How the GEO Inspection Method Score Is Assigned
A 0 to 100 score computed by a fixed, public rule from the measured criteria of the GEO Inspection Method, always with coverage, measurement date, and method version.
- Code
- MET-BA-GEO8NOTA-001
- Version
- 1.1.0
- Status
- Current
- Effective
- October 19, 2026
- Next review
- April 17, 2027
- License
- CC BY 4.0
TECHNICAL NOTE: All information contained here reflects my experience and my working method, and you are solely responsible for any changes on your website or your clients’ websites. We assume liability exclusively under partnerships formalized by contract. If in doubt, visit our Legal Notice.
The 8-Layer GEO Inspection Method score is a 0 to 100 number computed by a fixed, public rule from 36 measured criteria, always published alongside coverage, measurement date, and method version, so anyone can reproduce the math from the evidence.
This page is part of the 8-Layer GEO Inspection Method. The method defines what to inspect in each layer. Here I show how each inspection turns into a number.
It applies to version 3.2.1 of the method. This page is at version 1.1.0, and the scoring rule for this version is named nota-1.1.0.
What the Score Is, and What It Is Not
The score is my engineering view turned into a yardstick. It was built from three things:
- reference documentation: the OECD and European Commission Joint Research Centre handbook on composite indicators (2008), for how to weight, aggregate, and test robustness; Google’s Lighthouse performance scoring, as a precedent for a score with public weights; and the technical specs that set thresholds for individual criteria (web.dev, RFC 9309, Google Search Central, schema.org);
- technical knowledge accumulated across inspections of my own website, from October 6 to October 16, 2026;
- my engineering judgment, as the person responsible for the method.
It is not an official Google standard or anyone else’s. It is not universal and is not endorsed by any big tech company. It is a working yardstick, versioned, with every choice explained.
Whenever a choice is mine (rather than coming from external documentation), I say so in the text.
And what is the score not?
It is not a guarantee of Google rankings or citations in any AI assistant. It measures whether machines receive, understand, index, and cite the website, with clear criteria, sources, and dates. A high technical score does not compel anyone to cite you.
The OECD handbook itself acknowledges that weights are, at bottom, value judgments. That is why I publish the weights, the reason behind each one, and a cross-check with equal weights.
The old yardstick was replaced. Before this method, I used the AI Vision Map yardstick, with 5 dimensions. It was replaced on October 16, 2026 by this 8-layer version. From it I kept only the core principles:
- the formula is fixed and deterministic;
- no AI calculates the score (AI may help explain a finding, never assign the number);
- coverage appears right next to the score.
The weights from the old yardstick do not apply to this method and are not published here.
Principles of the Score
Before the formula, the principles. The rules come directly from them.
- Missing data on a critical area must weigh on the score. Without measurement, a criterion scores zero.
- A high score never hides a severe flaw. That is why the cap (ceiling) exists and why every score shows its coverage.
- A score of 80 to 90 or above belongs to sites with solid data, applied testing, and above-average structure. It is not a participation trophy.
- No number without a measured source. I do not use estimates.
- The math is reproducible. With the evidence and this page, someone else arrives at the exact same number.
Think of it as a business owner:
What good is a 90 score if the bot meant to read your site gets blocked every single day?
The Formula, Step by Step
The score comes out in five steps, always in the same order: criterion status, layer score, global score, cap, and rounding.
1. Status of Each Criterion
| Status | Value | How it appears |
|---|---|---|
| Pass | 100 | Pass |
| Partial | 50 | Partial |
| Pending | 0 | Pending |
| Unmeasured | 0 (counts as Pending) | “to be measured” |
An unmeasured criterion counts as Pending (0) and displays as “to be measured”. The status comes from the published threshold for each criterion in the corresponding layer of the method.
