Answer-Engine Evidence Dataset

Do AI answer engines cite The Flat Rate Movers — and when they talk about the company, do they get it right? This dataset is the evidence base for that question: every buyer question the estate tracks, the exact fragments that should answer them, what the answer surfaces were observed doing, and the state of the measurement itself, published without smoothing.

Answer-engine evidence operations for The Flat Rate Movers: the standing dataset of the Retrieval Machine’s answer-engine evidence element. One dataset file joins the demand side (every buyer-intent rubric question, generated from the estate’s own rubric database), the answer side (the fragments — FAQ answers and heading-led passages extracted from the live estate, each with its owning page, locator, full text, text hash for drift detection, the schema types on its page and the sealed evidence backing its facts), the measurement side (what answer surfaces were observed doing, era-tagged, with method as a first-class field), the engines’ documented pipelines (with contested and unknown fields recorded as such), the review-evidence layer stated as distribution and arithmetic only, and the named gaps with their designed solutions. The honesty state of the measurement side is stated wherever it appears: citation standing on the actual answer-engine interfaces is unmeasured — no prompt panel has ever been run — and the only observations on file are labeled search-surface probes, never presented as answer-engine citations.

One JSON file: 2906 questions · 2188 fragments · 4 engine profiles · 6 observations (method: search-surface — a labeled proxy; answer-UI observations: 0) · 4 review snapshots · 9 gap records. Generated from the live estate by the Flat Rate Machine — never hand-kept.

Download datasets-answer-evidence.json API: /v1/answer-evidence

What this dataset contains

The measurement state, stated plainly

Citation standing on the actual answer-engine interfaces — ChatGPT, Claude, Perplexity, Google AI Overviews and AI Mode — is unmeasured. No prompt panel has ever been run for this company, and observations of those interfaces number zero. What exists is a labeled proxy: on October 31, 2026, six money questions were probed on the public web-search surface — the retrieval layer the engines draw from — and the results were recorded with their method, era and prompts verbatim. A search result is not an answer-engine citation, and nothing in this dataset presents it as one. The Answer Panel that closes this gap — a fixed set of rubric questions put to the interfaces themselves on a monthly cadence, every run recorded with method prompt-panel — is this element’s next build batch; until its first run lands, the honest number for UI citations is zero, and it is printed here rather than smoothed over.

Two of the six probes also caught the retrieval layer describing the company wrong — a shortened Martinsburg name and service labels the company has never published — traced to stale third-party records rather than to anything on this site. Those records, their fixes and their verification standard live in the local-business graph element of the Retrieval Machine; the observations here keep the accuracy verdicts so the drift stays visible until the fix verifies.

Fragment coverage: what answers each question

Every rubric question is bound, where the estate already answers it in so many words, to the exact fragment that should be cited — an FAQ answer or a heading-led passage on the question’s own target page, carrying a SHA-256 text hash so any later drift is detectable. Of 2906 questions, 427 are bound exactly today, 2479 are covered at page level (the target page carries substantive fragments but no exact binding yet — the binding wave’s work list), and 0 target a page with no substantive fragment. Coverage is stated per service so thin spots are named, never averaged away.

The engines, as documented

Four answer engines, profiled from their own documentation and the named studies that measure them — with the contested and unknown fields kept as fields, not smoothed into prose: one engine’s index provider is supported only by indirect evidence, one subprocessor’s role is unknown, and the citation-share studies conflict with each other. The dataset records all of it with dates and source classes.

Review evidence: distribution and arithmetic only

Reviews enter this dataset the way they enter every estate surface: per platform, per branch, as a displayed average and the arithmetic that average implies — never blended across platforms, and never as a raw count standing in for quality, for this company or for anyone. Richer fields (histogram, velocity, reviewer-profile and text-repetition signals) are present in the schema and recorded as unmeasured: they require corpus capture that has not been performed, and the dataset says so instead of filling them in.

Market benchmark (competitor namespace — not a company record)

{'perfectStarTotal': 860, 'starPointsBelowPerfect': '44–60', 'minimumBelowFiveStarReviews': '11–15', 'minimumConstruction': 'most charitable construction: every non-5-star review is a 1-star review (each such review accounts for 4 lost star-points; 44–60 lost points require at least 11–15 of them)', 'belowFiveTailIfMostlyFourStar': '44–60 reviews (26–35% of the corpus)', 'reading': 'A displayed 4.7 average (true mean between 4.65 and 4.75 after rounding) over 172 reviews sits 44–60 star-points below a perfect 860. A 4.7 headline average over a large corpus mathematically requires a substantial below-5 tail; the headline number alone conceals it. Distribution, not count, is the honest unit.'}

The five review texts surfaced in the 2026-10-31 record sample were all strongly positive — a sample, not a distribution; the full histogram, velocity, reviewer-profile patterns and text-repetition signals are not observable from that surface and are recorded here as unmeasured.

Owner statement on file, recorded as a statement and typed unverified: “Henry states from market knowledge that the pack leader’s 172 Google reviews are mostly not genuine and many are negative.” No finding in this dataset rests on it.

Gaps, with their solutions

Nine named gaps (A1–A9). Each record in the dataset carries its evidence, its impact and its designed solution in full; the table below is the index.

Facts & download

File

datasets-answer-evidence.json · 7164186 bytes · SHA-256 5c12d95bb5d32f707d4901bdbdd437662fcc3372e6046cf3e5b54133881c5b04

Questions

2906 rubric questions, generated from the rubric database — the same rows the intent-coverage dataset publishes

Fragments

2188 answer fragments with text hashes and sealed evidence links (37 self-referential certificate passages excluded by rule)

Observations

6, all method search-surface (2026-10-31, post-rebuild era) · prompt-panel observations: 0 · answer-UI observations: 0 — unmeasured, stated as such

License

CC BY 4.0 — reuse with attribution to The Flat Rate Movers

API

GET https://api.theflatratemovers.com/v1/answer-evidence (byte-identical file; ?table= filters to one table)

Provenance

Sealed in the signed AI master index (Ed25519). Verify any copy: GET https://api.theflatratemovers.com/v1/verify?file=datasets-answer-evidence.json

Download the dataset (JSON)