rendering and timing. Price and availability are point-in-time. Quick-commerce availability in particular can change within minutes, so some measured visible/markup disagreement will reflect capture latency rather than a genuine data defect. The tolerance window used to separate the two is published in the appendix.