GEO-052·Issue 05·Strategy & Market·52 / 132
Which Categories Need the Most Evidence: Beauty, Electronics, Baby, Outdoor
Evidence budgets follow revenue, but risk decides citation. Score categories on failure cost, severity, trial dependence, and comparability, then fund the gap.
Issue focusEvidence of Trust: Reviews, UGC, and Third-Party Corpora
A four-line cross-border group selling into the US and Germany splits its Q2 evidence budget in April: cosmetics, charging accessories, baby monitors, hiking gear. Beauty earns most, so it takes the largest share. The monitor line defers paid safety testing. In June, asked whether the monitor can stay beside a crib all night, an assistant names two rivals. Both publish the encryption standard, the infrared switch delay, and the lab behind it. Yours says “peace of mind, engineered.”
Evidence money should follow the cost of a wrong answer, not revenue: the higher the consequence of a bad recommendation, the more the model quotes only what a stranger can re-check.
Order value is the wrong sorting key
“High ticket, more evidence” fails at both ends. A cheap video cable with a vague compatibility claim can black out a desk setup; a made-to-measure curtain costs more, yet the buyer only needs a measuring guide; a baby bib costs little and material safety is still a hard gate.
Low-risk categories get a recommendation and a buy link. High-risk ones become “if A, take this, but confirm B.” Conditions need quotable facts; without them the model recommends the best-documented brand, not the best.
Risk aversion follows: safety, health, compatibility, or after-sales liability make a wrong answer expensive, so the model hedges. “It depends” is not evasion; it is an empty slot only a standard number, a test condition, or a named responsible party fills, and a competitor’s number fills it first.
Ranking three to five candidates needs numbers in one unit, and messy conventions push the answer back to “well reviewed.” Whoever names the unit first gets quoted — see the four trust prerequisites.
Four questions, three levels
Three levels decide who goes first; five invent false precision.
| Question | High signal |
|---|---|
| Failure cost | A return means cross-border shipping or re-measuring a window |
| Consequence severity | A failure involves injury or legal responsibility |
| Trial dependence | Photos and copy cannot stand in for wear, swatch, or grip |
| Comparability | Listings disagree and reviews still ask how big it really is |
Severity carries the most weight: it decides whether the model says anything at all; the rest decide how specific it will be. A category low on all four earns little from more evidence.
Four categories, four jobs for evidence
| Category | Job of evidence | Fund first | Common misallocation |
|---|---|---|---|
| Baby | Entry ticket: no number, no shortlist | Standard number, test items, age range | Creator seeding |
| Electronics | Ranking input: missing data, missing rank | Test conditions, compatibility list, warranty scope | Unboxing videos |
| Beauty | Exclusion: buyers use it to subtract | Ingredient list, who should not use it, swatch conditions | Retouched imagery |
| Outdoor | Verification: first order on scenario, repeat on service | Field-test conditions, durability data, warranty execution | Spec dumps |
A baby product’s evidence is a ticket; without a standard number there is no shortlist to join, and a creator budget cannot repair that. Beauty evidence subtracts: “not for these skin types” outperforms “suitable for all.” Outdoor evidence must survive a follow-up asked at -10°C, not in a studio.
Evidence also does not travel: a baby test report is unreadable on an electronics page, and “tested at -10°C” on a baby page reads as a safety question. A site-wide template works for the category it was written for and dilutes the rest.
Example: one battery claim, two categories
Example — invented brands and specs. A team sells a baby monitor and an action camera; both pages say “8-hour battery.”
On the monitor page it carries nothing. Night buyers fear latency, dropouts, infrared switch speed, and encryption. Supply those and the model has a sentence for “can it be hijacked”; withhold them and the extract is “this brand also offers baby monitors.”
On the camera page battery is the variable, but only with conditions: 4K at 30fps, 25°C, stabilization off. Without them the model says “long battery life”; with them it answers whether the camera lasts a ski day.
Counter-example: a beauty brand spends on creator unboxing while the page has no full ingredient list and no “not for you if.” Asked about sensitive skin, the assistant quotes a competitor. The gap is not exposure; it is a re-checkable list and a stated boundary.
Four steps, then the template
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Score | High, medium, low on the four questions | Category, growth | One written reason per category; priority moves only when a level changes |
| 2 Inventory claims | “Military-grade,” “food-grade,” “industry-leading,” line by line | Content, compliance | Each line gains a condition, a standard number, or a link |
| 3 Fix five SKUs | The SKUs behind the five most-asked questions | Content, category | The same prompts, sampled twice, return conditions |
| 4 Move it | Specs, compatibility, and certification into tables and structured data | Engineering | Page, feed, and schema agree |
Step 3 is the one teams skip: rewriting the catalog feels thorough, but the return sits on the few SKUs people ask about. Write each repair as an answer unit, move spec fields into product schema, and sample with the AI visibility baseline method: same questions, same window, two rounds. Step 1 is a strategy and market call, not a content brief.
Where this breaks
First, low-risk commodity: cables, cases, jewelry, where price and look decide and evidence past a point earns little; spend on comparison pages instead.
Second, saturating categories: public specs and standardized units mean differentiation has moved to service; more parameters only lengthen the page.
Third, compliance red lines: efficacy in beauty, medical implication in baby, safety promises in electronics. A wrong sentence is worse than none, and rules still shift across markets as of 2026; route wording through legal.
Fourth, marketplace sales: platform AI reads listings, reviews, and fulfillment, so on-site evidence comes second.
Fifth, an unresolved entity or uncrawlable pages. Evidence with nowhere to attach is a PDF in a drawer; fix access before reordering categories.
This month, do one thing: score every category high, medium, or low on the four questions, take the widest gap, and rewrite five SKUs of unfalsifiable claims, each with a condition, a standard number, or a link. Ask the same questions in four weeks and see whether the answers quote numbers.