GEO-038·Issue 04·Content & Semantics·38 / 132
The PDP Fold: Which 80 Words Does AI Extract?
Most of what an assistant quotes from a product page comes from the first screen. Three extraction failures, a five-slot order, and a no-JS test.
Issue focusAnswer Units: Writing Category and Product Pages for GEO
Tuesday morning, a DTC stroller brand selling into the US and Germany fields a support ticket: does the stroller fit in an overhead bin? The assistant’s answer names a rival’s model, its folded size, and three airlines’ cabin limits. Their brand appears in one clause — available in four colors. The page opens with a full-width photo, “Travel light” across it, and folded dimensions on screen seven, inside an image.
The first screen is not a poster; it is the paragraph an assistant quotes — what the product is, who it serves, and what makes it different, before anything else loads.
The fold is an answer slot, not a poster
Extraction is chunking, not reading and then summarizing. A retrieval system strips navigation, promo bars, and recommendation rails, then splits what remains by heading and paragraph. The first content block becomes the first chunk, and it fixes the model’s reading of what the product is, who it is for, and what it is compared against. This is the answer unit framework applied to a product page: claim, evidence, difference, next step, compressed.
The fold’s job is not time on page; it is being quoted correctly. A hero that yields no standalone fact is worth nothing at an AI entry point.
The counter-example is everywhere: a full-bleed photo with “New” and “Free shipping” in the corner, alt text product-image-01. Text baked into an image is information to a human and blank to an extractor. A plain paragraph with three spec rows gets cited.
Eighty words is not a truncation limit any platform published. It is a working budget for one answer paragraph — what it is, who it is for, why it differs — about 500 characters. Run long and two ideas compete for one quote; the second loses.
Three places the quote stops
| Failure | What happens | Fix |
|---|---|---|
| Position | Template order sets the candidate text; blocks below the gallery rank later | Put one self-contained answer paragraph first |
| Boundary | A heading split stops merging; age range and weight limit in adjacent paragraphs return only one | Keep a comparison fact and its qualifier together |
| Form | Text in images, hover layers, collapsed details, script-injected specs return nothing | Render specs as HTML; move image facts to captions |
Rendering comes first: if specs arrive by JavaScript, no rewording helps — see whether an assistant can render your store. The rest is writing.
Boundary is the failure teams miss: the sentence is correct, just split in the wrong place. Form is cheapest to audit — a fact inside a picture does not exist. HTML semantics and paragraph boundaries cover both.
Order beats wording
A model restates in the order it reads, and a buyer rules out in the same order: category, hard numbers, terms, limits, proof.
| Slot | What goes there | Passes when | Common failure |
|---|---|---|---|
| 1 | Position sentence: category, scene, differentiator | Reads without the photo | “Redefining travel” |
| 2 | Three to five hard specs, with units | Comparable across tabs | “Ultra-light” |
| 3 | Price, stock, shipping, returns | Matches structured data | Price in script; returns in an FAQ |
| 4 | Compatibility and limits | Names what it does not fit | Benefits only |
| 5 | Verifiable evidence | Test, certification, or warranty on the page | “Industry-leading” |
Slot 4 is the one cut in review, on the theory that limits cost conversions. It is the question a buyer asks next in a chat box, and what makes the paragraph read like a spec sheet instead of ad copy.
A four-step playbook
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 | Audit: fetch without JavaScript, strip tags, read the first 300 words | Engineering, content | You can count standalone sentences |
| 2 | Reorder: rewrite the position sentence and spec line to the five slots | Content | Names category and scene, no adjective stack |
| 3 | Add limits: exclusions, compatibility, box contents | Category operations | At least one explicit “not for” |
| 4 | Test: five real buyer questions against that text alone | Growth | No wrong values; cited sentence logged |
Step 3 is skipped most often, and the result reads like every competitor’s paragraph.
Step 4 measures comprehension, not ranking. curl does not execute JavaScript, so its output is close to the text layer an extractor sees. Paste that text into a model and ask what this is, who it is for, and three key specs; wrong answers point to paragraph boundaries or wording, not the model. Re-run five real buyer questions and log the sentence cited; citation share is a separate job — score with an answer quality rubric before a model grades its own output.
Example: from a slogan to a paragraph
Example — hypothetical brand.
Before, the fold held three words and two badges: “Travel light”, “2026 New”, “Free shipping”. Folded size sat on screen seven, inside a photo.
After, the first paragraph reads:
“Folding travel stroller, 23 × 18 × 9 in (58 × 45 × 23 cm) folded, with a cabin-bag size chart for three airlines on this page; for children from 6 months to 4 years and up to 48 lb (22 kg); folds one-handed and stands on its own. 15 lb (6.8 kg) without the rain cover or cup holder. Not for infants under 6 months; twins need the double model.”
One quotable paragraph, or three clean facts: folded size, age range, weight. A shopper decides fit in one read; a model restating them cannot swap the weight limit with the age range. The photo stays — it just stops carrying the data.
Where this breaks
Visual categories. Fashion, jewelry, gifts, and home decor sell on the image first; a spec block in place of the hero costs conversions. Keep the photo; put the paragraph beneath it.
Marketplace listings. On Amazon the fold belongs to the template; bullets and A+ modules do not move, so fix title, attributes, and reviews instead.
Configurable products. Custom furniture has no single price before configuration; state the range and starting price, not a price that does not exist.
Established model queries. Buyers arriving with a model number need compatibility, warranty, and delivery first, not a brand introduction.
Regulated categories. Supplements, baby products, and medical devices put limits and warnings above benefits; that outranks any extraction gain.
The content and semantics hub covers this line.
This month, pull the ten SKUs that carry the most revenue, disable JavaScript, and paste the first 300 words of each into one sheet. Mark every SKU without a sentence answering what it is, who it is for, and why it differs. Rewrite those, add three spec rows, then ask a model five real buyer questions — and fix one wrong value at a time.