LLMention
← Back to ScannerMethodology

How the GEO score is calculated

Every rule on this page is the rule the scanner actually executes. The list below is generated from the same data the analyser runs on, so the published method and the live method cannot drift apart.

Browse sections

By Alex XuMethod last revised

How is the total score calculated?

The scanner runs 40 explicit checks across twelve dimensions. Each check is worth a fixed number of points inside its dimension. A dimension’s score is the fraction of its available points that were earned, and the total is those twelve scores combined by weight:

total = Σ ( dimension_score × dimension_weight / 100 )

where dimension_score = earned_points / available_points × 100
and   Σ dimension_weight = 100

There are no score floors. A page that satisfies none of the checks scores near zero rather than being lifted to a respectable-looking minimum. Points are relative weights within a dimension: where two entries below describe the two outcomes of a single check, only one of them applies to any given scan.

There is one thing that is not a floor and is worth separating from one: a small number of checks are marked not applicable to a particular kind of page, and are removed from the calculation entirely rather than counted as a failure. A policy page is not a worse page for containing no expert quotations; the question does not apply. When a check is excluded, its points leave both sides of the fraction and the remaining dimensions are renormalised against their own sum, so the total stays a percentage of what could actually be assessed. Every exclusion is listed against the check it applies to below, and the tests assert that each one names a real check and that no exclusion applies to a homepage — which is what makes this change score-neutral for every scan the tool performed before page types existed.

There are two ways a check can reach that state, and the difference is worth knowing when you read a score. Page-type exclusions are declared against the check they apply to in the table the analyser executes and this page renders below: a privacy policy is not asked for expert quotations or statistics because it is a policy, and an article body is not asked for the site’s About and Contact links because a single-page fetch never sees the chrome around it. Rule-level inapplicability is decided by the rule itself from what it found on the page rather than from the kind of page it is: hreflang asks a page to declare alternate language versions, a page that declares none cannot be judged on that question, and one fetched page cannot tell a single-language site from a site that has translations it has not declared. Neither kind is a failure and neither is a discount. In both, the points leave the numerator and the denominator, the remaining weights are renormalised, and the total stays a percentage of the checks that could actually be assessed — which is why a perfect score means “nothing that applies to this page failed”, never “every rule in the catalogue passed”.

What changed, and when

A method that never changes is not being maintained, and one that changes without saying so cannot be checked after the fact. Every change that could move a score is recorded here with its date and its reason, because a score is only comparable to another score produced under the same rules:

  • — hreflang is no longer scored when a page declares fewer than two language codes. It asks a page to declare alternate language versions, which a single-language site has none of, so the old rule cost such sites 2 points for a question that does not apply. The scanner fetches one page and cannot tell a single-language site from one that has not declared its translations, so absence is now marked not applicable and excluded from the score. A page that declares exactly one code still fails.
  • — A check can be marked not applicable to a page type it cannot judge, and is then excluded from the score rather than failed. Six exemptions were published for legal, contact and article pages - a privacy policy is not a worse page for quoting nobody - and the remaining dimensions are renormalised, so a root path scored exactly what it scored before.

The list is short on purpose and it is the record of the method rather than of this site: copy edits and new worked examples do not appear. The one entry a reader of the study has to know about is the first — that table was measured before it, which is why the study page states both dates and leaves the rows as measured instead of re-scoring them under the current rules.

What do the grades mean?

GradeScoreReading
A90-100Excellent readiness — the page has what a citation needs
B80-89Good readiness — most of what a citation needs
C70-79Average readiness — some of what a citation needs, and several dimensions need work
D60-69Poor readiness — significant GEO gaps
F0-59Critical readiness — most of what a citation needs is missing

What the scanner cannot tell you

A score is only useful if its limits are stated. These are the ones that matter:

  • It does not ask any AI engine about you. Nothing here measures whether ChatGPT currently recommends your brand. It measures whether your pages are in a state that makes being cited possible. Those are different questions, and conflating them is the most common overclaim in this category.
  • A refused request is not proof of a block. Large sites verify crawlers by IP address rather than user-agent, so a scanner’s request can be refused even when real GPTBot traffic is served normally. Whether a site blocks AI crawlers is answered here by robots.txt, which is the site’s stated intent, not by guessing from one response code.
  • The score describes one page unless a site scan is asked for. A single-URL scan cannot see site-wide problems on interior pages. A site scan reads a few more pages the sitemap names — capped at five by default and ten at most — scores each of them with exactly these rules, and reports every page’s score separately, so a homepage number is never passed off as a site number.
  • Performance checks are coarse. Real Core Web Vitals need a browser; this scanner deliberately does not run one. That is why the Delivery dimension carries the lowest weight.
  • Weights are a judgement, not a measurement. The dimension weights reflect published research and the fact that unreadable pages cannot be cited. They are stated in full so you can disagree with them specifically rather than vaguely.

