The AI Visibility Assessment: how thoroughly we measure a clinic

Updated 22 October 2026 (Agent 11.1: robots.txt result on the Overview, #3587; developer brief says "check whether" unless confirmed). Previously updated 16 October 2026 (Agent 11.1: live probing described as the four patient conversations; how it differs from other AI scores). Previously updated 7 October 2026 (Agent 11.1: catch-up from every lane after their state-doc updates; scoring rebuild of 21 Sep, no score-trend claims, no citation claims for schema or credentials, Sites, content plan, support). Plain-English briefing. The assessment is the front door of the product and the evidence base for everything that follows. Written 24 July 2026. Checked against the state docs 2 October 2026 (Agent 11.1: schema wording guard). Last verified: 24 September 2026 (Agent 11: fixed criterion count removed per the rubric rebuild). Previously: 19 August 2026 (Agent 11).

What it is

The AI Visibility Assessment answers one question with evidence: how visible is this clinic to AI search today, and exactly why? It is not a quick scan or a single-tool score. It is a multi-source, multi-stage audit that produces a benchmarked, itemised picture of the clinic's standing, and every subsequent recommendation the platform makes traces back to something the assessment measured.

The reason for the thoroughness is simple: the content plan only works if it targets what actually moves the needle for this specific clinic. Random content creation is the industry default; evidence-directed content creation is the product.

What gets measured

1. The full website, page by page. We crawl the clinic's entire site and build a complete page inventory: every URL, its title, headings, and content. Each page is classified by type (condition page, treatment page, service page, blog post, team page, structural page) and individually scored against our quality checklist. This isn't a homepage-only skim: a 147-page clinic site gets 147 individual page evaluations.

2. Quality scoring against a research-grounded checklist. (Do not quote a fixed criterion count; it changed with the September 2026 rubric rebuild and the customer-facing figure is the dynamic per-page "assessed against N signals" line. See 09.) Every page is scored 0 to 100 against a unified, evidence-based checklist, using the checks that apply to its page type (never state a number of checks; the count is dynamic and page-type specific). The checklist was rebuilt on measured evidence on 21 September 2026: double-counted checks merged, unevidenced checks dropped, and the strongest measured driver added, which is how well a page answers the related questions patients actually ask. Two further checks test whether a page is readable without JavaScript and whether AI crawlers can reach it. Every clinic's page scores moved on 21 September for reasons unrelated to their site. Customers see each page's score, its band (Critical, Weak, Fair, Strong, Excellent), a Clinical Credibility band (Strong, Fair, Needs work) and a count of things to fix; never the itemised checks. Top Priorities was removed from the Page Inventory (about 5 Oct) and the old E-E-A-T grade was withdrawn as unreliable: do not defend any E-E-A-T grade a customer saw before October. The same checklist scores the clinic's existing pages, scores every page we later generate, and flags anything that falls short before publishing (it warns, it doesn't block the customer from publishing): one consistent quality standard through the whole platform.

3. Live AI visibility probing. We don't guess whether AI recommends the clinic: we ask, the way a patient would. We don't ask an AI tool "do you know this clinic?", which only shows whether it has heard of the clinic. Instead we play out four real patient conversations, each going from "what could this be?" through to "who should I see near me?", on ChatGPT, Claude, Perplexity and Google: 64 answers in all. The four topics come from what the clinic's own website says it treats, including its specialties. We record whether the clinic is named, where, and who is named instead. Answers vary from run to run, so results are judged over several measurements, and they stay comparable while the clinic's four topics stay the same.

How this differs from other AI scores. Most AI visibility checks ask one-off questions, often about the clinic by name, on a single run. Ours follows a patient's whole conversation, from symptoms to "who should I see near me?", on the clinic's own conditions, and reads results over several measurements. Briefing 11 has the full comparison and wording.

4. Reviews and reputation. Google Business Profile review data is collected and analysed: volume, average rating, velocity. Reviews are one of the strongest local trust signals AI search uses, and they also feed the structured data we later attach to the clinic's content (a real aggregate rating: never fabricated; if a clinic has no reviews, no rating is claimed).

5. Traditional search positioning. Search-engine results positioning for relevant local healthcare queries, via an independent SERP data provider, because traditional and AI search visibility interact, and because ranking data feeds the prioritisation of which pages to refresh.

6. Search demand data. Search-volume enrichment from an independent keyword data provider, so plan prioritisation can weigh real patient demand rather than guessing which topics matter.

7. Structured data audit. Whether the site carries machine-readable schema describing who the clinic is, who its practitioners are and what conditions it treats, and whether that schema is correct, complete and consistent with the visible content. (Never claim schema improves AI visibility or citation; our research measured no such effect. It is good technical practice for search engines.)

8. Local entity and competitive context. Discovery of relevant local entities and how the clinic sits within its local healthcare landscape.

9. The team. Practitioner detection and profile analysis: who works at this clinic, what credentials are visible, how complete and machine-readable each practitioner's professional identity is. Practitioner profile depth is measured as a first-class part of the assessment, because it is what lets content carry specific, true clinical detail and lets patients see who stands behind it. Since 6 Oct the Registration Numbers check also reads the numbers a clinic shows on its own website (team pages, bylines, structured data), so clinics that display AHPRA or HCPC numbers are no longer told they are missing (other clinics see the correction at their next assessment).

10. Deep research backstop. Where automated collection can't reach (unusual site structures, sparse data), a deep-research agent supplements the automated pipeline, so an unusual site still gets a complete assessment rather than a degraded one.

Belt and braces: multiple independent sources

No single tool's opinion decides anything. The assessment cross-references: our own full crawl, live AI-engine probing, Google Business Profile data, an independent SERP provider, an independent search-volume provider, the DOI registry (for citation verification downstream), and deep-research supplementation. Where sources disagree or data is missing, that is surfaced honestly: the platform flags "needs verification" rather than silently guessing. The internal design rule is: no fabrication, anywhere, ever. A clinic with no reviews shows no rating. A page that can't be verified isn't marked as fine.

What the customer sees

The output is a scored report the customer can actually read: an overall site health score, an AI visibility score with the real citation rate behind it, what's working well, priority improvements in plain language, per-category breakdowns (technical foundations, local trust signals, professional credibility, content and discovery), and the complete page inventory with individual page scores. Nothing is a black box: an admin can drill from the headline number down to the specific criterion a specific page failed.

robots.txt (live, #3587). The Overview says whether the clinic's robots.txt file is there and which search and AI crawlers it allows or blocks. "Unknown" appears only when the clinic's site refused our checker, with the reason.

The developer brief only says something is missing when our checks confirmed it. Otherwise it says "check whether" (live 22 Oct).

Benchmarking and what we can claim about change over time

Every assessment writes a dated snapshot. When the clinic re-measures (a lighter periodic re-run that reuses the stored discovery, re-crawls the pages and refreshes the score-driving inputs), the new result is stored alongside the old. Each measurement is a dated snapshot. AI Visibility results (the 64 answers) are kept for each run; when a clinic's four patient problems are refreshed, a new baseline starts and earlier results are marked as measured on the previous questions. Page and Site Health scores are snapshots too: we cannot yet show an honest page-score history or before-and-after trend (Agent 6, 7 Oct 2026: the dated measurements are not yet carried through to the comparison). Do not promise score tracking, show a trend, or imply we can prove improvement. A Full assessment's four patient problems reflect the clinic's specialties (up to two specialties plus at least two everyday conditions, all four only for a single-niche clinic; since 5 Oct). Each run uses only that run's answers.

How this connects to everything else

Common questions this briefing answers