2. Layer Score
Simple average of the criteria in that layer. Every criterion in a layer carries equal weight inside it.
layer score = sum of criterion values ÷ number of criteria
3. Global Score
calculated global score = sum of (layer weight × layer score) ÷ 100
Layer C8 has two blocks with their own weights: indexing (10) and AI citation (5). Each block uses the average of its criteria.
4. Cap
displayed global score = lower of (calculated global score) and (active cap ceiling)
Today, the only cap is the INS-01 cap, with a ceiling of 70. I explain it further down in the cap section.
5. Precision and Rounding
I compute everything at full precision. I round only on display, to one decimal place, half up (64.29 becomes 64.3; 3.75 becomes 3.8). I never round an intermediate value.
Because of that, the displayed per-layer contributions may not add up exactly to the displayed global score. The difference is rounding, not arithmetic.
Weights, Criteria, and Points per Criterion
There are 8 layers, weights summing to 100, and 36 criteria. Each code below links to the corresponding layer in the method, where the Pass, Partial, and Pending thresholds live. Since method version 3.2.0, every criterion has a fixed code (from C1.1 to C8.2) and its own anchor on the method page. A published code never changes its number: a retired criterion stays marked as deprecated.
| Code | Layer | Weight | % of score | Criteria | Points per criterion |
|---|---|---|---|---|---|
| C1 | CDN and firewall | 15 | 15% | 5 (C1.1 to C1.5) | 15 ÷ 5 = 3.00 |
| C2 | Server and performance | 10 | 10% | 3 (C2.1 to C2.3) | 10 ÷ 3 = 3.33 |
| C3 | HTML and accessibility | 10 | 10% | 4 (C3.1 to C3.4) | 10 ÷ 4 = 2.50 |
| C4 | Metadata | 10 | 10% | 5 (C4.1 to C4.5) | 10 ÷ 5 = 2.00 |
| C5 | Schema and entities | 15 | 15% | 7 (C5.1 to C5.7) | 15 ÷ 7 = 2.14 |
| C6 | Graph and internal links | 15 | 15% | 6 (C6.1 to C6.6) | 15 ÷ 6 = 2.50 |
| C7 | Access by AI agents | 10 | 10% | 4 (C7.1 to C7.4) | 10 ÷ 4 = 2.50 |
| C8 indexing | Indexing | 10 | 10% | 1 (C8.1) | 10.00 |
| C8 citation | AI citation | 5 | 5% | 1 (C8.2) | 5.00 |
| Total | 100 | 100% | 36 |
Points per criterion = layer weight ÷ number of criteria. That is how much a Pass criterion contributes to the global score (a Partial contributes half). In C5, for instance, each criterion is worth 15 ÷ 7 = 2.14 points.
I publish this column on purpose. It exposes a consequence of equal weighting within each layer: a criterion in a layer with fewer criteria carries more weight in the global score than one in a layer with many. The single indexing criterion is worth 10 points; each schema criterion is worth 2.14.
Criteria are defined by function, not by brand. “CDN and firewall” applies to any provider. Cloudflare is what I use, the best practice I recommend, and my case study; another CDN or server logs with the same data count equally.
Why These Weights?
C1, C5, C6, and C8 each carry 15 because they are the layers with the most direct impact on visibility:
- C1 decides whether the bot receives the page;
- C5 and C6 tell the machine who is who and how everything connects;
- C8 is the outcome (indexing and citation).
C2, C3, C4, and C7 each carry 10 because they are necessary conditions, yet on their own they move the outcome less.
This is my engineering choice, and it is versioned. The weights do not come from any external documentation or multi-site dataset calibration: they are my judgment as the owner of the method. Changing any weight requires a new version of this model, with the rationale logged in the history. The cross-check with equal weights below shows whether that choice changes the story, and the sensitivity test next shows how much each weight moves the rule.
Weight Sensitivity Test
Because each weight is my judgment, I show what happens when it moves. The test is straightforward: one layer weight rises or drops by 5 points, and the other seven adjust in proportion to their weights so the sum stays 100. In C8, the 5 points split just as they do today: two thirds for indexing and one third for citation.
