When There Are No Reviews, What Makes AI Trust Your Product?

Friday afternoon, a support inbox at a US direct-to-consumer power brand gets a two-line email from a Munich shopper: “Has anyone actually tested this brand?” The SKU launched on the German storefront three weeks earlier with zero reviews. That week the growth lead runs a prompt through a chat assistant — “portable power station for camping that I can fly with” — and gets four familiar names. The brand’s page is among the retrieved sources, never in the answer.

A new SKU does not lose to better reviews; it loses to missing checkable facts — AI decides whether to trust an unknown product from a number, a test condition, a named party who is liable, and whether all three agree everywhere.

Zero reviews is an information problem first

When an assistant answers “is this worth buying”, it works in two passes: it narrows the candidates to products it can describe with specifics, then checks that those specifics corroborate. A zero-review SKU drops out at the first pass — not because it is bad, but because there is nothing quotable to say.

First-party evidence does not replace word of mouth. It supplies the layer underneath: sentences a model can extract, identifiers it can verify, a legal entity to hold responsible. The E-E-A-T framing in Google’s rater guidelines asks the same question: who makes the claim, and what backs it?

The counter-example is everywhere: “EU certified” in white on red, no certificate number, no covered models, no issuing body. To a model that weighs the same as “industry-leading” and “crafted with care” — decoration. A shopper reads it the same way; the gap surfaces only when something goes wrong.

Four kinds of evidence, sorted by how hard they are to fake

Evidence	What a model can extract	Passes when	Fails when
Test reports	Lab, standard, sample model, conditions, result	Number matches the SKU on sale	A lab photo, no clue what was tested
Certifications	Mark, certificate number, covered models, market	Number traceable, models match, market and validity	“Certified in multiple countries”
Warranty and returns	Term, scope, liable party, cross-border process	Term per market, repair address, who pays shipping	“Global warranty”
Performance claims	A figure and its measurement basis	Method, load, ambient temperature stated	“10 hours of runtime”, no conditions

They are not equal in weight. Certification answers “may you sell this here”, a test report “what can it do”, warranty “who pays when it fails”, performance “where are the limits”. Qualification and liability are hardest to fake, so cold-start pages lead with them; performance comes last, because the next revision invalidates it fastest.

Verifiability is three conditions

Number, conditions, consistency. Remove one and the evidence falls back into copywriting.

Number means a traceable identifier: certificate number, report number, standard, clause link. Conditions means the setup behind the figure: sample model, method, load, ambient temperature, market. Consistency means the same fact appears in the same form on the product page, manual, packaging, support script, and dealer listings.

Teams skip consistency, and it costs the most. When a model finds two runtime figures for one SKU, it rarely picks the better one; it cites neither. Contradiction reads as risk, and risk is easier to route around than resolve. Tying evidence to the product entity — model, variant, market — is what makes it usable downstream; the Product Graph shows how those relationships fit together.

Never let the only copy of a report live in a scanned PDF: image text is invisible to retrieval. Put the summary on the page; attach the PDF as a file.

A four-step playbook

Step	Action	Owner	Passes when
1 Inventory	List reports, certificates, warranty, test data; model and market per row	Quality and compliance	Every row maps to a live SKU; gaps left blank
2 Add conditions	Fill in number, issuing body, standard, test setup, covered models	Quality and content	Who issued it, which model, under what conditions
3 Write the citable paragraph	60 words below the buy box: qualification, capability, liability, public summary link	Content	Claim readable without the PDF
4 Sweep	Spot-check 10 SKUs against page, manual, packaging, support script	Operations	Differences zero, or documented with a reason

Step one should not start with copy. Lay the evidence out first, and some SKUs will have nothing to show — a finding for the range review, not a gap for content to paper over. Turning facts into buying reasons is its own craft; see turning attributes into buying reasons.

Example: three adjectives replaced by three facts

Example (hypothetical, no real brand, no measured outcome): a 500 Wh power station launching in Germany with zero reviews, its page saying “EU certified”, “long warranty”, “high capacity”. The rewrite: “CE declaration of conformity covering models X1 and X1 Pro, document number on the support page”; “24-month warranty in Germany, repair address in Hamburg, return shipping paid by the brand”; “512 Wh rated capacity, tested at 25 °C and −10 °C, method linked”. Qualification, then liability, then capability — the order matters more than the word count.

Where this breaks

First, a product with a structural flaw. Evidence makes a problem clearer: a report stating its conditions can just as easily show your runtime collapsing in the cold. Precondition: a product that survives being described precisely.

Second, sensory categories. Fragrance, fit, fabric hand-feel: first-party evidence says what the thing is, not whether it suits you. Third-party experience still carries that.

Third, categories that close inside a marketplace: where the platform reads listings, reviews, and fulfillment metrics, off-site evidence carries little weight. Fix listing fields first.

Fourth, regulated categories. Medical devices, children’s products, food-contact materials: the required evidence is set by law, so clear compliance before rewriting a sentence.

Fifth, an active trust crisis. Recall, chargebacks, a regulator’s letter — handle the incident; an evidence page will not rescue it.

Sixth, the commonest misuse: evidence that only certifies itself. Adjectives stamped by the brand, no number, no issuer, nothing to cross-check. A model treats that as marketing, not evidence. The bar is “someone else could verify this”, not “we said it sincerely”.

To see whether it worked, watch which sources get cited in the same prompt set, not where you rank; the authority hub collects this line, and an AI visibility baseline covers measurement.

This month: pick ten SKUs with zero or single-digit reviews, put their reports, certificates, and warranty terms on one table, and write each a sentence with a number, a model, and a condition — below the buy box, with a public summary link. Any SKU you cannot write it for gets labelled “evidence gap”, not padded with adjectives.