The questions
Every scan asks 20 questions, built from fixed templates (hlsqt-v1.1) drawn from a buyer-interview taxonomy of how healthcare buyers research vendors. Two companies in the same category answer the same questions, which is what makes their results comparable. We never let a model invent the questions.
| Kind of question | How many | What it measures |
|---|---|---|
| Open category questions | 8 | Whether you are named when nobody mentions you. The headline measure. |
| Problem-led questions | 4 | Whether you are named when a buyer starts from their problem. |
| Head-to-head questions | 4 | How you are compared with the competitors you named, one competitor per question. |
| Questions about you by name | 4 | How you are described, and whether it is accurate. |
Every question ends by asking for named vendors, because questions that invite general advice get answered with criteria rather than names. Before any question is asked, a validator rejects the set if your name has leaked into a question that is supposed to be open, if a slot is unfilled, if a head-to-head question names more than one competitor, or if any question runs past 500 characters.
The AI assistants
ChatGPT, Gemini, Copilot and Google AI Overviews are asked through Cloro, which returns what a person sees in each product, including the sources it cites. AI Overviews come from a Google search; when Google shows no overview for a question, that question is left out of Google’s count rather than scored as naming no one. At least one assistant that always searches the web and cites its sources must be in every scan; without one there is no citation measure, so the scan refuses to run.
Runs
These models do not give the same answer twice, so every question is asked three times on every assistant. Your first read comes from run 1, and is marked directional. The final read is the mean of all three runs, and we publish the run-to-run spread beside it.
Who gets named
Mentions are counted by exact name matching, with no model involved. Names that are also ordinary words (Epic, Cedar, Clarity and the like) only count when the surrounding sentence is about software or your category. Visibility is the weighted share of open and problem-led questions in which you are named. Your position, the order AI names you in among the competitors you entered, is reported separately and never folded into the score.
What AI cites
Every cited link is reduced to its domain and classified from a fixed table of publications; an unknown domain is classified once by a model and added to the table. Earned coverage means trade press, review sites, clinical literature. Citation presence is the share of answers that cite sources in which your own site is one of them.
How you are described
The passage around each mention is labelled, from recommended strongly to mischaracterized, and every label carries the exact words that earned it. With fewer than 5 mentions we do not score sentiment at all and say so: a figure from two sentences is not defensible.
The score and the tiers
The composite weights visibility 0.55, citation presence 0.3 and sentiment 0.15. When sentiment is gated the composite uses the first two, reweighted, rather than scoring the missing measure as zero. The composite is never shown alone: two companies in the same tier can have opposite problems.
| Tier | Band | What it says |
|---|---|---|
| Leader | 80 to 100 | AI names you first. |
| Contender | 60 to 79 | AI recommends you alongside the leaders. |
| Overlooked | 15 to 59 | AI knows you, but rarely recommends you. |
| Invisible | 0 to 14 | AI doesn't name you in answers about your category. |
The bands apply to your percentile against a reference cohort of health IT vendors measured the same way. Until that cohort is complete, they apply to the composite itself and the tier is marked provisional.
What we never do
- Change a result. Each scan’s result is stored once and never rewritten; a rescan returns the stored record.
- Let a model compute a figure. The written briefing is checked, and any number in it that is not in the measured record is refused.
- Show you recommendations on this page. It says where you stand; the audit says what to do about it.
The question templates
- What are the best [category label] vendors for a [buyer type]?
- Which three [category label] platforms should a [buyer type] shortlist, and why?
- Who are the leading providers of [category label] software in the US healthcare market?
- Which vendors offer [keyword] for healthcare organisations?
- We are running an RFP for [category label]. Which vendors should be on the invite list?
- What [category label] options exist for a mid-sized [buyer type] without a large IT team?
- Which [category label] vendors integrate most easily with an existing Epic or Cerner environment?
- Which newer or emerging [category label] companies are worth watching?
- How can a [buyer type] solve [pain], and which vendors are named most often for it?
- A [buyer type] facing [pain]: which vendors would you shortlist, and is building in-house credible?
- Which [category label] vendors publish real outcomes data a [buyer type] could build a business case on?
- Which [category label] vendors have the best record on failed or stalled implementations?
- How does [brand] compare with [competitor] for [category label]?
- Is [brand] or [competitor] the better choice for a [buyer type]?
- What are the main alternatives to [brand], and how do they differ?
- We currently use [competitor]. Would moving to [brand] be worth it?
- What is [brand] and what does it do?
- Is [brand] a good [category label] vendor? What do customers say?
- What are the main criticisms or limitations of [brand]?
- Is [brand] a good fit for a [buyer type], and who is it not right for?