With full access to all sources, the maximum possible score remains 100 under any variation: if every criterion passes, the weights do not change the result. So the test looks at two things that depend only on the rule, not on any website’s data:
- how much each criterion in the layer is worth in the global score;
- the ceiling of the public-data-only diagnostic, the product in which part of the criteria cannot be measured (maximum coverage of 71.0% and maximum score of 68.1 under the published weights; see below).
| Layer | Tested weight | Points per criterion in layer | Max score with public data only | Difference | Max coverage with public data only |
|---|---|---|---|---|---|
| Published weights | 68.1 | 71.0% | |||
| C1 | 20 | 4.00 | 65.3 | -2.8 | 68.0% |
| C1 | 10 | 2.00 | 70.9 | +2.8 | 74.0% |
| C2 | 15 | 5.00 | 68.9 | +0.8 | 72.6% |
| C2 | 5 | 1.67 | 67.2 | -0.8 | 69.4% |
| C3 | 15 | 3.75 | 69.9 | +1.8 | 72.6% |
| C3 | 5 | 1.25 | 66.3 | -1.8 | 69.4% |
| C4 | 15 | 3.00 | 68.7 | +0.7 | 71.5% |
| C4 | 5 | 1.00 | 67.4 | -0.7 | 70.5% |
| C5 | 20 | 2.86 | 70.0 | +1.9 | 72.7% |
| C5 | 10 | 1.43 | 66.2 | -1.9 | 69.3% |
| C6 | 20 | 3.33 | 69.0 | +0.9 | 71.7% |
| C6 | 10 | 1.67 | 67.2 | -0.9 | 70.3% |
| C7 | 15 | 3.75 | 67.8 | -0.3 | 71.2% |
| C7 | 5 | 1.25 | 68.4 | +0.3 | 70.8% |
| C8 | 20 | 13.33 (indexing) and 6.67 (citation) | 66.0 | -2.0 | 68.8% |
| C8 | 10 | 6.67 (indexing) and 3.33 (citation) | 70.1 | +2.0 | 73.2% |
How to read the table: on the C1 row with weight 20, the remaining weights scale down proportionally (C2, for example, moves from 10 to 9.41), each C1 criterion becomes worth 4.00 points, and the public diagnostic ceiling drops from 68.1 to 65.3. The difference comes from full-precision math and is rounded only for display.
What the test shows. No 5-point shift moves the public diagnostic ceiling by more than 2.8 points. The layers that move it most are the ones public data measures least (C1 and C8, which depend on logs and Search Console). That aligns with the rule: without access to the underlying sources, the score cannot rise. The test uses no website data, neither mine nor a client’s. The effect of weights on an actual score appears in the equal-weights cross-check published beside each score.
No Double Counting
Three explicit safeguards:
- redirects appear in C3 (the URL published in the sitemap returns 200 without redirecting, without soft 404, and without a 301 ending in 404) and in C6 (a body link passes through a 301 or hits a 404). Those are distinct objects: the published URL and the link pointing to it. There is no standalone redirect criterion;
- the “status 200” check for sitemap URLs lives strictly in C3. The consistent indexability criterion in C4 looks only at noindex and canonical;
- content without JavaScript (C3) and the machine-readable version (C7) are two distinct criteria by design. Search engines and most AI crawlers read the HTML page, and there is no public evidence that all of them read
llms.txtor Markdown versions. The machine-readable version complements; it does not replace.
Indexability and Yoast. On my site, Yoast SEO manages the sitemap: a page marked noindex drops out of the sitemap automatically. The consistent indexability criterion exists for what Yoast cannot see:
- noindex sent by the origin server or CDN in the
X-Robots-Tagheader; - noindex injected by the theme or another plugin;
- a URL in the sitemap whose canonical points to another page.
