A scoring tool pointed only at other people’s websites is a marketing gimmick. The same tool pointed first at its author’s own site is a methodology. This is the first entry in the teardown series, and the target is us — three pages of algoholic.com through the exact engine that powers Am I in the AI?, with the weak dimensions published unedited.
The numbers below are real engine output against the production build of June 11, 2026 (engine v1.0, structural analysis — the same eight dimensions, the same weights, the same code path every visitor’s URL runs through).
The three pages
| Page | Score | Grade | Words analyzed |
|---|---|---|---|
Home (/) | 73/100 | Retrievable | 1,730 |
Flagship paper (/research/statement-level-visibility) | 83/100 | Citation-grade | 5,320 |
Consulting (/work-with-me) | 67/100 | Retrievable | 840 |
The ordering is the first finding, and it is the one the whole archive predicts: the research page outscores the marketing pages. The flagship paper — 5,300 words of quantified, sourced, entity-anchored prose with full schema — is the only page of the three that reaches Citation-grade. The pages built to sell score 10–16 points below the page built to argue.
Where the home page bleeds: qualifier proximity, 2/12
The single worst dimension on the home page. Of 40 quantified sentences, only 8 carry their qualifier — the window, the sample, the comparison — in the same sentence. The other 32 are hero stats and chronology captions: big numbers (“~700 posts”, “27 years”, “14 models”) presented as standalone fragments, with the context that makes them defensible living in a different block of the page entirely.
For a human scanning a designed page, that separation is fine — the eye travels. For a retrieval chunker, every one of those 32 numbers is an orphan: a claim that arrives at the model without the evidence attached. The home page is doing, at the layout level, exactly what the chunking note warns prose writers against.
The fix we are taking: the hero stat block and chronology captions get qualifier-carrying rewrites — “~700 posts in the nekuda archive since 1999” instead of a bare number with a label below it. Cheap edits, measurable delta; we will republish the score after.
Where /work-with-me bleeds: chunk survival, 0/12
Zero out of twelve — the worst single cell in the entire teardown. Only 4 of 62 “paragraphs” on the consulting page sit in the 30–140 word self-contained band, because the page is built from UI fragments: pricing-card labels, engagement-type bullets, form microcopy. Almost nothing on it is a retrievable span that states a complete claim.
Honest caveat, because the engine deserves the same scrutiny it applies: part of this is measurement artifact. A conversion page made of cards and labels is not failing at prose — it was never prose. The engine reads UI fragments as broken paragraphs, and a designed page will always score below a written one on this dimension. We considered suppressing the dimension for pages like this and decided against it, because the underlying point survives the caveat: a page made entirely of fragments gives a retrieval system nothing to carry away. A consulting page that contained even three complete, quantified, self-contained statements about the practice would have something in the context window when a model is asked “who should audit my AI visibility?” Right now it has two.
The fix we are taking: a short “what the practice actually does” prose block on /work-with-me — three statements, each one chunk-survivable, each one a claim a model can lift whole.
The four superlatives we had to go find
The engine flagged 4 unfalsifiable superlatives on /work-with-me and 2 on the flagship paper. We went looking. On the paper, both hits are mentions — the durability table literally discusses the phrase “industry-leading” as a failure mode, and the engine cannot tell use from mention. (Logged as a known v1.0 limitation; it stays, because special-casing our own pages is how scoring tools die.) On /work-with-me, two were real: a “best-in-class” and a “cutting-edge” in engagement descriptions, written before the taxonomy paper documented that confident assertion is the statement shape most likely to be silently absorbed. They contradict our own research. They are gone in the next content pass.
What survives the teardown
The strong dimensions are strong for documented reasons. All three pages max or near-max entity specificity (100–215 named-entity mentions per page) and front-loading, and all three carry full Person/Organization JSON-LD with a sameAs graph — the structural identity layer the entity-disambiguation note calls non-optional. The flagship paper additionally maxes machine-readable identity (ScholarlyArticle + BreadcrumbList) and takes 10/12 on chunk survival, which is what 121 deliberately self-contained paragraphs buys you.
The scoreboard we are accountable to
| Dimension (worst page shown) | Score | Committed fix |
|---|---|---|
| Qualifier proximity — home | 2/12 | Qualifier-carrying stat rewrites |
| Chunk survival — /work-with-me | 0/12 | Three chunk-survivable practice statements |
| Superlative liability — /work-with-me | 3/6 | Delete the two real superlatives |
| Source anchoring — /work-with-me | 7/17 | Cite the published research from the sales page |
This teardown gets a follow-up with re-scored numbers once the fixes land. If the deltas are small, that gets published too — the engine’s credibility is worth more than our score.
Run your own page through the same engine: Am I in the AI? · Methodology: /methodology · Want the 14-model probe behind the structural score? Work with me.