Intro
A five-star resort is described as four-star. A 78-room property is described as a 120-room hotel. A signature chef left two years ago and AI engines still cite his menu.
These are not edge cases. They are what a brand fact accuracy AI search audit surfaces on almost every premium hospitality account Capston reviews. AI engines reuse the highest-authority source they can find, and “highest authority” is rarely the brand’s own canonical page.
This page explains what a fact accuracy audit covers, the four-status taxonomy used to grade each claim, and the propagation chain that fixes have to travel before AI engines update their answers.
Why AI engines repeat incorrect brand facts
AI engines do not check facts against the brand. They generate the most probable answer given the corpus they were trained on plus the live sources they retrieve at answer time.
That has three structural consequences for premium brands.
First, stale beats current when the stale version has more inbound links and longer indexing history. A 2019 trade press feature outranks a 2025 brand page in the retrieval layer, so the 2019 numbers win.
Second, aggregators outweigh owners. OTA listings, travel guides, and review platforms each republish a slightly different fact set. AI engines blend them, and the blended version drifts from what the brand actually offers.
Third, silence is interpreted. When the brand-owned page does not state a fact explicitly (rating, ownership, capacity, policy), the engine fills the gap with whatever the wider web says — which may be wrong.
This is why fact accuracy is one of the eight dimensions in the AI visibility scoring system. Without it, the rest of the score sits on shifting ground.
What to audit
A brand fact accuracy audit grades AI answers claim by claim. For premium hospitality, the recurring claim categories are:
- Rating — star rating, classification body, year awarded.
- Capacity — room count, suite count, villa count, restaurant covers, meeting capacity.
- Location — address, district, distance to landmarks, transport context.
- Signature experiences — spa rituals, on-property activities, seasonal programs, partnerships.
- Policies — family policy, pet policy, accessibility, dress code, cancellation terms.
- Dining concepts — restaurant names, chef names, cuisine type, awards, opening hours.
- Awards — Michelin, Forbes, Condé Nast, World’s 50 Best, year of recognition.
- Ownership — operator, owner group, brand affiliation, management changes.
Each AI answer is parsed into individual factual claims. Each claim is then matched against the canonical source — the brand’s own authoritative page, document, or registry entry. This is the same evidence trail captured by the AI answer evidence layer.
The four-status taxonomy
Every claim is flagged with one of four statuses. The taxonomy is deliberately narrow so the work after the audit is unambiguous.
- Confirmed — the AI answer states the fact, and it matches the canonical source. No action.
- Partial — the AI answer is directionally right but incomplete or imprecise (right rating, wrong year; right room count, wrong suite split). Action: enrich the canonical page so the precise version is the most retrievable one.
- Wrong — the AI answer states something that contradicts the canonical source. Action: locate which corpus source is feeding the error, correct it where editable, dispute it where not, and reinforce the correct version upstream.
- Missing — the AI answer omits a fact that the brand considers material (a signature experience, an award, a policy). Action: make the fact explicit on the canonical page and propagate it through the chain below.
The four-status grid produces a flat list of editable items. That list is what drives the fix queue. It is also the structure the Capston Hospitality Scorecard uses to weight fact accuracy in the final score.
The propagation chain
A fix on the brand website does not, by itself, change AI answers. AI engines retrieve from many sources, and the new version has to travel through the chain before the answer set stabilizes.
The chain has four stages.
- Brand-owned canonical pages — the property site, the brand site, the press kit, the fact sheet. This is the source-of-truth layer. Every audited claim must have one canonical home here, with a stable URL.
- Wikidata entity — the structured entity record that AI engines lean on heavily for unambiguous facts (location, classification, parent organization, founding date). The Wikidata item is updated with sources pointing back to the canonical pages.
- Trade press updates — recent, dated, authoritative articles that restate the corrected fact in context. AI engines weight recency, so refreshed coverage moves the average.
- OTA and aggregator profiles — booking platforms, travel guides, association directories. Each is updated with the corrected fact and, where possible, a citation to the canonical page.
The order matters. Updating an OTA before the canonical page is set is how brands accidentally create new conflicts. Canonical first, entity second, press third, aggregators fourth.
How long fixes take
Fact propagation is slower than most marketing teams expect. The realistic horizon is two to three quarters from canonical fix to majority-stable answers.
The drivers of that timeline:
- Recrawl cadence — AI engines re-index priority sources on weeks-to-months cycles, not days.
- Model retraining vs. retrieval — facts retrieved at answer time can shift within weeks; facts baked into the model itself only shift on the next training cycle.
- Aggregator review queues — OTA and directory edits often go through manual review, which adds weeks per source.
- Competing signals — a single well-linked legacy article can keep the wrong version alive for months even after every editable source is corrected.
Brands that audit quarterly and fix in the right order see the curve flatten by quarter two and stabilize by quarter three. Brands that fix only the website see drift come back within a single retrieval window.
How this fits into Capston Core
Brand fact accuracy is one of the eight scoring dimensions in AI visibility scoring. The audit runs on every brand engagement that follows the Capston Core methodology, and the fix queue is tracked in the AI answer evidence layer so before/after states are traceable per claim.
→ Back to Capston Core
FAQ
How many claims does a typical audit produce?
For a premium hospitality property, between 60 and 200 distinct claims across the eight categories, depending on portfolio scope and answer volume.
Which AI engines are checked?
The same engine set used in the overall scoring — ChatGPT, Perplexity, Google AI Overviews, Gemini by default, plus market-specific engines where relevant.
What if the canonical page does not exist yet?
Creating the canonical version is the first remediation step. Without a source of truth on the brand’s own domain, the rest of the chain has nothing to anchor on.
Can Wikidata be edited directly by the brand?
Yes, with sourced citations. Capston supports clients in framing the edit so it is accepted by the Wikidata community and survives review.
Final CTA block
Find out which facts AI engines get wrong about your brand.
Audit your brand facts
Read the methodology