An anonymized adaptation: a large content footprint and no inclusion in a shared question set
An anonymized and lightly normalized adaptation of a public audit. The firm is a management consulting and intellectual-products firm with a large public content footprint. The page count is rounded and nothing else identifies it. Its identifier in the toolkit is CASE-MC-01.
The counts show recognition and a zero on questions nobody reviewed
| Observation | Count |
|---|---|
By-name answers the audit judged correct |
15 of 15, each citing the firm’s own site |
Buyer-style answers that named the firm |
0 of 280 |
Indexable pages |
About 650, many times the engineering firm’s |
Structured data, and a file listing the site for language models |
Both present; their validity was not checked |
No association score appears, because no finder has been defined
The 280 answers came from a question set written for a shared audit. Many of the questions concern services this firm does not sell, so a zero on all of them proves little. A score needs questions that come from the firm’s own Finder Specification, and none has been written. The scorecard scores Entity Recognition only, leaves eight dimensions unscored, and computes no overall score.
The two zeros in the book’s second and third examples come from different question sets, so they are not a comparison. The example does not test the explanation that the firm is missing because it lacks enough content. It shows only that a large site with structured data and a file for language models can be absent from a shared question set.
The findings keep the observation, the hypothesis, and the check apart
Observation. The firm was identified correctly in every by-name answer and named in none of 280 shared buyer-style answers. Hypothesis. Recognition and association are separate properties here, and nothing in the counts says which questions the firm should have been an answer to. Experiment. Write the Finder Specification from the firm’s own client conversations, derive a benchmark from it, freeze it, and run it. Interpret nothing about association until then.
Observation. The site has many times the pages of the engineering firm, and the firm was named in none of the shared answers. Hypothesis. Volume may be unrelated to whether any page states a problem in a finder’s words, with evidence a finder can check and a source an engine can cite. This zero does not test that. Would disconfirm. A page-to-question map showing that most of the roughly 650 pages answer specification questions directly and with corroboration. Then volume and addressability go together here, and the zero needs another explanation. Experiment. Map every page to a question. Count the pages that answer none. Run the evidence chain on the questions that do have a page. Report volume, addressability, and inclusion side by side.
The first intervention comes after the specification and the freeze
IV-1 (content, days 31 to 60, after the Finder Specification is written and the benchmark is frozen): map each existing page to one Finder Specification question, consolidate overlapping pages into one canonical page per question, and repair the weakest link in the evidence chain for the highest-weight questions. It adds no page that has no finder question behind it, and it names the result that would show it failed: unchanged inclusion on the problem and provider questions on every engine at the retest. All data is in the toolkit download under examples/management-consulting-firm/.