Product Copy for Agents: Short, Precise, Comparable

Composite example. A pet-hydration brand runs a dry run with an agent: a Chicago buyer asks which cat fountain keeps two cats drinking over a long weekend. Two units come back with a tank volume and a runtime; the third — the brand’s best seller — comes back as “tank capacity not provided.” The page opens with “Big tank, calm cat, clean water, always,” capacity in a photo caption.

An agent compresses each product into one comparison-table row: the spec sentence, then the exclusion, then the brand line; invert that order and the prettier the copy, the less of it gets used.

What it wants is a row, not a paragraph

An agent filters, compares, orders. It reads the feed first — price, stock, dimensions — returns to the page for short spans that stand alone, then reads the description to explain why this one. The first two decide whether you can be compared; the third only decides wording. Fetch or parse failure ends it earlier — see agentic commerce readiness and the product feed as landing page.

Short, because five products in one comparison leave little room each, and a sliced paragraph loses its subject first: “It supports fast charging” can attach to the cable above it. One sentence, one fact.

Comparable, because the output is a table and every cell needs the same footing. Write “large tank” while a rival writes “3.2 L,” and your cell reads “not provided.” The buyer sees the blank.

The usual failure is structural: brand story first, specs in an image three screens down. Extraction reads text, the figures vanish, and the cell falls back to “no comparable specs.”

Three blocks, three jobs

The spec sentence decides whether you can be compared, the boundary whether you should be excluded, the brand why you.

Block	What goes in	Passes when	Fails as
Spec	Who, value, unit, basis	Headline hidden, still a cell	“Huge tank”
Boundary	Where it does not fit; which spec is worse	Lets the agent exclude	Strengths only
Brand	Who it serves, what it replaces	One sentence, no repeated specs	“Committed to pets worldwide”

Order follows retrieval: what and how much, when not to buy, why believe it. One sentence per block, under thirty words; longer, and the model quotes the block whole, cutting the boundary.

Price, shipping, stock and warranty stay out: policy fields, validated against live systems and the returns page, see transaction trust signals.

Delete these words:

Don’t write	The agent stores	Write instead
Huge tank	no field	3.2 L tank volume
Fast charging	no basis	0–80% in 30 min (25 °C, stock adapter)
Aerospace-grade aluminum	unverifiable	6061 aluminum, 1.5 mm wall
From, up to, approx.	no range	floor and ceiling, or median + sample

Adjectives become fields, superlatives need a source or nothing, fuzzy ranges become a method. Empty the sentence and there was no fact.

Example: one fountain, three sentences

Example — invented brand, model and figures.

Before: “The Acme P2 smart pet fountain pairs an extra-large tank with a silent pump and intelligent dry-run protection.” The agent stored: tank capacity not provided.

After, as three paragraphs:

“The Acme P2 cat fountain holds 3.2 L, about six days for two adult cats drinking 500 ml a day, and measures 28 dB(A) at one meter in a room with a 25 dB(A) floor.”

“The pump cannot be taken apart; the unit is rinse-only. Households with four or more cats should look at the larger P3.”

“Acme makes cat drinking fountains only. The P2 is the entry model and shares filters and pump with the P3.”

Capacity gains a value and a basis; “silent” becomes decibels and distance; the pump limit is stated, not hidden. Model, capacity and compatibility stay untouched — the fields agents merge entities on.

Sales blocks the boundary sentence, fearing it talks buyers out. In an agent channel it does the reverse: asked about downsides, the agent hunts it, and competitors rarely have one.

Five steps from field list to comparable sentence

Step	Action	Owner	Passes when
1 Pick fields	30 days of questions, tickets and returns → 5–7 compared fields	Content, category	Traces to a question
2 Source values	Master data, test reports, policy pages; “not tested” if none	Category, QA	Traces to a record
3 Draft blocks	Spec, boundary, brand; one sentence each	Content	All three answer, heading hidden
4 Extraction test	Three sentences alone in two assistants: what, why, who not for	Content	Answered, no “claims”
5 Ship and re-ask	Replace the first PDP block; re-ask in two weeks	Content, growth	Quotes a value or a limit

Step 2 is the one copywriting cannot rescue. Invent a number and step 4 exposes it: the model repeats the fiction, and the buyer finds out on return.

Step 4 is the only falsifiable step. No answer means rewrite; answers that drift mean the referent is loose, so replace “it” with the model name.

Step 5: one screenshot is not evidence, and a changed answer is not a changed order; agent sessions need their own conversion and refund view, as in agent traffic measurement.

Localize the units, not the facts

The same machine shows 3.2 L in English, 3200 ml in German. The volumes agree, but conversion is not guaranteed, so the model may file two specs; “super large capacity” keeps both blank.

Change	Do not change
L/ml, lb/kg, °F/°C, plug and voltage	model name, values, materials, test conditions
Local certification and warranty wording	the certification, the promise
Category words local buyers use	the primary scenario and the exclusion
Currency, shipping threshold, return window	the conclusion and the downside

Hard rule: one master record for every version’s numbers. When they contradict, the model is wrong whichever it picks.

Where this breaks

First, the page is not fetchable or rendered; copy with no reader cannot be compared.

Second, the data does not exist; do not invent a range, source the test first.

Third, high-touch and made-to-measure categories — the agent only pre-screens, so “fits everyone” is not a conclusion.

Fourth, marketplace fields are locked: long version on your PDP, short in the listing, same entity terms.

Fifth, handing the blocks to an agency as a fill-in form: every cell has words, none is comparable.

This piece sits on the content and semantics line: comparable first, then quotable.

This month, tag each sentence of your top ten PDP first paragraphs spec, boundary, brand or adjective. Delete the adjectives, source missing boundaries from support and returns, then re-ask the same agents and see whether “not provided” is gone.