Methodology
Every dimension, every weight, and an honest list of what the scan cannot tell you. If you think a weight is wrong, you have everything you need to argue with it.
30 checks across eight dimensions, weighted to 48 points and normalised to a 0–100 score. Each check returns pass, warning or fail; a warning earns half credit. Two more checks are added only when they fire — a site that refuses our reader outright, and one serving a broken certificate chain — so a scan reports 30 findings, or 32 in those cases.
| Dimension | Weight | Share | What it measures |
|---|---|---|---|
| Structured data | 10 | 21% | Whether an engine can read your business as data rather than inferring it from prose — and whether you use a specific type (Dentist, Attorney, Restaurant) rather than a generic one. |
| Identity & entity clarity | 8 | 17% | Whether the page states who this business is, unambiguously, in both the visible copy and the markup. |
| NAP completeness | 8 | 17% | Phone, a complete postal address, and opening hours — as structured data, not just text. The most common gap we find in the field. |
| Reviews & reputation | 6 | 13% | Whether a real reputation exists in a form an engine can read and cite. Most businesses have the reputation and not the markup. |
| Answerability & content | 6 | 13% | Whether any page answers the literal question a customer types, in a form an engine can lift. |
| Authority signals | 4 | 8% | Linked, claimed profiles — and enough of them to corroborate the business across independent sources. |
| Technical foundation | 4 | 8% | HTTPS, canonical, viewport, sitemap. Genuinely table stakes, and weighted accordingly. |
| AI crawler access | 2 | 4% | Whether the four agents we currently test — GPTBot, PerplexityBot, ClaudeBot and Google-Extended — are blocked in robots.txt. Be clear about what each governs: PerplexityBot feeds Perplexity's index, while GPTBot and ClaudeBot govern model training and Google-Extended governs Gemini training and grounding rather than inclusion in Google Search. The agents that decide search citations, OAI-SearchBot and Claude-SearchBot, are not in this check yet. That is a gap in our engine, not a judgment that they do not matter. Rare, binary and serious when it fires — so it is a red flag, not a fifth of your grade. |
| Score | Tier | Meaning |
|---|---|---|
| 80–100 | Recommendable | Engines can find, read and trust the business well enough to name it confidently. |
| 60–79 | Findable | Engines can identify the business, but gaps cost it recommendations against better-described competitors. |
| 40–59 | Unconvincing | Engines can see the site, but too much of what a recommendation needs is missing or unreadable. |
| 20–39 | Unreadable | Engines cannot describe the business well enough to name it. |
| 0–19 | Invisible | Nothing on the page gives an answer engine enough to recommend the business by name. |
A methodology page that admits its limits is more useful than one that claims completeness — and this list is the part we would want to read first if someone else had written it.
An earlier version of this rubric was wrong in a way worth publishing. It weighted AI crawler access at more than a fifth of the grade, and because almost no site blocks those crawlers, nearly everyone collected those points for free. Combined with checks that could only ever return "warning" rather than "fail", the arithmetic floor for any site on HTTPS was 46 out of 100 — the bottom two tiers were literally unreachable, and a practice with no structured data at all could score 68 and be told it had "a solid base".
It also failed sites for being more specific than required: a dental practice correctly marked up as Dentist failed a check for LocalBusiness, because the comparison was a string match rather than a type hierarchy. And its contact check passed on the mere presence of the word PostalAddress anywhere in the HTML — which meant our own site, with no phone number on it at all, passed a check for having a phone number.
The current scale weights what a recommendation actually depends on, and most checks can now fail outright. Scores went down across the board, including our own. That is the point of publishing the rubric rather than the score.
With 30 checks across eight dimensions, weighted to a total of 48 points and normalised to a 0–100 score. Two further checks are added only when they fire — a site that refuses our reader outright, and one serving a broken certificate chain — so a scan reports 30 findings, or 32 in those cases. The dimensions and weights are published in full on this page. The scan measures readiness to be recommended — whether an answer engine can identify a business, read its facts, and find corroboration for them — rather than whether the site ranks.
Because it is binary and rare. Almost no small business blocks these agents, so weighting the check heavily inflates everyone else's score for something they were never going to fail. It is reported as a red flag when it fires instead of quietly padding a fifth of every grade. Be precise about what a block costs, too: blocking a training crawler such as GPTBot or ClaudeBot does not remove you from ChatGPT or Claude search results, and Google-Extended does not affect inclusion in Google Search at all.
We have published one scan set against the current rubric, and it is small: six Boulder dental practices in September 2026, scoring 35 to 79 with a median of 64 and half of them under 60. Their technical-foundation median was 94 against a reputation-markup median of 9 — a site that loads cleanly and says almost nothing machine-readable about its reputation. Six practices is not a distribution, and we will replace this answer with a real one once enough sites have been scanned to state one. Scores above 80 are uncommon and usually mean someone has done deliberate structured-data work. If a scan returns a high score for a business that is plainly invisible in AI answers, we would rather hear about it than not — the rubric is published so it can be argued with.