On a website using another SEO plugin, or built on another stack, I verify how that tool manages sitemaps and indexability before scoring.
Coverage: the Score Never Stands Alone
Every score reads like this: “score X, coverage Y%”.
coverage = sum of points of measured criteria ÷ 100
A criterion marked “to be measured” does not count toward coverage. By contrast, a measured criterion that landed on Pending does count: it was measured, it simply missed the threshold.
A score of 60 with 95% coverage tells one story. A score of 60 with 40% coverage tells a very different one.
The INS-01 Cap: Ceiling, Not Deduction
As long as INS-01 remains open, the displayed global score is capped at 70.
INS-01 is the block my origin (Hostinger’s LiteSpeed server behind Cloudflare) enforces against verified AI bots: 429 for GPTBot and meta-externalagent, with no exceptions allowed. The CDN is not returning those errors. Having a CDN alone solves nothing when the origin turns bots away.
While it stays open, AI bots cannot read the website reliably. That compromises three parts of the yardstick:
- C1, CDN and firewall: 15 points;
- C7, Access by AI agents: 10 points;
- C8, AI citation: 5 points.
ceiling = 100 - (15 + 10 + 5) = 70
In the standard calculation, INS-01 hits three layers: C1, C6, and C7.
- C1: criteria for 429 or 403 on verified bots and errors on discovery files (including
/okf/*); - C6: “nodes read by verified bots” (how many sitemap URLs each bot managed to read);
- C7: “AI bots actively crawling” (successful request volume per bot) and
llms.txtfor verified bots.
“Nodes read” and “bots actively crawling” draw from the same source (verified bot logs), yet measure different things: one measures reach (how many pages), the other measures volume (how many requests per bot). A bot can make many requests and still reach few pages. So both stay, and I declare the overlap right here.
An honest caveat about 70: the ceiling formula accounts for C1, C7, and AI citation. It does not include C6, even though “nodes read by verified bots” is also affected by INS-01. Because the ceiling is a cap rather than a deduction, that creates no extra penalty: the C6 criterion already pulls down the regular sum, and the cap only clips whatever exceeds 70.
Why a ceiling instead of a deduction?
INS-01 already zeroes out or reduces criteria inside the regular sum (in C1, C6, and C7). If I also subtracted 30 points, the same defect would be penalized twice. So the regular calculation stays untouched, and the cap only trims anything above 70.
And why have a cap if the math already penalizes it?
Because a weighted sum is compensatory: a deficit in one layer gets offset by strength in another. The OECD handbook describes that exact property of linear aggregation. Without the cap, high scores in schema and HTML could mask a site the bot cannot even read.
Exit condition: the cap lifts once a 7-day re-measurement, running the exact same query, shows the C1 verified-bot 429 criterion (code C1.1) within the Pass threshold. Partial is not enough: at Partial, bots still cannot read the website reliably.
In versions 1.0.0 and 1.1.0 of this model, the INS-01 cap is the only cap that exists, and it applies to my website.
Cross-Check with Equal Weights
Alongside the official score, I publish the score under equal weights: 12.5 for each of the 8 layers. In that calculation, the C8 score is the weighted average of its two blocks (2/3 for indexing and 1/3 for citation), so the layer stays one single layer.
equal-weights score = sum of (12.5 × layer score) ÷ 100
If both scores land close together, the weights are not changing the story. If they drift apart, I explain why. That is the sensitivity analysis the OECD handbook recommends to demonstrate the robustness of weight choices.