The twelve dimensions

01AI Crawler Access

16%

If a crawler cannot read the page, nothing else on this list can matter. This is the only dimension where a failure makes the rest of the score moot.

Why this weight: Highest weight of the twelve. A page that is blocked or unreadable has no path to being cited, so it dominates the total.

  • 2 ptA request for /robots.txt returns HTTP 200 with a non-empty body. A 200 with an empty or whitespace-only body counts as not served, because it publishes no policy at all.
    Not satisfied: No readable robots.txt at the domain root.
  • 6 ptParsing robots.txt into user-agent groups, none of gptbot, claudebot, perplexitybot, oai-searchbot or google-extended is disallowed from /, AND the homepage served a crawler-shaped request with HTTP 200. An exact agent group takes precedence over the * group, the longest matching pattern wins and Allow wins ties, an empty Disallow value matches nothing, and * is treated as a wildcard. When the homepage refused the request, the verdict follows a second probe of the same URL made with a browser-shaped User-Agent: if that one returns 200 the refusal is aimed at identified crawlers, and the check fails under robots-ai-blocked; if it is refused as well, the refusal is about the scanning address rather than the site, and this check scores 2 of 6 as unverified.
    Not satisfied: At least one AI crawler is disallowed from the site root, or the server refuses identified crawlers outright.
    Note: A site with no robots.txt at all passes this check only when the homepage was actually served, because with no file no crawler is disallowed by it. That is a deliberate reading: absence of robots.txt is a policy gap, not a block, and the gap is reported separately by robots-present. A refusal by the server is the stronger evidence and now overrides the file: robots.txt saying nothing is not permission if the request is rejected before it is read.
  • 6 ptThe inverse of the check above, reported by name so the blocked agents appear in the finding. It is also reported when the server refuses identified crawlers while serving a browser-shaped request, because the effect on those crawlers is the same.
  • 1 ptrobots.txt contains a Sitemap: line.
    Note: The presence of the line is the test, not what it points at: the URL is not fetched, so a Sitemap: line naming a file that does not exist passes here and fails sitemap-valid. The two results are reported separately rather than merged, because a declaration that exists and a file that resolves are different facts.
  • 2 ptThe homepage returns HTTP 200 to a non-browser request that identifies itself honestly.
    Not satisfied: A non-200 homepage response. The access verdict itself is reported by robots-ai-allowed or robots-ai-blocked, which account for what the refusal was aimed at.
  • 2 ptThe homepage has no meta robots directive containing noindex.
    Note: The directive is read out of the served HTML and nowhere else, and the pattern expects the name attribute to come before the content attribute: `<meta name="robots" content="noindex">` fails this check while `<meta content="noindex" name="robots">` passes it. An X-Robots-Tag: noindex response header is not read at all, because response headers are not part of the document this check is given.
  • 3 pt/sitemap.xml returns a body containing <urlset> or <sitemapindex>.
    Not satisfied: No sitemap, or a response that is not valid sitemap markup.
    Note: The body only has to contain `<urlset` or `<sitemapindex` somewhere. The addresses inside are not read, not counted against the 50,000-entry limit and not fetched, so a sitemap listing pages that return 404 passes. A response that arrives but is not sitemap markup is a partial pass worth 1 of the 3 points.
  • 1 ptrobots.txt contains a Content-Signal directive with a value on the same line, for example `Content-Signal: search=yes, ai-input=yes, ai-train=no`. The directive is matched case-insensitively at the start of any line.
    Not satisfied: No Content-Signal directive in robots.txt.
    Note: Weighted 1 of the 16 points in this dimension because adoption is early: the directive was proposed in 2025, and this scanner reads it as a stated position rather than as a signal any engine has committed to. It is scored because the convention is cheap to follow and a tool that recommends it should be able to show its own line, not because it moves a ranking.

02Machine Readability

12%

JSON-LD is how a model binds your brand name, domain and product into one entity instead of inferring three unrelated strings.

