Evidence discipline
Provenance & Confidence Standard
How BackTier tags every factual claim in its deliverables: what is verified, where it came from, how confident it is, when it was last checked, and which causal claims the evidence does not support.
Purpose
BackTier's credibility depends on the same discipline it asks AI systems to apply to everyone else: distinguish what is verified from what is inferred, disclose where a claim came from, and never assert causation an evidence base does not support. This standard governs how facts are tagged, sourced, and dated across every BackTier deliverable — audits, research reports, AI-answer testing logs, case studies, vertical research, and any sales or positioning claim about market conditions.
This is not a formatting nicety. A visibility-infrastructure company that publishes unlabeled inference as fact is exactly the kind of weak-evidence source BackTier tells clients AI systems learn to discount. The standard exists so BackTier's own output survives the scrutiny it recommends clients apply to their competitors.
The standard applies to AI Visibility Audits, AI-answer testing logs, BackTier research reports, case studies, vertical research and targeting documents, competitive positioning claims, and any internal document whose conclusions will surface in a client-facing asset. It does not apply to framework definitions, methodology descriptions, or BackTier's own stated positioning — assertions of what BackTier does, not factual claims about the world.
The tag set
Six tags, used inline and attached to the specific sentence or data point they govern — never applied blanket to a paragraph.
FACT
A specific, checkable claim about the external world: a citation observed in a model's answer, a competitor's page content, a schema type present on a site, a statistic published by a named source, a date, a dollar figure.
A FACT tag means: if someone checked this today, they would find it as stated, subject to the LAST VERIFIED date attached to it.
SOURCE
Where the FACT came from: a named source with a URL or document and the date accessed, or, for BackTier-run testing, direct observation with the test conditions stated.
No SOURCE tag, no FACT tag. An unsourced factual claim is treated as an inference at best and is rewritten or removed before publication.
CONFIDENCE
High, Medium, or Low, applied to every FACT and every INFERENCE.
High: independently corroborated by two or more sources, or a BackTier-run test with a documented, reproducible method. Medium: a single credible, named source, or an inference that follows directly and narrowly from High-confidence facts. Low: vendor-supplied data, a single anecdotal report, an inference resting on indirect evidence, or anything time-sensitive that has not been re-verified.
A claim with no confidence tag is not ready to publish.
INFERENCE
A conclusion BackTier draws from one or more FACTs. It is stated separately from the facts it rests on, never blended into the same sentence as if it were itself observed.
Format: state the inference, then reference which facts it derives from.
TEST
Specific to AI-answer testing. Captures the prompt tested, the platform, the date tested, location assumptions, and the raw model output, kept separate from any interpretation of that output.
A TEST entry is the observation; conclusions drawn from it are INFERENCE entries that cite the TEST.
LAST VERIFIED
The date a FACT was last confirmed true. Required on every FACT classified as ephemeral.
A FACT past its re-verification window without a current LAST VERIFIED date is stale: it must be re-checked or downgraded to Low confidence before reuse.
Durable vs. ephemeral facts
Every FACT is classified as durable or ephemeral. Durable facts are unlikely to change on any timeline relevant to the document: definitions, historical dates, methodology descriptions, structural claims about how a model type works in general, publicly documented company history. Durable facts still carry SOURCE and CONFIDENCE but do not require a re-verification cadence beyond normal periodic review.
Ephemeral facts decay on a timescale that matters to the reader: AI-answer outputs, search or answer-engine rankings, competitor citation presence, pricing, feature availability, structured-data support status, a vendor's claimed capabilities, current market-size figures.
Every ephemeral FACT requires a LAST VERIFIED date and an explicit re-verification window stated in the document. The defaults: AI-answer testing is re-verified within 30 days before reuse in a client deliverable; competitive and market figures within 90 days; anything published as part of a dated research report is understood to be a snapshot as of that date and does not need re-verification unless republished.
A document mixing durable and ephemeral facts says so plainly — the reader needs to know which claims are still true today and which were true as of a specific date.
Unproven causality
A specific flag, separate from the confidence scale, required whenever a claim — BackTier's own or a vendor's — implies that one action causes a change in AI-answer behavior: for example, that adding schema increases citation likelihood, or that a specific change caused a client's inclusion rate to rise.
BackTier does not currently have controlled, causally isolated evidence for these relationships. Visibility work is multivariate, model behavior is not fully observable, and before-and-after changes are confounded by everything else that changed in the interim. Any claim of this shape is flagged as unproven causality and rewritten to state correlation or observed sequence, not causation.
For example, rather than asserting that adding FAQPage schema got a client cited, the standard requires: this client added FAQPage schema in March; citation frequency in the tested prompt set increased between the February and April testing windows; the two are correlated in this observation; no controlled test isolates schema as the cause.
This rule applies with extra weight to vendor claims BackTier repeats or cites. A vendor's own causal claim about its product does not become a BackTier fact by repetition: it is flagged and attributed to the vendor as their claim, not BackTier's finding. This is the mechanical enforcement, at the sentence level, of BackTier's standing prohibition on guaranteed rankings and manufactured causation.
How this maps onto BackTier deliverables
AI-answer testing logs: each captured field — prompt, platform, date, location, entities included or omitted, citations, source quality, competitor implications — is a TEST or FACT entry; every interpretive line is an INFERENCE; any claim that a specific input caused the observed answer-layer outcome is flagged as unproven causality.
Audit outputs: findings sections are FACT, SOURCE, and CONFIDENCE entries; priority-fix and phased plan sections are inference-driven recommendations built on those facts, and say so.
Research reports: the standing structure already separates observation from interpretation at the section level. This standard adds sentence-level tagging inside those sections and requires the source ledger to double as the SOURCE registry — every FACT in the report is traceable to an entry in the ledger.
Case studies: the starting problem, diagnostic findings, and work performed are FACT entries; the observed outcome is a FACT with a LAST VERIFIED date as of publication, or an INFERENCE where the metric was not precisely measured. BackTier does not imply guaranteed causation between work performed and outcome observed — that relationship is treated as unproven causality by default unless a controlled comparison genuinely exists.
Vertical research and targeting: economics claims such as transaction value, market fragmentation, and buyer willingness to pay are FACT, SOURCE, and CONFIDENCE entries; the conclusion about whether a vertical is worth pursuing is an INFERENCE explicitly built from the facts above it.
Working rule for drafting
When producing any governed deliverable: write the FACT with its SOURCE first, tag CONFIDENCE, classify durable or ephemeral and date it if ephemeral, then write the INFERENCE separately underneath. If a claim implies causation, stop and check whether controlled evidence actually exists — if not, flag it and rewrite to correlation or sequence language before it goes further.
Client-facing final copy renders this discipline in BackTier's normal paragraph-driven prose, without visible bracket notation. The tags are a drafting and quality-assurance discipline, not reader-facing formatting — but the underlying claim-by-claim record of what is known, how well, since when, and what is merely suggested is reconstructable from the draft at any point before publication.