Full Example with Fictitious Numbers
Attention: every number in this section is an EXAMPLE. The website is fictitious, the statuses were chosen solely to illustrate the math, and nothing here is a measurement of my site or of any client.
| Layer (EXAMPLE) | Criteria status | Layer score | Weight | Contribution |
|---|---|---|---|---|
| C1 | 2 Pass, 2 Partial, 1 Pending | 300 ÷ 5 = 60.0 | 15 | 9.0 |
| C2 | 1 Pass, 1 Partial, 1 “to be measured” | 150 ÷ 3 = 50.0 | 10 | 5.0 |
| C3 | 3 Pass, 1 Partial | 350 ÷ 4 = 87.5 | 10 | 8.8 |
| C4 | 3 Pass, 1 Partial, 1 “to be measured” | 350 ÷ 5 = 70.0 | 10 | 7.0 |
| C5 | 4 Pass, 3 Partial | 550 ÷ 7 = 78.6 | 15 | 11.8 |
| C6 | 3 Pass, 2 Partial, 1 “to be measured” | 400 ÷ 6 = 66.7 | 15 | 10.0 |
| C7 | 2 Pass, 2 Partial | 300 ÷ 4 = 75.0 | 10 | 7.5 |
| C8 indexing | 1 Pass | 100.0 | 10 | 10.0 |
| C8 citation | 1 Partial | 50.0 | 5 | 2.5 |
Now, the step-by-step math (EXAMPLE):
- calculated global score: 9 + 5 + 8.75 + 7 + 11.7857.. + 10 + 7.5 + 10 + 2.5 = 71.5357.., displayed as 71.5;
- the displayed contributions add up to 71.6, not 71.5. That is the rounding in C3 (8.75 becomes 8.8) and C5 (11.7857 becomes 11.8) showing up. The official score comes from full-precision math;
- coverage: three criteria are “to be measured” (one in C2 worth 3.33; one in C4 worth 2.00; one in C6 worth 2.50). Coverage = 100 minus 7.8333.. = 92.1666.., displayed as 92.2%. The Pending in C1 was measured, so it counts toward coverage;
- cap: if this site were mine, with INS-01 open, the displayed score would be 70.0, because the math passed 70. If the calculation landed on 55.0, the displayed score would be 55.0; the cap leaves it untouched;
- cross-check with equal weights: C8 becomes (2 × 100 + 50) ÷ 3 = 83.33..; summing 12.5 × each layer score ÷ 100 gives 71.3839.., displayed as 71.4. It sits 0.1 away from the official score, so in this example the weights do not change the story.
Output in the published format (EXAMPLE): “score 70.0 (math 71.5, INS-01 cap), coverage 92.2%, method 3.2.1, calculation model 1.1.0, data from YYYY-MM-DD to YYYY-MM-DD, measured on YYYY-MM-DD, valid until YYYY-MM-DD”.
Missing Capability Scores Zero
In this version, as in 1.0.0, no criterion drops out of the calculation as “not applicable”. Whoever lacks the capability, or does not grant access to it, keeps the criterion at Pending (0), “to be measured”, and that goes into the report as a finding.
- Equivalent source counts equally: since each criterion is defined by function, another CDN, another firewall, or server access logs with reverse DNS bot verification count just as Cloudflare counts. The equivalence is logged in the evidence;
- without the capability: a website without a CDN is not exempt from C1; it lacks the protection C1 measures. I recommend, help configure, or carry out the Cloudflare setup;
- with no possible source: having zero logs of your own website is already a problem to solve.
Where the Evidence for Each Criterion Comes From
Every measurement becomes a record in the evidence file: criterion, origin, exact source, window, date and time with timezone, value (with numerator and denominator), status, caveat, and who collected it. The calculation reads only that file. Collection, validation, and calculation are separate stages.
