GEO-039·Issue 04·Content & Semantics·39 / 132
Category Pages Aren't Product Lists: Organize by Question Chains
AI lifts model numbers from your category page but no reason to choose. Five blocks, a question-chain map, and the PDP boundary that keeps both citable.
Issue focusAnswer Units: Writing Category and Product Pages for GEO
Friday morning, a DTC portable-power brand selling into the US opens Perplexity with a customer’s overnight question: two nights of car camping, a camera, a drone, a 14-inch laptop — how much capacity? The answer cites two review sites, a Reddit thread, and three competitors. Engineering finds nothing broken: pages crawl, markup validates, stock syncs. The page is twelve models ranked by price and one line of copy, “Power for wherever you go next.” The model can lift every model number and no reason to prefer any.
A category page is not a shelf but a judgment: which one, for whom, compared with what — and that judgment, not the SKU grid, is the part AI quotes.
At the category level, AI wants criteria, not names
The model is shrinking a candidate set, and names are not the bottleneck. Model numbers, capacity, price, and stock reach it from feeds and price trackers, often fresher than your page. A category page that re-typesets the feed is low-density, however clean it looks.
What the buyer’s second question needs is what no feed holds: which variables change the choice, how each is measured, and when not to buy. Only someone who sells the category can write that; otherwise the model cites a guide that may not shortlist you. A founding year and “engineered with care” change no decision.
One category page, one main question chain
The same term hides different constraints. A solo hiker asks about weight and recharge time; a family asks whether one unit runs a cooler and a projector at once. Cover both and you answer neither one precisely.
Find the main chain in tickets, site search, and reviews: keep the “how do I choose” questions with a constraint, take the recurring one. “Which one is best” carries no constraint and cannot be the main chain.
Branches belong on stable subcategory or comparison pages, never faceted URLs, which are unbounded and duplicate each other. If the chain is unclear, start with the question-chain mapping method.
Five blocks, in an order that does not move
Boundary, criteria, routing, comparison, evidence. Each block leads with one sentence a model can lift.
| Block | Question | Facts required | Keep out |
|---|---|---|---|
| Category boundary | What this class solves, and what it doesn’t | Fit and non-fit cases | Founding story |
| Selection criteria | Which variables change the choice | Unit, measurement, threshold | Adjective stacking |
| Situation routing | Which tier fits my constraint | Constraint-to-tier map, with “don’t buy” | One SKU’s specs |
| Side-by-side comparison | Where tiers differ, what the price step buys | A few fields, measured alike | PDP paragraphs |
| Evidence and next step | Why believe it, how to verify | Test conditions, warranty, returns | “Market leader” claims |
Cap criteria at five; seven usually means the deciding variable is missing. Keep table and paragraph boundaries machine-visible, as in HTML semantics for extraction.
- Pick the chain. Tickets, site search, reviews into one main question, at most three branches. Owner: content. Passes when the constraint is in the buyer’s words.
- Write the five blocks in order, five criteria or fewer. Owner: content and category. Passes when a reader can stop after block three and still choose.
- One landing page per follow-up. Owner: content and SEO. Passes when no two pages answer the same question.
- Re-read with JavaScript off. Owner: growth. Passes when the first screen answers the main question.
Which page answers which question
One question, one home: does the statement hold for one SKU, for two, or for the class?
| Buyer’s question | Answered by | Category page’s job |
|---|---|---|
| Which class fits my situation | Blocks two and three | States the situation and the non-fit |
| How do these two differ | Comparison page | Links out, repeats no specs |
| Will it run on 230V | Compliance note plus PDP | States the boundary |
The spec table is where the two pages fight. Copy it, let a price drift, and the model may drop both. The category keeps comparison fields; the PDP keeps full parameters. Both need a clear answer unit, from different angles: “how to choose” versus “what this is.”
Example: a two-night camping question chain
Example — a hypothetical portable-power brand, no real company, no figures. Main question: powering a camera, a drone, a laptop for two nights of car camping.
The boundary block says this is camping power, not a home outage backup; a space heater belongs on another page. Criteria: capacity versus output wattage, devices running at once, recharge time, capacity lost to cold. Routing splits three ways — solo lightweight, family road trip, van conversion — naming the fitting tier and who should not buy the largest. Comparison puts three tiers on capacity, weight, ports, recharge time, and what the price step buys. Evidence covers cell chemistry, test conditions, certification, warranty, and EU service.
Three links: two subcategory pages, one comparison page. Facet URLs are noindexed. Acceptance is not a ranking: with JavaScript off, do the first 300 words and one table answer “how much capacity”? Judge citations with a repeatable rubric, not a screenshot.
Where this breaks
Fewer than ten SKUs, or made-to-order. The PDP plus one comparison page is enough. Revenue concentrated on a marketplace. In-platform AI reads listings, reviews, fulfillment; your site page does not move it. Fix the listing. Regulated categories — medical devices, child safety, anything financial. Give dimensions, not recommendations; add a consult-a-professional line and clear it with legal. Price and stock that move daily. Give a price band, not a hardcoded number, and point to the live PDP price. Buyers who only compare price. Decision content changes nothing.
The misuse worth naming: forty category pages from one template, five identical blocks, only the noun swapped. That is a duplicate set with better formatting. One question separates the two — does this page contain a constraint and a piece of evidence that exist only in this category?
Do one thing this month, take the highest-revenue second-level category, merge ninety days of tickets and site search into one constrained “how do I choose” chain, and write 600–800 words across the five blocks with three links to pages you already have. Then read the first screen with JavaScript off, as a model would. Prove one category before copying to the second. The rest of this line sits on the content and semantics hub.