How every number
is measured
Every weight and threshold below is read straight out of the scoring code, so this page cannot drift from what actually runs against your site.
The Global GEO Score
One number from 0 to 100, a weighted average of three independent pillars. Each pillar is measured on its own and knows nothing about the other two.
| Pillar | Weight | What it averages |
|---|---|---|
| GEO page quality | 40% | Every page in your latest completed scan |
| Citation authority | 35% | The citation score of those same pages |
| AI visibility | 25% | The questions in your most recent completed AI visibility batch |
A pillar with no data is excluded, not scored zero
Its weight is redistributed across the pillars that were measured. A project with no Lens produces no AI visibility data by design; counting that as zero would have capped its score at 75 and reported a well-built site as weak. What the number describes is your site, not our coverage of it — so the dashboard says Not measured and tells you how many of the three pillars are behind the score.
Grade bands
GEO page quality 40% of the composite
Scored per page from the crawled HTML — no LLM involved, so the same page always scores the same. Five components, each 0–100 and each worth 20% of the page score.
Structure Clear hierarchy an assistant can follow 20%
- An H1 is present
- The page uses lists
- The page uses tables
- Semantic HTML elements are present
- Flesch reading ease, graded: 60+ scores full marks, 30–40 scores a quarter
Semantic depth Enough substance to be worth quoting 20%
- Word count — over 800 scores full marks, 300–800 partial, under 300 nothing
- A meta description is present
- More than two internal links
- At least one outbound link — the page cites something itself
Interpretability Machine-readable meaning 20%
- Attributed quotes are present
- Any schema.org structured data at all
- Schema of a type assistants actually use: Article, Product, FAQPage, HowTo, BreadcrumbList
- Alt text coverage — counted only on pages that have images
Conversational Answers shaped the way a question is asked 20%
- A heading is phrased as a question
- Question-and-answer blocks, up to three of them
Engagement Visual and formatting substance 20%
- The page has images
- More than three images
- Lists or tables break up the text
- More than five paragraphs
A signal a page cannot have is not a signal it failed
Each component scores as a proportion of the points that actually apply to that page. Alt text is only counted where there are images — a text-only page is not marked down for the absence of a problem. The weight is redistributed over the remaining signals, the same way the composite handles an unmeasured pillar.
Citation authority 35% of the composite
Assistants prefer pages that show their work. Every outbound link and quoted claim on a page is extracted, the source domain is classified by authority, and the page is scored out of 100 across eight components — these are points, not percentages, and they sum to 100.
| Component | Points | What earns them |
|---|---|---|
| Authoritative sources | 20 | How much of what you cite is Tier 1 or Tier 2 |
| Evidence-backed claims | 20 | Statistics and figures that carry a source |
| Source diversity | 15 | How many distinct domains and authority tiers you draw on |
| Quotations & attribution | 10 | Quotes that name who said them |
| Freshness & recency | 10 | How recent the cited material is |
| Primary sources | 10 | Original research, filings and documents rather than coverage of them |
| AI extractability | 10 | Whether a claim and its source can be lifted out together |
| Schema & metadata | 5 | Structured data describing the citations |
Authority tiers
Each cited domain is placed in one of four tiers, and the tier sets how much that citation counts for. A domain we have not classified starts in Tier 4 rather than being thrown away.
| Tier 1 | Institutional / High Authority | counts 100% |
| Tier 2 | Strong Industry Authority | counts 75% |
| Tier 3 | Commercial / Medium Authority | counts 45% |
| Tier 4 | Weak / Unverified | counts 15% |
AI visibility 25% of the composite
The other two pillars measure your site. This one measures the answer: real buying questions are put to an assistant, and we score whether it brings you up.
1. The questions
Written from what your site says it sells, the market it serves and your audience lenses — in your market's own language. Most of them deliberately do not name you: the measurement is whether an assistant brings you up unprompted, and a question that names you guarantees a hit and measures nothing.
2. The answer is grounded in a real search
Each question is run through a web search first, and the assistant may only recommend sites that actually came back. Asked from memory alone, a model has no reliable recall of a small or regional market and invents businesses fluently — which would score invented rivals against you, and an invented mention of your own domain as a win. If a grounded answer names your domain when the search never returned it, the mention is recorded as invented and not counted. Where no search backend is configured the answer is marked ungrounded on the report rather than quietly mixed in.
3. Each answer is scored out of 100
- Your brand is named40
- Your URL is cited30
- Named in the first sentence+20
- Second sentence / third sentence+12 / +5
- Named more than onceup to +10
The brand is matched as a whole word, with diacritics folded on both sides — a Romanian assistant writes "Pușcălupi" for the brand behind puscalupi.com, and an exact match would call that a miss. Matching a bare substring instead would score "algorithm" as a mention for a site on go.com.
4. Who gets recommended instead
Every answer is also checked against your competitor list, so the report shows how often each rival was named against how often you were. Scoring zero because assistants recommend nobody and scoring zero because they recommend three named rivals every time are different findings, and only the second is something to act on.
Competitor GEO score
A competitor's GEO score is their homepage only — one page, fetched live when you press Scan, put through exactly the same page quality scorer as your own pages. We do not crawl third-party sites.
Compare it against your homepage, not your site average
Your own pillar score is the average of every page in your latest scan. A competitor's is a single page — usually their strongest, since homepages are built to be. Reading one against the other is not a like-for-like comparison, so the competitors page shows your homepage score next to theirs and that is the number to benchmark on.
- The score is only ever written from a real fetch. A competitor that has never been scanned shows —, not a zero.
- A failed scan leaves the previous score untouched, so a temporary network problem does not read as "this competitor scores 0".
- Competitors found automatically come from real search results, filtered to sites that were fetched and confirmed to exist — the model never supplies a domain.
Technical SEO audit
A separate track from the composite score — it does not feed the Global GEO Score. It is the classic technical checklist, scored out of 100 and calibrated against Lighthouse.
What a scan covers
- Pages are discovered from your start URL and your sitemap, then from links found on the pages already crawled.
- A single scan reads at most 50 pages. On a larger site the score describes the pages that were read, not the whole site.
- Only HTML pages are scored. PDFs, images and feeds are skipped rather than scored badly.
- Each page is analysed, scored and audited independently, so one page that fails to load does not fail the scan.
Something here not matching what you see?
These are the exact weights the scoring code runs with. If a number on your dashboard does not line up with this page, that is a bug and we want to hear about it.
Get in touch