GEO-113·Issue 10·Content & Semantics·113 / 132
Brand Narrative in AI Answers: A Consistency Check
AI assembles your brand from your site, listings, reviews, and press, then averages the conflicts. A seven-field card, a four-step audit, and a fix order.
Issue focusMeasurable: AI Visibility, Attribution, and Experiments
In September, the founder of a US direct-to-consumer e-bike brand read an AI answer about her company: “an entry-level commuter brand — adequate range, budget pricing.” Nobody had written that. The site says “city commuter, folds in 15 seconds”; the listing says “all-terrain fat tires”; one review calls the bike “great for short errands,” another “struggles on hills.” Every source is defensible alone. Assembled, they describe a stranger.
AI does not quote the sentence you want remembered; it averages your site, listings, press, and reviews into a lowest common denominator — so consistency is a data problem about seven fields, not a copy problem.
Why a model averages you
A model describing a brand merges before it writes. Eight disagreeing sources leave three safe moves: repeat the most frequent claim, keep the most conservative claim, retreat to the broadest category word. Each shaves off the difference.
Numbers degrade fastest. A site claim, a marketplace attribute, and a test figure become a range plus “varies with conditions.” Not false, but it is where you surrender the conditions under which the bike performs.
Off-site material carries weight in that merge: a review reads as more checkable than an adjective on your About page, because the writer had no reason to flatter you. You control only whether the facts inside it agree, which is why the work starts in the brand entity graph: name, category, origin, and scale aligned before any description is rewritten.
Seven fields, not one paragraph
Consistency lives at the field level, not the sentence level. Each field takes one value across every source; wording, language, and units can change, the fact cannot.
| Field | One-value rule | Typical conflict |
|---|---|---|
| Category | What you are and are not | Site says commuter, listing says all-terrain |
| Buyer and setting | Who, where, what annoyance it removes | First-time commuters here, hardcore riders there |
| Differentiator | A claim someone could dispute | “Rides better” everywhere, testable nowhere |
| Price band | One range, one currency, tax stated | $1,499 on site, $1,299 on the marketplace |
| Key specs | Same test conditions, same units | Three range figures, no protocol |
| Origin and scale | Production site, team size, founding year | “Since 2016” vs. “a decade of experience” |
| Warranty and returns | Years, coverage, return shipping | Two years on site, one on the listing |
Fields one and seven get skipped and matter most. Category decides which candidate set you land in; warranty decides whether you get recommended to a warranty-sensitive buyer. Get those wrong and the other five sharpen a picture of someone else.
The audit: four steps, one afternoon
An audit is transcription, not a rewrite.
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Sample | Homepage and top three PDPs per locale, three marketplace listings, 12 months of reviews and press | Content, growth | Every source has a URL and a date |
| 2 Copy verbatim | Only the sentence describing the brand or product — no paraphrase | Content | Each line points back to its source |
| 3 Fill the card | One cell per field; flag conflicts in red with both sources | Content, category | One value per field, or a marked conflict |
| 4 Rank | Sort by times cited × how checkable the claim is | Growth | A fix list, highest impact first |
Step two stops “I think they wrote something like that”: memory smooths conflicts, transcription does not. Sample as an AI visibility baseline would.
Example: one range figure, three numbers
Example — brand, specs, and prices are invented; no real product described.
Before: the site claims “55 miles, eco mode, 77°F, no cargo”; the listing says “75-mile range”; a reviewer measured “about 37 miles” on a hilly route with a loaded rack. The answer: “roughly 50 miles, fine for short trips.”
Only the site number carries conditions; the other two have no protocol, so the model treats them as equally credible.
Define one repeatable figure — test standard, assist level, weight, temperature, terrain — and use it in the PDP spec table, marketplace attributes, and feed. Pull the unsupported 75 miles rather than editing the reviewer’s 37 up to 55. Offer that reviewer a spec sheet as a citation, not a request to change the verdict; a reversal on request reads as less credible.
Afterward the answer tends to converge: “rated 55 miles in eco mode with no cargo at 77°F; one test measured about 37, with load and hills reducing it.” The brand is not flattered, but the buyer can see which number applies to their commute.
Fix order: cited sources first
Sequence matters: rewrite the homepage first and the model still cites the old listing and review, so you add a fourth version instead of removing three.
- Structured fields first. Feed values, marketplace attributes, PDP spec tables — each change removes one contradiction, and machines read these directly as machine-readable evidence.
- Then high-frequency third parties. Your lever is material, not editorial control: a spec sheet, a correction note, quotable numbers.
- Then the site narrative and its locales. Slowest and least risky; the facts are already fixed.
- Then re-test. Same prompt set before and after, looking for convergence on the same fields, not a ranking. Convergence is a signal, not a guarantee.
Where this breaks
First, one paragraph everywhere is not consistency. Homepage, category, and PDP need their own conditions; a copied block ignores context and gets cited less.
Second, a repositioning leaves old corpus behind. Separate the eras with “since 2024” rather than claiming “first” for both.
Third, platform fields are constrained and third parties are not yours. You can correct your own fields, not delete their sentence.
Fourth, a brand with no off-site mentions has a coverage problem first. A tidy card only documents that nobody has spoken yet.
Fifth, the worst fix is consistent and wrong. Keep a flagged conflict with its protocol attached rather than unifying on a flattering number.
Where the buyer asks changes which sources get read, so this connects to the four AI search entry points. Browse the line on the content and semantics hub.
This week, pick three sources — homepage, one marketplace listing, one third-party review — copy the sentence each uses to describe you into the seven-field card, and fix only the highest-weight conflict. Ask again in two weeks.