Why this weight: Second highest. It is deterministic and fully in your control, and it is what makes entity disambiguation possible.

  • 5 ptAt least one <script type="application/ld+json"> block parses as JSON and declares at least one @-keyword (@context, @type, @graph or @id). A block containing only {} parses but declares nothing, so it does not count.
    Note: The first twenty JSON-LD blocks in the document are the ones examined, so a page that emits a block per component is read only as far as its twentieth. A block that fails to parse does not stop the ones after it from being read, because each is parsed on its own.
  • 4 ptParsed JSON-LD contains an @type of Organization, WebSite, Person or LocalBusiness.
    Note: Types are collected from every node in every block that parsed, nested ones included, so a Person node inside an Article's author property satisfies this check on its own: a page can pass it without declaring an Organization anywhere. The evidence line names the types that matched, so the finding says what it saw rather than asserting that a brand entity exists.
  • 3 ptParsed JSON-LD contains an @type of FAQPage, Article, BlogPosting, HowTo, Product, SoftwareApplication or BreadcrumbList.

03Content Depth

11%

Generative engines select sources that answer a question thoroughly. Thin pages are rarely retrievable regardless of how well they are marked up.

Why this weight: High, because depth is a precondition for retrieval rather than a bonus on top of it.

  • 5 ptVisible text, after removing script, style and comment content, contains at least 800 word tokens. 300-799 is a partial pass.
    Not applicable to legal, contact pages. Excluded from the score rather than counted as a failure.
    Note: The count runs over everything left after tags are removed, so navigation, the footer, a cookie notice and repeated boilerplate are all included: a page can reach 800 words on chrome and very little content. That is the limit of a count which measures how much text a crawler receives rather than how much of it answers a question.
  • 3 ptThe page contains at least 4 H2 elements. 2-3 is a partial pass.
    Note: The heading text is never read, only the element and its level, so four H2s that all say "Overview" are four sections to this check. Nothing asks whether the sections divide the content along anything a reader would recognise.
  • 3 ptThe page contains both a <ul>/<ol> and a <table>. Having only one is a partial pass. Script and style bodies are excluded first, so markup that only appears inside a string does not count.
    Note: Any list and any table in the response count, including the navigation menu and a table used for layout: the check asks whether the page uses the two structures a parser can lift, not whether the content is inside them.

04Citability & Evidence

11%

Statistics, quotations and cited sources are the interventions with the largest measured effect in the published GEO research.

Why this weight: High and research-driven: this is the dimension most directly tied to measured visibility gains.

  • 4 ptBody text contains at least 5 numeric claims: percentages, currency amounts, multipliers of the form 3.2x, thousands-separated figures, or raw numbers of five digits or more. A bare four-digit number is not counted, because it is a year far more often than it is a finding.
    Not applicable to legal, contact pages. Excluded from the score rather than counted as a failure.
    Note: Counting bare three-digit numbers used to make any page with a copyright line or a product ID look quantified, which is the opposite of what the check is for. The count runs over the whole response, chrome included, and each match counts separately, so five percentages satisfy it as readily as five different kinds of figure. One to four numeric claims is a partial pass worth 2 of the 4 points.
  • 3 ptThe page contains at least one <blockquote> element.
    Not applicable to legal, contact pages. Excluded from the score rather than counted as a failure.
    Note: The element's presence is the whole test. Nothing reads the text inside it and nothing checks that it is attributed to anybody, so a pull-quote of the page's own sentence passes. Telling a citation from a styled paragraph needs a judgement this scanner does not make.
  • 3 ptThe page links to at least 2 external hosts on the authority list: arxiv.org, doi.org, nature.com, science.org, acm.org, ieee.org, springer.com, sciencedirect.com, any .gov or .edu host, wikipedia.org, github.com, developer.mozilla.org, web.dev, nih.gov, who.int or europa.eu. Own-domain links are excluded.
    Not applicable to legal, contact pages. Excluded from the score rather than counted as a failure.
    Note: The count is of links, not of distinct hosts, so three links to the same authority domain satisfy a rule that names two sources - the evidence line prints the distinct hosts, which is how a page can pass with only one domain on it. Nothing follows the links either: whether the cited page says what this page claims is not something a single fetch can establish. Exactly one authority link is a partial pass worth 2 of the 3 points.
  • 1 ptThe page declares a rel=canonical link.
    Note: The declaration is read and the address in it is not. Nothing compares the canonical with the URL that was fetched and nothing follows it, so a page pointing its canonical at a different page, or at one that returns 404, passes. Confirming the target would be a second request, which a single-URL scan does not make.

