GEO-049·Issue 05·Authority & Off-site·49 / 132
Four Trust Prerequisites for AI to Recommend You
AI recommends what it can verify. Four prerequisites — evidence, consistency, independence, freshness — a 40-minute self-check, and a repair order by cost.
Issue focusEvidence of Trust: Reviews, UGC, and Third-Party Corpora
A DTC robot-vacuum brand selling into the US and Germany watches a marketer run a prompt: “Best robot vacuum for a 750-square-foot apartment with hardwood floors and a long-haired cat?” Three brands come back. Theirs does not. The page has a solid rating, a two-year warranty, and the certification numbers a German buyer checks. The diagnosis is “more reviews,” so the team orders five creator videos — the one fix that addresses none of the failures.
AI recommends the brand whose claims survive four machine-checkable prerequisites — evidence, consistency, independence, freshness — and because they are serial, one failure discounts the other three.
Trust is four checks, not a tone of voice
A model has no “liking” step: naming a brand takes on the buyer’s risk of choosing wrong, so confidence attaches only to what can be verified.
| Prerequisite | What is checked | The common failure |
|---|---|---|
| Evidence | Condition, unit, method someone could repeat | “Aerospace-grade cells” |
| Consistency | Site, listing, reviews, knowledge panel agree | Two-year warranty on site, one year on the listing |
| Independence | Who says this on a domain you do not own | Every asset sits on your domain |
| Freshness | Which year this number is from | A 2024 price table live in 2026 |
The hard part is reconciliation, not writing: budgets go to writing, so brands look strongest on evidence and weakest on consistency. E-E-A-T is a quality-rater concept, not a ranking factor; these four rows are its machine-executable form.
Evidence and consistency: reproducible first, uncontradicted second
The evidence threshold: a third party could redo the sentence and get the same result. A usable claim has a fixed skeleton — object, condition, number, unit. “At 70°F, on hardwood, standard mode, it ran 118 minutes on one charge” is repeatable; “all-day power” is zero information. With nothing to extract, the model falls back on the safest true line: the brand also offers products in this category.
Consistency beats richness. Two versions of one entity do not make the model choose; they lower confidence in the entity. Price, warranty, certification IDs, and model aliases are the hot spots; master data that is not unified contradicts the best-written page at the next feed sync.
The test is crude: cover the logo, read your page, the listing, and one independent review side by side. Where they disagree, your credibility is the lowest, not the average.
Independence and freshness: someone else says it, in the present tense
First-party pages are an application; third parties are the endorsement. A model cannot measure sincerity, but it can see how evidence is distributed. Everything under your domain is one source and the risk. Media, labs, communities, and marketplace reviews cover different parts of the question; a portfolio beats one viral post.
The shortcut is manufacturing a third party — a new domain, reviews under different bylines. But domain registration, business entity, payment account, and image licensing leave relationships; once linked, independence is the first prerequisite to collapse and the hardest to rebuild.
Freshness is cheaper to fix and more lethal to skip: an expired certificate, a discontinued model, a changed price, a shortened return window all reset the first two. In the agentic layer it becomes a transaction problem — stock, delivery days, and return terms must be right now — and it fails silently in the price and availability markup.
A forty-minute self-check, then repair in cost order
Score each row zero or one; every row spans two teams.
| Prerequisite | Check | Passes when | Owner |
|---|---|---|---|
| Evidence | Cover the brand, read five core pages | Three condition-and-number sentences per page | Content |
| Consistency | One SKU across site, listing, two off-site sources | Model number, warranty, certification ID, price definition match | Ops + merchandising |
| Independence | List sources by type, deduplicated | Three source types, not all on your domain | Marketing |
| Freshness | Page dates, certification expiry, price, stock | Key numbers dated; expired items removed | Merchandising + engineering |
Zero or one: fix visibility before outreach. Two: repair the cheapest failing row. Three: shortlist-ready. Four: make the definitions a recurring audit, not memory. The score sequences work; as a target, it produces evidence that passes by one point.
Fix in cost order, not table order.
- Reconcile. Grid model, warranty, certification ID, and price across four channels; keep one definition and name who changes what. A blank cell is an undecided question.
- Rewrite. Five core pages get two or three condition-and-number sentences; delete adjectives without a subject.
- Date it. Stamp key numbers, pull expired certifications and discontinued models, reconnect price and stock to master data.
- Reach out. Add the missing source type — one media review, one lab test, one community thread — before chasing volume.
Acceptance is counted, not ranked: empty cells, a page that reads with the logo covered, numbers untouched for twelve months, source types.
Example: one product, four versions of the truth
Example — a fictional cleaning-appliance brand; names and figures are invented.
The site says 180 minutes of runtime, the marketplace listing says 150, and a 2024 review still says 120. Warranty: 24 months on the site, 12 on the listing. Asked three times, the assistant omits the brand twice and ranks it last once, noting “runtime figures vary between sources.”
Fix the master data first: one runtime definition, one warranty term, published test conditions, a date on every number. A week later the brand may still be absent, but “figures vary” should be the first sentence to disappear — a result an AI visibility baseline can hold you to.
Where this breaks
Low-consideration categories. Phone cases, cables, storage boxes: AI rarely compares brands at that level, and evidence work returns less than crawlability and listings.
A new product with no off-site mentions. Independence cannot go zero to one overnight. Build first-party evidence and crawlable pages first; buying reviews to pass the row sets all four to zero.
Marketplace-only revenue. Platform AI reads listings, reviews, and fulfillment; your site’s evidence still counts, but it is not the deciding input.
Regulated categories. Medical devices, children’s safety, food-contact materials: compliance conclusions cannot be replaced by freshness or independence, and legal review outranks all four rows.
These four rows are the foundation of authority and off-site work and open the Evidence of Trust issue. If your pages still describe instead of prove, start with the answer unit framework.
This month, do one thing: for your best-selling SKU, build a four-column grid for runtime, price, warranty, and certification across your site, the marketplace listing, and two off-site sources. Mark every disagreeing cell. Fix numbers and dates first, sentences second, and date each figure you touched.