GEO-098·Issue 09·Technical Foundation·98 / 132
The Product Feed Is the New Landing Page
Agents filter feed records before they ever open your page. A field-priority checklist, comparable titles and attributes, and the errors that drop SKUs.
Issue focusReady to Transact: Agentic Commerce and Product Feeds
On a Thursday, the ecommerce lead at a DTC air purifier brand sees a split screen. Paid impressions slide, organic sessions hold flat, and the SKU that carries the quarter never appears when an assistant is asked for “a purifier for a 450 sq ft bedroom, under 30 dB, with a filter under $60 a year.” Five answers come back, none of them hers. The page is fine: CADR table, noise test, filter cost, returns. The fault is in a file nobody opened this month — the feed row still says out of stock after the restock, the hero carries a “Spring Sale” banner burned in, the title reads AirPure Multi-Function Home Air Purifier (Best Seller), and room coverage is free text: “large rooms.”
An agent’s first gate is not your landing page but your feed row: it filters fields, keeps what passes, then opens a page — a flawless page behind a failed field is traffic you never receive.
A page persuades; a feed survives filtering
A landing page assumes a human arrived; a feed assumes a machine must pull fifteen records from fifty thousand.
Three layers choose. Filtering is attribute matching, not semantic guessing: “under 30 dB” compares against a number, and if your noise field is empty because the page says “whisper quiet,” the row fails; the model has no duty to infer. Comparison follows: shortlisted rows sit side by side, and only structured values fill cells; synonyms in a title render as blanks. Verification is last: an agent needs availability now, a price with a stated tax basis, a delivery window, and a return term, each in its own field.
The KPI is filterable, comparable, verifiable, not complete; a feed written as a small essay scores zero on all three.
Field priority: a missing P0 is not a listing
Rank fields by what breaks when absent, not by spec-sheet order.
| Priority | Fields | If missing | Acceptance test |
|---|---|---|---|
| P0 | Unique ID, price and currency, availability, crawlable URL, hero image | Never shortlisted, or suppressed | Matches live page; URL returns 200 |
| P1 | Brand, GTIN, category, three to five filter attributes | Listed, then loses every sort | Values from one master record |
| P2 | Shipping cost and speed, return window, variants | Agent cannot promise fulfillment, picks a rival | Destination-market basis |
| P3 | Certifications, bundles, accessories, review summary | Costs upsell and trust, not entry | Verifiable, with a source |
P0 is the ticket, and usually the field updated by hand every Friday while teams polish P3.
P1 is the sieve. Buyers exclude on three to five attributes — room coverage, noise, filter cost, voltage, compatible model — which need their own fields, not words inside an infographic.
P2 is the last mile: shipping and returns written as a paragraph become unpredictable in a table. The per-market arithmetic has its own model in localized product feeds.
Titles, attributes, and images as comparable values
A feed title is an index, not copy: brand + category noun + the differentiating value + the setting, in the words buyers filter with.
Counter-example: 2026 New AirPure Multi-Function Air Purifier, Home Air Cleaner, Quiet, Large Room, Best Seller. Asked for 30 dB or 450 sq ft, that row has no number to match. Better: AirPure AP-450 Air Purifier, 450 sq ft, 24 dB, HEPA H13.
Attributes take one value. If the feed says HEPA and the page says HEPA plus a pre-filter, an agent can pick either and be wrong; being wrong once reads as a bait-and-switch.
Images need a clean hero on plain background, no burned-in text, and a stable URL. A signed CDN link that expires returns 403, which means no image. Variant images must map to variant attributes, or the agent cannot tell what it shows.
Example: one row, before and after
Example — brand, prices, and specifications are invented; no real product is described.
Before: AirPure 002-A Multi-Function Air Purifier (Best Seller); availability out of stock, untouched for nine days; price 199, no currency; a “Spring Sale” banner in the hero; brand, GTIN, coverage, noise, and filter cost blank; a campaign redirect for the link.
After: AirPure AP-450 Air Purifier, 450 sq ft, 24 dB, HEPA H13; availability synced from the warehouse; 229.00 USD, tax basis stated and matching the page; coverage 450 sq ft; noise 24 dB; filter 59 USD a year; voltage 120 V; shipping free, 2–4 business days, US warehouse; returns 30 days; clean hero and unsigned URLs.
Nothing in the second record is an adjective; every value can be filtered, compared, checked.
Five steps, and the rejections that keep coming back
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Mine the queries | Turn 30 days of chat, site search, and ads into five filter attributes | Growth, category | Each attribute has a feed field |
| 2 Fix P0 | Availability, price, URL, hero image reconciled to the site | Engineering, ops | 50 sampled rows, zero disagreements |
| 3 Rewrite titles | Top 50 SKUs by revenue: brand, category, value, setting | Content | Values match the filter fields |
| 4 Validate | Validator plus your own rules: units, currency, variants | Engineering | Zero rejects; every warning closed |
| 5 Test backwards | Ask ten constrained questions; check whether you appear | Content | Shortlisted in six of ten |
Only step 5 is falsifiable; a validator proves a field exists, not that it is enough.
The rejections repeat: a price that disagrees with the page, availability stale after a restock, promo text burned into the hero, variants collapsed into one row so “do you have it in white” is unanswerable, a blank required attribute, a link that redirects to the homepage, inches and centimeters in one catalog, no local-language copy. None is an optimization; each is a repair.
Where this breaks
First, the feed does not replace the page. Once a row passes, the answer unit still has to be written.
Second, bespoke categories resist fields. Custom furniture, made-to-measure apparel, anything needing a survey: the feed covers only part of the screening.
Third, feeds lag. Price and availability rarely sync exactly, and promotion weeks are when they disagree most; put acceptable staleness in an SLA.
Fourth, more fields is not better. A wide record with contradictions loses to a narrow one that is right.
Fifth, agent-traffic attribution is unsettled, so do not credit a feed relaunch with an order count. Start from AI referral attribution.
This month, do one thing: export the feed rows for twenty best-selling SKUs into a sheet with four columns — availability, price, hero image URL, three filter attributes — ask an assistant one constrained question per attribute, and watch which rows it keeps. The layer above is agentic commerce readiness, which opens this issue; the rest of the lane is on the technical foundation hub.