05Answer Readiness

10%

Answers are extracted as spans, not pages. A question-shaped heading followed by an immediate answer is the easiest thing for a retrieval system to lift intact.

Why this weight: Substantial, and largely mechanical to fix.

  • 4 ptAt least 3 H2 or H3 headings either end with a question mark or begin with how, what, why, which, when, who, where, is, are, do, does, can, should or will.
    Note: A heading qualifies by ending in a question mark or by starting with one of those fourteen words, so "What we do" counts as a question-shaped heading and "Our approach" does not, whatever either one actually answers. One or two qualifying headings is a partial pass worth 2 of the 4 points.
  • 3 ptParsed JSON-LD contains a FAQPage node. If instead the page uses <details>/<summary> markup without the schema, that is a partial pass.
    Not satisfied: No FAQPage schema. Note that this check reads structured data, not the word "faq" or "question" appearing anywhere in the HTML.
    Not applicable to legal pages. Excluded from the score rather than counted as a failure.
    Note: The node only has to be present. A FAQPage whose mainEntity list is empty, or whose questions appear nowhere in the visible text, satisfies this check, because comparing the markup against the page is a judgement the scanner does not make.
  • 3 ptOf the first 30 heading-then-paragraph pairs (H2/H3 optionally followed by wrapper elements then a <p>), at least 60% have an opening paragraph of 80 words or fewer.
    Note: Only the first thirty pairs and the first 200 KB of the document are examined, because this is the most expensive pattern the scanner runs. A pair counts when an H2 or H3 is followed by wrapper elements and then a paragraph, so a heading followed by a list contributes nothing. Fewer than three pairs fails the check; three or more pairs with under 60% short openers is a partial pass worth 1 of the 3 points.

06Trust & Authority

10%

E-E-A-T signals decide whether a model treats a claim as safe to repeat rather than something it should hedge.

Why this weight: Weighted equally with answer readiness: attribution is what separates a quote from a rumour.

  • 2 ptThe homepage was retrieved over https://, rather than only over plain http://.
    Note: The scheme reported is the one the request completed on. https is tried first and http only when the first attempt produced no response at all, so a site that answers over http is scored on that response rather than credited with a scheme it did not serve.
  • 3 ptHrefs on the page include both an about-style path (about, company, team, who-we-are) and a contact-style path (contact, support) or a mailto: link.
    Not applicable to article pages. Excluded from the score rather than counted as a failure.
    Note: The paths are matched as substrings, so /about-us, /company/team and /who-we-are all count as about-style, a mailto: anywhere in the document counts as contact, and a link to /support counts whether or not it reaches a person. Having only one of the two is a partial pass worth 1 of the 3 points.
  • 3 ptAuthorship is signalled by any of: a Person node in parsed JSON-LD, an author property with a non-empty value in parsed JSON-LD, rel="author", or a visible byline matching "by Firstname Lastname".
    Note: Any one of the four signals is enough and none of them is verified: a byline-shaped string anywhere in the visible text, such as "by Jane Doe" in a caption or a comment, satisfies the check without any person existing behind it. That is the limit of judging authorship from a single page.
  • 2 ptParsed JSON-LD contains a sameAs property with a non-empty value.
    Note: The property has to be present with a value in parsed JSON-LD; the profiles in it are not fetched, so a sameAs list of accounts that no longer exist passes. It does not have to sit on an Organization node either - a sameAs anywhere in the parsed graph counts, because what it identifies is the entity the rest of the graph refers to.

07Semantic Structure

8%

Heading hierarchy and landmark elements are how a parser locates section boundaries at all.

Why this weight: Moderate. Mostly a one-time fix.

  • 3 ptThe page contains exactly one non-empty <h1>. Zero or several is a partial or failed result.
    Note: Zero H1s fails the check outright and scores none of the 3 points; more than one is a partial pass worth 1 of the 3. Headings with no text are dropped before they are counted, so an empty `<h1></h1>` neither satisfies the check nor breaks it.
  • 3 ptAt least 3 of main, article, section, header, nav and footer appear as elements.
    Note: The test reads element names, not ARIA roles, so a page built from `<div role="main">` and `<div role="navigation">` scores nothing here however it reads to assistive technology. Fewer than three of the six is a partial pass worth 1 of the 3 points, and there is no failing outcome: this check can cost a page at most two of its three points.
  • 2 ptThe raw server response contains at least 100 words of visible text without executing JavaScript.
    Note: The hundred words are counted across the entire response, so navigation, the footer and cookie text are part of the total. A page whose article is assembled by JavaScript can clear the threshold on its chrome alone: the check proves the response is not empty, not that the article is in it.

