This is the last piece in the series, and it is the one where the tool has to talk about itself without advertising. The useful question is not "what else can it score" but "what can it not tell you, and what would be worth measuring next". Both answers are more specific than they sound.
What this tool still cannot tell you
- Whether an engine will cite you. The free scan measures readiness: whether a page is in a state that makes citation possible. A high score is not a citation; a low one is not proof of invisibility.
- Why a page scored low. The evidence line says what the scanner saw; it does not say which cause mattered, or what the effect of the cause is in a given engine.
- What was used for training. Robots rules and content signals state a position; they do not prove what a provider did with content it may already have.
- What people actually ask. The free scan has no prompt bank and no demand data. The paid measurement uses a question bank the buyer approves; it is not a sample of real user queries.
- Competitors’ share of voice. The free scan does not run any answer engine. The paid measurement counts mentions in a fixed bank; that is a different measurement, not a market share.
- Whether a fix worked. A re-scan can show that a check now passes. It cannot show that a citation changed, because readiness and citation are different measurements and the second one needs a before/after run.
- Content truth or quality. Every markup rule is a presence check. A page can pass them with false claims, a weak argument, or a translation nobody would read.
- Rendering and mobile experience. The scanner fetches; it does not run a browser. JavaScript-only content is detected, but visual usability is not measured.
- Site-wide behaviour. One URL is one page. A site can have a perfect homepage and a broken checkout, and this scan will not see the checkout.
- Whether a context file is read. llms.txt, markdown alternates and content-signal are declarations. Whether an engine reads them is not observable from here.
What this series actually measured
Thirty pieces, one per scheduled slot, all English. The measured pieces in the series were deliberately small and scoped:
| Piece | Measurement | Scope |
|---|---|---|
| Answer-first writing | 101 heading-paragraph pairs, 91 openers ≤80 words | 16 English blog articles on 2026-10-26 |
| Meta descriptions | 33 of 40 in the 50-160 range | A systematic 40-URL sample of this site’s sitemap |
| Study re-read | 9 blocked rows = 7 robots + 2 refusals | The published 30-row dataset of 3 October 2026 |
| Redirects | 5 legacy URLs, 1 hop each | This site’s own redirect table on 2026-10-29 |
| TTFB | 0.75-3.8s TTFB; x-nextjs-cache MISS | 4 URLs, 3 runs each, one network, 2026-10-29 |
None of these is an industry benchmark, and each says so in its own piece. What they have in common is that a reader can reproduce them: the URLs, the dates and the method are in the text, and the raw data for the study is published.
What I would measure next
- A random or stratified sample. The convenience sample of 30 homepages was useful for finding mechanisms, not for prevalence. A sample drawn from a defined population, with page types stratified, would make every future claim stronger.
- Citations with a frozen question bank. The paid measurement already runs three runs per question. The next step is a before/after design on a page that changed, with the same bank fingerprint, so the comparison has a denominator and a date.
- Profile verification, automated. The sameAs rule checks presence. A follow-up fetch could check that each URL resolves, names the entity and links back; that is automatable and would close the rule’s stated gap.
- Markdown twin drift. Compare the H1, heading outline and word count of each twin with its HTML page in CI. Drift is the failure mode the declaration cannot see.
- Internal link graph. Orphan pages and pages with a single inbound link are cheap to find from a crawl and are a real part of discoverability.
- Page-type rules. A product page and a documentation page have different high-value checks; the current method applies one set to every page.
- Consent and paywall detection. Whether the body is in the response is checkable; whether a banner hides it is the next layer.
How the next measurement should be reported
The rules are already in the series: a sample size in the sentence, a count instead of a percentage, a date, the raw rows behind the count, and a line about what the measurement does not cover. What I would add is a fixed fingerprint for the question bank, so a before/after comparison can prove the questions did not change between runs.
And a refusal: no "visibility score" that compresses questions, engines and runs into one number. The compression is what makes the result look like a ranking, and it is the one claim this product does not want to make.
What the tool will not promise
It will not promise a citation, a ranking, or a training opt-out. It will not replace analytics, rank tracking or a human reading the page. It will not tell you what an engine thinks; it will tell you what a fetch saw and whether the page is in a state that makes being understood possible.
That is a narrower product than the category usually advertises, and it is the one this site can defend. The series ends here, but the measurements do not: the study can be re-run, the scan is public, and the counts are checkable.