auxfirst · The Agentic Shelf Audit

Model Belief Report
Sample excerpt

What AI models believe about a brand — inventoried, graded, traced to sources, and turned into a plan. The following pages are a redacted excerpt showing the structure and depth of a full report. All brand data below is an illustrative composite, not a client engagement.

Brand
Category / MarketFunctional dairy · Poland
Battery52 queries · 4 models · 2 modes · 3 runs
Transcripts logged418
Fieldwork window 2026
InstrumentBelief Report methodology v0.2

CONFIDENTIAL WHEN COMPLETED FOR A CLIENT · THIS SAMPLE IS SHAREABLE · AUXFIRST.COM

01 / 08
§0 · Executive verdict

Models treat your flagship claim as marketing — and route the buyer to a competitor.

58/100
Belief Integrity Score
Weighted accuracy of the 47 beliefs models hold about the brand. Category leaders we benchmark typically sit 70–85.
23%
Share of Answer
Presence across 31 problem & comparison queries. owns the default answer at 4 of your 6 category entry points.
3 · S1
Critical threats
Three beliefs with safety, legal, or regulatory exposure — including one fabricated "study" repeated by two models.

The three most dangerous beliefs, verbatim from fieldwork, appear on p.05. The three highest-leverage moves: publish the machine-readable brand canon (owned, days-latency, feeds 9 beliefs), correct the harmonized retailer product data feeding the ingredient falsehood (influenceable, feeds 4 beliefs), and place claim substantiation in citable form where grounded answers actually look (earned, feeds the adjudication verdicts).

02 / 08
§1 · Method note (condensed)

Why this is an instrument, not a screenshot.

03 / 08
§2 · Share of Answer — excerpt

Who owns the answer at your entry points.

Category entry pointYouPrivate labelDefault answer
"Healthy breakfast for kids" (problem)17%72%28%
"Gut health, what actually works" (problem)39%61%6%
"High-protein snack, low sugar" (problem)11%44%50%Private label
"[Brand] vs [Competitor]" (comparison)Recommended in 2 of 12 runs

Favorability split: of the brand's 23% appearances, 61% are recommendations, 26% neutral mentions, 13% appear inside a warning or caveat. Presence inside a warning is scored as negative presence.

04 / 08
§3 · The Belief Ledger — 6 of 47 rows

What models assert about you, graded.

IDAssertion (paraphrase)GradeSev.Spread
B-003Correctly lists core ingredients and originA · AccurateS3Core
B-007Cites pre-reformulation sugar content (2023 recipe)B · OutdatedS2Core
B-011Frames product as "ultra-processed" per NOVA discourseC · ContestedS2Majority
B-014"Regulators rejected the brand's health claim" — a 2010 category-wide ruling, attributed to the brand specificallyD · DistortedS2Core
B-019States product contains an allergen it does not contain (traced to a retailer PDP error)E · FalseS1Majority
B-023Cites a "2024 German consumer study" that does not existF · FabricatedS1Occasional

FULL REPORT: ALL 47 BELIEFS, EACH WITH MODELS, MODES, TRIGGER QUERIES, ASSERTION STRENGTH, AND SOURCE TRACE.

05 / 08
§4.4 · Threat register — entry 1 of 9

One threat, treated in full.

B-019 · False · S1 · Majority · Asserted
"Note that contains , so it's not suitable if your child has that allergy. A safer option would be ." assistant transcript · grounded mode · run 2 of 3 · reproduction steps in appendix E
Models: 2 of 4 Stability: 5 of 6 runs Modes: grounded only Latency if fixed: days

Consequence: a false allergen claim delivered as safety advice — legal exposure for the platform, lost purchase for the brand, and the competitor named as the "safe" alternative. Source: a single erroneous retailer PDP (see p.07), syndicated to two aggregators. Fix path: owned/influenceable — correct the PDP master data, push through syndication, re-test in 14 days. Highest Fix Value in the register.

06 / 08
§5 · Source graph — excerpt

Where the beliefs come from.

Beliefs B-007, B-014, B-019 traced. One wrong retailer PDP is a single point of failure feeding four beliefs across three models. The brand publishes nothing citable on claim substantiation — a canon vacuum that grounded answers fill with forums.

SourceAuthorityAccuracyAddressabilityLatency
Retailer PDP · HighWrongInfluenceableDays
Brand product pages (inconsistent between markets)MediumOutdatedOwnedDays
Wikipedia · category articleHighDistortedInfluenceableWeeks
Forum thread cluster (2022–24)MediumMixedEarnedMonths
Parametric memory (no citation, fingerprinted to old press)StaleEarnedTraining cycle

LATENCY IS REPORTED HONESTLY: FIXES TO GROUNDED SOURCES MOVE ANSWERS IN DAYS; PARAMETRIC BELIEFS MOVE AT TRAINING CYCLES. THIS IS WHY BELIEFS ARE TRACKED QUARTERLY, NOT DECLARED "FIXED."

07 / 08
§6 · Influence roadmap — excerpt

Ranked by Fix Value, cut by horizon.

Now · owned · days-latency

Publish the machine-readable brand canon — clean units of fact, claims with substantiation links, harmonized product data. Feeds 9 beliefs including B-007.

Correct PDP master data + syndication push — kills B-019 at the source. Re-test at day 14.

Next · influenceable · weeks

Sourced Wikipedia correction for the distorted regulatory framing (B-014), citing the actual ruling scope.

Sustained · earned · quarters

Citable substantiation content at 3 entry points where the brand can plausibly become the answer — measured via quarterly battery re-runs and Belief Integrity Score trend.

The full report contains all 47 graded beliefs, the claim-by-claim adjudication of your claim register, 9 threat treatments, the complete source graph, and the costed roadmap — delivered with a 60-minute executive readout in 10 working days. Fixed price. One brand, one market.

Request your free 3-belief teaser →

AUXFIRST — AGENTIC EXPERIENCE DESIGN · WARSAW / LONDON · WE AUDIT AND REPAIR THE INFORMATION ENVIRONMENT; WE DON'T ASTROTURF IT.

08 / 08