08Metadata & Discoverability

7%

Title, description and canonical control what a search or answer surface can show about the page.

Why this weight: Moderate, and cheap to satisfy.

  • 3 pt<title> is between 15 and 65 characters long.
    Note: The length is measured in characters of the title element as the HTML contains it, with entities left undecoded, so an `&amp;` counts as five characters while rendering as one. A title that exists but falls outside 15-65 characters is a partial pass worth 1 of the 3 points; only a page with no title element at all scores none of them.
  • 2 ptThe meta description is between 50 and 160 characters long.
    Note: The measured length is the content attribute as written, so entities count in their source form and a description can be inside the range as a reader sees it and outside it as the scanner measures it. A description that exists but falls outside 50-160 characters is a partial pass worth 1 of the 2 points.
  • 1 ptAt least one meta property beginning with og: is present.
    Note: One property from the family is enough, and which one it is does not matter: a page carrying only og:title passes as surely as one carrying og:title, og:description, og:image and og:url. The check answers whether the page speaks Open Graph at all, not whether its tags are complete.
  • 1 ptThe <html> element carries a lang attribute.
    Note: The attribute's presence is tested, not whether its value is a language anyone speaks, so `lang="english"`, `lang="EN"` and `lang="zz"` all pass. Reading the value for a region subtag is the separate lang-region check, which is why the two can disagree about the same attribute.

09AI Context Files

5%

A cheap and optional signal: one URL that lets a site state its own context for machines. Publishing it is easy; whether any engine reads it is not something this method measures.

Why this weight: Deliberately low. The check is worth 4 of this dimension's 6 points and the dimension carries 5% of the total, so the file is capped at 3.3 of the 100. It is scored because the convention is cheap to follow, not because it moves a ranking.

  • 4 pt/llms.txt returns more than 20 bytes and contains an H1, a > summary line and at least 3 markdown links. Missing any of those is a partial pass.
    Note: Four properties are examined: more than 20 bytes, an H1, a > summary line and at least three markdown links. What the file says is not judged, so three links to pages that do not exist pass the structure test.
  • 1 ptrobots.txt is readable, so a crawler policy is published. Readable means a body with non-whitespace content: an HTTP 200 with an empty or whitespace-only body publishes no policy and fails this check, exactly as it fails robots-present.
    Note: This is the same file the robots-present check reads, scored a second time in a different dimension: once because crawler access is what the policy governs, and once because a machine-readable context file is published. Adding robots.txt therefore moves two checks in two dimensions, and because that makes this the second score for one file, a site that serves an empty one can no longer collect this point for it.
  • 1 ptThe document head contains a <link> with rel="alternate" and type="text/markdown", in either attribute order.
    Not satisfied: No markdown alternate is declared on the page.
    Note: This reads the declaration only. The scanner fetches one URL, so it does not follow the alternate to confirm that it resolves, and a page can pass this check while pointing at a link that 404s. That limitation is stated here rather than hidden; closing it means a second request in the fetcher.

10Freshness

5%

Dated content is deprioritised in generated answers, and an undated page gives an engine nothing to reason about.

Why this weight: Low to moderate: a real but secondary signal.

  • 3 ptA machine-readable date is found in dateModified, datePublished or article:modified_time. Within 12 months is a full pass; older is a partial pass. If only the HTTP Last-Modified header carries a date, that is a partial pass.
    Note: The first date found in source order wins: dateModified is checked before datePublished, which is checked before article:modified_time. They are not compared with each other, so a fresh datePublished beside a stale dateModified is read as stale. A date within twelve months is worth the full 3 points, a recent Last-Modified header 2, and an old or unparseable date 1.
  • 2 ptThe current calendar year appears in the visible body text.
    Note: The current year is looked for anywhere in the visible text, including a site-wide footer that no individual page updates: a page whose own content is from 2024 passes on a footer reading © 2026. The year does not have to sit next to the word copyright, and it is taken from the scanner's clock rather than from anything the page declares.

