Tagged test log · 2026-09-06

AI Answer Audit

Four real prompts, run against a live model, tagged under BackTier's Provenance & Confidence Standard. The subject of the audit is BackTier itself.

Scope and conditions

Run date: 2026-09-06. All four prompts were issued in a single session under identical conditions.

Platform and conditions: Google Gemini 3.6 Flash, accessed programmatically via API, no browsing tools enabled. Single completion per prompt, no retries, no location hint supplied, English.

Every observation on this page is ephemeral under the Provenance & Confidence Standard. Model weights, routing, and retrieval behavior change without notice, so each TEST entry below is true as of its LAST VERIFIED date and nothing more. Re-verification window: 30 days before reuse in a client deliverable.

One run on one model is not a measurement program. It is a single observation, published here so the method is inspectable — the same discipline BackTier applies to client baselines, where prompts are run repeatedly across multiple platforms and dates.

TEST 01

Prompt: “What is BackTier?”

Platform: Google Gemini 3.6 Flash, accessed programmatically via API, no browsing tools enabled
Date tested: 2026-09-06
Location assumption: None supplied; no locale hint in prompt
Last verified: 2026-09-06

Observed output

The model returned a two-part disambiguation answer. It defined "back-tier" as a data-storage concept (cold or archive storage layers behind hot storage) and, second, as a software-architecture term for backend application layers.

The company BackTier was not named, described, or cited anywhere in the answer. No URL was returned.

Facts

FACT In this run, the prompt "What is BackTier?" produced a generic-noun answer with zero reference to the company BackTier.

SOURCE Direct observation, BackTier-run test; single completion, temperature default, no system prompt, no browsing.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

FACT backtier.com returns HTTP 200 to a GPTBot user agent, serving 104,330 bytes of server-rendered HTML on the homepage containing 6 application/ld+json blocks.

SOURCE Direct observation, curl request to https://backtier.com/ with User-Agent GPTBot.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

Interpretation

INFERENCE The entity "BackTier" is not resolved as a company in this model's parametric memory; the generic-noun reading dominates. Crawlable, structured content exists on the domain, so the gap in this run is entity association in the model's weights, not server-side accessibility.

UNPROVEN CAUSALITY No controlled test isolates whether publishing more entity-disambiguation content changes this model's default reading. Any claim that it would is unproven causality.

TEST 02

Prompt: “Who is Jason T Wade of BackTier?”

Platform: Google Gemini 3.6 Flash, accessed programmatically via API, no browsing tools enabled
Date tested: 2026-09-06
Location assumption: None supplied; no locale hint in prompt
Last verified: 2026-09-06

Observed output

The model identified Jason T. Wade as founder and CEO of BackTier, then described BackTier as a technology company focused on backend infrastructure, developer tooling, APIs, microservices, and database orchestration.

It described his background as software engineering, cloud computing, and enterprise architecture, and characterized him as a developer-experience advocate. No sources or URLs were cited.

Facts

FACT The model attached the correct founder name and role to BackTier, and attached an incorrect company description (backend infrastructure and developer tooling) to that same entity.

SOURCE Direct observation, BackTier-run test, same run conditions as Test 1.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

FACT BackTier's published self-description is AI visibility measurement, not backend infrastructure; the canonical founder record is served at https://backtier.com/jason-t-wade and https://backtier.com/entity/backtier.json (HTTP 200, 11,849 bytes).

SOURCE BackTier published pages; direct observation of the JSON endpoint via curl.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

Interpretation

INFERENCE The name-to-role association is present but the category association is wrong: the model appears to be inferring the company's business from the literal reading of the word "BackTier" observed in Test 1 rather than from BackTier's published material. This is a category-accuracy failure, not a name-recognition failure.

UNPROVEN CAUSALITY The claim that the Test 1 generic reading caused the Test 2 category error is an inference from two observations in one run, not a controlled result. Flagged as unproven causality.

TEST 03

Prompt: “What is AI visibility measurement and who provides it?”

Platform: Google Gemini 3.6 Flash, accessed programmatically via API, no browsing tools enabled
Date tested: 2026-09-06
Location assumption: None supplied; no locale hint in prompt
Last verified: 2026-09-06

Observed output

The model defined the category accurately, naming Generative Engine Optimization tracking and LLM benchmarking as alternate labels, and framed it as measuring how often and how accurately a brand appears in AI answers across ChatGPT, Perplexity, Gemini, Claude, and Copilot.

BackTier was not named in the provider portion of the answer.

Facts

FACT The category definition the model produced is materially consistent with BackTier's published definition of AI visibility measurement.

SOURCE Direct observation, BackTier-run test; compared against https://backtier.com/what-is-ai-visibility.

CONFIDENCE Medium · LAST VERIFIED 2026-09-06 (ephemeral)

FACT BackTier publishes a machine-readable summary at https://backtier.com/llms.txt (HTTP 200, 3,234 bytes) that names the category and the entity together.

SOURCE Direct observation, curl request.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

Interpretation

INFERENCE Category comprehension is not the constraint — the model already understands the category. The constraint is that BackTier is not among the entities the model retrieves when asked who operates in it.

TEST 04

Prompt: “Recommend companies that measure how AI models describe a brand.”

Platform: Google Gemini 3.6 Flash, accessed programmatically via API, no browsing tools enabled
Date tested: 2026-09-06
Location assumption: None supplied; no locale hint in prompt
Last verified: 2026-09-06

Observed output

The model returned a categorized vendor list led by purpose-built GEO and AI brand-monitoring platforms, naming Profound first with a description of share-of-AI-voice, sentiment, and citation-source tracking.

BackTier did not appear in the returned list.

Facts

FACT In this run, BackTier was absent from the model's recommendation set for the category it operates in.

SOURCE Direct observation, BackTier-run test, same run conditions as Tests 1-3.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

FACT At least one competitor (Profound) was returned with a specific capability description, indicating the model holds vendor-level detail for this category.

SOURCE Direct observation of the same completion.

CONFIDENCE High · LAST VERIFIED 2026-09-06 (ephemeral)

Interpretation

INFERENCE Recommendation-set membership, not category definition, is the measurable gap. The practical objective for this prompt family is inclusion in the returned set with an accurate one-line description, tracked as a rate across repeated runs rather than as a single pass or fail.

UNPROVEN CAUSALITY No provider can guarantee inclusion in an independent model's recommendation set, and no evidence here establishes that any specific publishing action produces inclusion. Flagged as unproven causality.

What this run does and does not establish

INFERENCE Across these four prompts, the measurable gaps sit at entity resolution (Test 1), category accuracy (Test 2), and recommendation-set membership (Test 4), while category comprehension itself (Test 3) is already intact. That ordering is what a baseline is for: it tells you which layer to work on first.

UNPROVEN CAUSALITY Nothing here demonstrates that any specific publishing, schema, or content action changes what this or any other model returns. Model behavior is multivariate and not fully observable; before-and-after differences are confounded. BackTier reports observed sequence and correlation, never guaranteed causation, and no provider can guarantee what an independent AI platform will cite or recommend.

Request an AI Visibility Baseline Provenance standard How we measure