Turn Specs into Evidence: Tests, Comparisons, and Reproducible Methods

Consider a composite scene: a bike-computer brand shipping from Shenzhen to the US and Germany pastes its page beside three rivals into a chat assistant and asks how battery life compares. Each rival returns with a clause: brightness, GPS rate, temperature, units tested. Theirs gets half a sentence — “the brand claims up to 20 hours.” Two days later a return request arrives: the page says 20 hours, I get nine. The figure came from a bench run no rider could repeat.

A spec reports what you achieved; evidence reports under what conditions, by what method, and whether a stranger could rerun it — and a number without a method is read as a claim, then settled by the customer at home.

Specs are the vendor’s language, evidence is the buyer’s

A model comparing products looks for a shared coordinate system. Two numbers measured under different conditions leave it two choices: present both, or drop the one it cannot place. Side by side, your figure reads as claim against a rival’s measured result; dropped, you leave the answer. The gap is not the number; it is the protocol.

Buyers meet the same gap: someone in Ohio cannot lift the product, so page and reviews are the whole object. If the page gives outcomes only, one review line — “nowhere near what the listing says” — becomes the only third-party description, and assistants quote it first because someone else said it.

The usual counterexample bolds “20 hours,” turns it into a hero icon, and adds “all-day battery.” Same claim three times: no condition, no method, a string with no source. The page does not need a third number; it needs the first condition.

Four elements of an evidence sentence

Each element must survive being lifted off the page — the answer unit test on a number.

Element	Question answered	Passes	Fails
Sample	What was tested	“X1 Pro, three samples”	“this product”
Conditions	In what environment and workload	“25°C, screen on, one GPS fix per second, Bluetooth off”	“normal use”
Method	How the figure was produced	“recorded to auto-shutdown, median of three units”	“lab tested”
Boundary	Where it stops being true	“about 14 hours under the same load below 0°C”	left out

Order matters: lead with the conclusion and readers keep the number and lose the qualifier; extraction does the same.

The cheapest element is the boundary. Asked about downsides, assistants hunt for the limit; rivals list strengths. One honest limit buys credibility for the paragraph.

Conditions a stranger could rerun

Three things must be complete: environment (temperature, humidity, line voltage), workload (load, mode, frequency), and statistic (peak or median, unit count, before or after aging). Peak and typical belong in separate sentences; only the peak turns a best case into a promise.

Comparison tables earn trust from a shared protocol, not design. Give competitor figures a source and a date. Run every row under the same conditions. Write “not disclosed” where data is missing: an empty cell reads as zero, an estimate reads as fact. The commonest self-inflicted wound is a polished table setting your measured result beside a rival’s marketing number. A table that admits where a competitor wins gets cited more than one that wins everywhere — a test record, not an ad.

Example: one runtime claim, rewritten

Example — the model names and figures below are invented; no real product is described.

Before: “All-day 20-hour battery. Full GPS, zero range anxiety.”

After: “The X1 Pro ran at 25°C with the screen always on, one GPS fix per second, and Bluetooth off; three production units recorded until auto-shutdown, median runtime 19.5 hours (shortest 18.2, longest 20.4). Below 0°C under the same load, expect about 14 hours; with Bluetooth and phone notifications on, about 16.”

Three edits: a number becomes a median with a range, environment and workload attach, and a limit is stated. The headline falls from 20 to 19.5 — a concession that turns a claim into something verifiable. Nobody refuses a product that manages 14 hours in the cold; they feel misled when the page never said so.

Five steps from spec sheet to evidence sentence

Step	Action	Owner	Passes when
1 Pick the specs	Three specs buyers ask about most, from tickets and search	Content, support	Each maps to a real question
2 Find the record	Test reports, inspection samples, repair logs; if none, test a batch	QA	You can name units, rig, environment
3 Rewrite	One sentence each for sample, conditions, method, result, boundary	Content	It stands with the heading removed
4 Lock the protocol	Spec, protocol, source, owner, verified date	Content, engineering	EN and DE pages carry identical figures
5 Re-ask	Run the same prompts three times; note what gets quoted	Growth	A condition is quoted, not a slogan

Step 4 is the one teams skip, and it decays: after a supplier or firmware change, an old condition sentence becomes a fresh contradiction. Keep the sheet as text, not a PDF scan — evidence a machine can read separates a documented test from an invisible one. Step 5 needs sampling; an AI visibility baseline keeps the prompt set still.

Where this breaks

First, the spec is not the decision variable. On consumables, jewelry, and gifts, price, look, and reviews decide; conditions only lengthen the page.

Second, the data cannot be published. Supplier terms and custom specs are often under contract; “tested to this standard, report on request” beats an invented range.

Third, the field is too small. Marketplace bullets cannot hold a condition clause without pushing keywords out — full sentence in A+, short version in the listing.

Fourth, the spec moves. New material, firmware, or supplier makes it stale; contradiction damages more than omission — see the four trust prerequisites.

Fifth, regulated claims. Medical, children’s, food-contact, and environmental claims draw their own scrutiny; a protocol is not a compliance review, and a wrong sentence is worse than a missing one.

This piece belongs to the content and semantics hub. This month, take one SKU, copy its three most prominent numbers, and rewrite each with sample, conditions, statistic, and boundary. Replace the originals, then run the same prompts in two weeks and see which sentence the assistant quotes.