11International Readiness

3%

Language and region markup decides which language market a page can be retrieved in at all.

Why this weight: Low for a single-market site, and the first thing to fix if you serve more than one language.

  • 2 ptAt least 2 distinct hreflang language codes are declared. Repeated attributes for the same language do not count, and x-default is a default marker rather than a language, so a single-language site declaring `en` plus `x-default` does not satisfy this. A page that declares no hreflang at all is not scored on this rule: with one page fetched, the scanner cannot tell a single-language site from a multilingual one that has not declared its versions, so absence is marked not applicable rather than failed. Declaring exactly one code still fails.
    Note: The codes are read from the attributes and the pages they point at are not fetched, so two hreflang tags naming language versions that were never built satisfy this check. It is a declaration check, and like the markdown alternate above it, it is one this scanner cannot close from a single fetch. The rule changed on 2026-10-23: before that date, no declaration was scored as a failure, which cost every single-language site 2 points for a question it was never asked. The change is recorded on /methodology/, which is where a reader should check it rather than here.
  • 1 ptA language tag carrying a region subtag appears in <html lang>, in og:locale or in an hreflang attribute - for example en-GB, en_GB or de-AT. A bare language such as "en" does not satisfy this; that is what html-lang checks.
    Note: This check previously ran the identical <html lang> regex as html-lang, so one attribute earned points in two dimensions while the rule promised a region.

12Delivery & Mobile

2%

A slow or non-mobile-readable page is dropped before any content analysis happens.

Why this weight: Lowest. The checks here are coarse on purpose; full Core Web Vitals need a real browser, which this scanner deliberately does not run.

  • 1 ptA meta viewport tag is present.
    Note: This reads the declaration, not the value. A viewport tag with `content="width=10000"` passes exactly as `width=device-width, initial-scale=1` does, because deciding whether a viewport configuration is usable needs a rendering engine and this scanner deliberately does not run one.
  • 1 ptThe uncompressed HTML response is under 500 KB.
    Note: This measures the uncompressed HTML document and nothing else: images, fonts, CSS and JavaScript are separate requests and are not counted, so a 400 KB page with three megabytes of images passes. It is also the coarsest band in the method - 499 KB scores the point and 501 KB scores none of it, with nothing in between.

Primary sources

Where this method makes a judgement about what generative engines favour, it follows published research rather than folklore. Read the sources directly:

  • GEO: Generative Engine Optimization (KDD 2024) — The paper this site's citability weighting follows. Its abstract reports visibility gains of up to 40%, and its Table 1 measures the best of the tested methods against the no-optimization baseline: 41% on position-adjusted word count and 28% on subjective impression, with quotation, statistics and citation additions among the strongest.
  • What Generative Search Engines Like and How to Optimize Web Content Cooperatively — Introduces AutoGEO, which extracts generative-engine preference rules and rewrites content for more traction.
  • What Gets Cited: Competitive GEO in AI Answer Engines — A controlled two-document testbed, 252,000 trials across six models: topical relevance and list position were the biggest drivers of being cited first, with explicit price information and a recent timestamp helping consistently.
  • llms.txt standard — The convention itself. Cited so you can read the primary source rather than our summary of it.

This site holds itself to the same standard it measures: it publishes its own llms.txt, declares a single Organization entity in JSON-LD, and labels its own dates. You can verify all of it by scanning this domain.

Questions about the method

Is the total an average of the twelve dimensions?

No. Each dimension is scored as the fraction of its own points that were earned, and the twelve are then combined by weight. A dimension worth 16% moves the total three times as much as one worth 5%.

Why is a context file weighted so low?

Because nothing we can measure shows it moving a ranking, and the arithmetic is small: the check is worth 4 of that dimension's 6 points, and the dimension carries 5%, so the file is capped at 3.3 of the 100. It is scored because the convention is cheap to follow.

Can I reproduce a score?

Yes. A scan is a single GET against the public API, and the same inputs produce the same score. The study over 30 homepages names the exact endpoint it used.

Evidence and sources

GEO can boost visibility by up to 40% overall in generative engine responses, and in the paper's Table 1 the best of the tested methods improve on the no-optimization baseline by 41% and 28%. — Generative Engine Optimization, KDD 2024

The weightings on this site follow that measurement rather than taste, and the parts of the picture a single-URL scan cannot see are stated rather than left out.

Primary sources