GEO-003·Issue 01·Content & Semantics·3 / 132
What a Product Looks Like Inside an AI Answer
An AI shopping answer is assembled from five fields, not one page. How entity, specs, price, stock, reviews, and citations decide if you stay shortlisted.
Issue focusThe Entry Point Shift: How AI Search Rewires Global E-commerce Traffic
A buyer in a 610-square-foot Austin apartment asks a chat assistant for a standing desk: shallow enough for a hallway nook, steady with two 27-inch monitors, under $500 delivered. Four brands come back with three specs each. The brand lands third, warranty line reading “not provided.” The warranty exists, in a FAQ image on their site. Nobody wrote anything false. The answer is still wrong.
A product inside an AI answer is not a page but an assembly of five fields — entity, attributes, price and stock, reviews, citations — and a missing layer drops you off the shortlist rather than lowering your rank.
Five layers, not one page
The instinct is to reread the page and decide it reads well. It does. That was never the question. “Which one should I buy” means rebuilding your product from facts that slide into a slot. Tone and brand narrative fit no slot.
| Layer | What it answers | If missing | What you control |
|---|---|---|---|
| Entity | Category, brand, competitor set | Filed under “other brands” | Title, category, naming consistency |
| Attributes | Size, load, compatibility | Reaches the shortlist, gets skipped | Spec tables, units, definitions |
| Price and stock | Price, availability, delivery | Absent at the quote stage | Feed, on-page price, shipping terms |
| Reviews | How it is used, what fails | Specs without trust | Review summaries, third-party testing |
| Citations | Where the claim comes from | Model credits someone else | Crawlable pages, off-site corpus |
The layers are sequential. A wrong entity voids the other four. Missing attributes make an accurate price unusable in comparison. Reviews and citations decide whether the model dares recommend you, not whether it sees you. Spend only on the last layer and you are doing PR for a subject that does not exist.
Why one missing layer removes you
The model’s process is dull: split the question into constraints, find entities that satisfy them, fill attribute slots, then let reviews and citations decide whether the result earns a sentence.
Constraints are hard. “Small apartment” maps to desk depth and minimum height; “under $500 delivered” maps to landed price. Where one has no single answer, the model does not guess — it swaps in a product that does. Substitution is cheap; guessing is often wrong.
Contradiction is worse than absence. 176 lb on your site, 220 lb on your marketplace listing, and the model likely uses neither, unable to establish which is true. Unverifiable and nonexistent behave the same way.
Counterexample: warranty, returns, and installation buried in a brand story about putting customers first. People parse that. Parsers extract no field, so the product is identified, then judged missing on warranty length.
Three failure patterns, three fixes
A page without a coherent entity. Brand A on the site, A-US in the marketplace, an old name in reviews. The model cannot reconcile three subjects. Fix naming and the external profile first, using the test in does your brand entity exist.
Specs that cannot be compared. Inches mixed with centimeters, load that never says whether it is static or mid-adjustment. Five standardized fields buyers filter on beat forty filled to the edges.
Reviews with nothing quotable. Hundreds of “great desk, fast shipping” carry no scenario, drawback, or comparison. The model lifts no supporting sentence and cites a third party. Change the questions you ask when requesting reviews.
Not a gap, but a misread: “strongest load in its class” is handled as a claim, and unverifiable superlatives invite doubt.
Which layer to fix first
Start from “can this be extracted,” not “is this well written.”
| Step | Action | Owner | Verify by |
|---|---|---|---|
| 1 | List the five fields buyers filter on | Content, category ops | 20 real questions; each has one clear answer |
| 2 | Write each as a standalone sentence and a spec table | Content | Cover the context; does it hold? |
| 3 | Kill conflicts across site, listing, feed | Content, marketplace ops | Same field, value, unit in all three |
| 4 | Add one honest “when not to buy this” line | Content | It answers a real exclusion question |
| 5 | Move facts out of images and JS-only modules | Engineering, content | Readable with JavaScript off; see the AI crawler map |
Skip step 1 and step 2 becomes a well-liked essay. Skip step 3 and every added detail multiplies contradictions. Skip step 5 and the first four steps exist only for humans.
Example: one standing desk, five layers
All names, specs, and prices are illustrative, not a real product.
Before: the entity resolves, the only attribute is “height adjustable,” and every review says “works great.” The model names the brand and leaves the spec row blank.
After, layer by layer:
- Entity: one brand name, category, and model series across site, storefront, reviews
- Attributes: 28–47 in height, 176 lb static, 132 lb moving, tops to 47 in
- Price and stock: $399 delivered, 3–5 days from a US warehouse
- Reviews: surface detailed ones — “two 27-inch monitors on a 47-inch top; slight wobble above 43 inches”
- Citations: “176 lb static, 132 lb in motion” as one sentence, reviewers given the same basis
The goal is not a guaranteed recommendation. It is that for “small apartment, under $500, two monitors,” the model has something to write and to cite. Being written in precedes being compared.
Where this breaks
Non-standard, experience-led categories. Custom furniture, art, anything needing a trial. Buyers will not hand that decision to an AI.
Revenue concentrated in one marketplace. If most orders come from a platform AI that reads only listings and reviews, fix those before the content layer.
The entity does not exist yet. If the name does not resolve and nothing off-site mentions it, entity work comes first; a spec sheet for an unrecognized subject is wasted.
The layers treated as a fill-in-the-blank form. Completed cell by cell, you get five layers with words in them and none that can be cited. Coverage is not the test; surviving without context is.
Price and stock are not a writing task. Content defines the basis; engineering connects the feed.
Why entry points matter first: how AI search split e-commerce traffic. Whether the fixes worked is a sampling question, not a screenshot: AI visibility baseline. This line continues on the content and semantics hub.
If you only make one change this month: take your best-selling SKU, run five real buyer questions through two assistants, and log every field that came back empty or wrong. Fix the most common one, and write it as a sentence that holds with no context around it.