Data Origins
- P, plugin: OKF Bundle Manager installed on the client’s WordPress site. Requires administrator privileges to install. It is not mandatory for measuring; it improves accuracy across C3 to C6 and generates the evidence file. It is a proprietary project, not for sale, with no public code;
- X, public check and external crawl: public collection run by us, with no client account access. Simulates a browser and bots with a declared user-agent. Needs client permission because it can trigger their firewall;
- A, read-only client account: CDN (Cloudflare, in my setup, or equivalent) or server logs, Search Console, hosting, and WordPress. GA4 does not feed the score;
- C, client-provided: raw exports sent by an authorized team member (dated CSV or JSON in the format I specify). Screenshots and summaries do not count. The evidence carries the tag “client-provided”, and I state plainly: I do not vouch for the integrity of data I did not collect at the source, because it may arrive trimmed or skewed;
- M, dated manual test: fixed prompts in AI assistants and Rich Results Test. On my own site, the citation prompts are the example list published in C8 of the method, following the logging rule for every answer (date, assistant, Cited or Not cited). The citation percentage is calculated over total answers in the round (prompt × assistant × day), not over prompts. For clients, upon engagement, prompts are written custom.
Collection Matrix
The matrix follows the exact stages of the inspection, with the fixed code and official name of each criterion matching the method page. Any criterion with origin A also accepts C (raw data sent by the client), carrying the tag and caveat.
| Stage | Layer | Criterion | Primary origin | Accepted equivalent source | Without source |
|---|---|---|---|---|---|
| 1 | C1 | C1.1 · Verified AI crawlers without 429 or 403 (cap criterion; INS-01) | A (CDN, verified bots) | Server logs with reverse DNS | Pending |
| 1 | C1 | C1.2 · Discovery files error-free for verified bots (home, robots.txt, llms.txt, sitemaps, and /okf/*) |
A (CDN by path) | Server logs | Pending |
| 1 | C1 | C1.3 · Fake bots filtered before origin | A (CDN and firewall) | Logs | Pending |
| 1 | C1 | C1.4 · 5xx at the edge | A (CDN) | Logs; uptime monitor | Pending |
| 1 | C1 | C1.5 · DNS and TLS: certificate valid for more than 14 days, HTTP and www/non-www redirecting to canonical HTTPS in 1 hop, no mixed content | X (public check) | ||
| 2 | C2 | C2.1 · Origin TTFB (p75) | X (PageSpeed Insights API) | Origin timing in CDN | |
| 2 | C2 | C2.2 · Core Web Vitals (p75 per template) | X (PSI: CrUX field or declared lab) | ||
| 2 | C2 | C2.3 · Edge caching | A (CDN) | X: cache headers | Partial at best with X alone |
| 3 | C3 | C3.1 · One H1 per page and proper hierarchy | X (crawl) | P | |
| 3 | C3 | C3.2 · Semantic HTML and accessibility | X (crawl and Lighthouse) | ||
| 3 | C3 | C3.3 · HTTP status of sitemap URLs: no soft 404 and no 301 ending in 404 | X (crawl) | P | |
| 3 | C3 | C3.4 · Main content in server HTML without depending on JavaScript | X (crawl with and without rendering) | ||
| 3 | C4 | C4.1 · Title and meta description present and unique | X (crawl) | P | |
| 3 | C4 | C4.2 · og:title and og:description consistent with title | X (crawl) | P | |
| 3 | C4 | C4.3 · Correct canonical | X (crawl) | P | |
| 3 | C4 | C4.4 · Consistent indexability: no sitemap URL with noindex (meta robots or X-Robots-Tag header) and all with canonical pointing to themselves |
X (crawl) | P | |
| 3 | C4 | C4.5 · Honest dates: “Updated on” and dateModified change only on genuine content edits |
P or WordPress read access (revision history) | X: public dateModified only flags suspicion |
Pending |
| 3 | C5 | C5.1 · JSON-LD present and parseable | X | P | |
| 3 | C5 | C5.2 · Person and Organization with unique @id and clean sameAs |
X | P | |
| 3 | C5 | C5.3 · Author linked by @id |
X | P | |
| 3 | C5 | C5.4 · Valid BreadcrumbList | X | P | |
| 3 | C5 | C5.5 · Duplicate entities | X | P | |
| 3 | C5 | C5.6 · Required fields per type | X and M (Rich Results Test) | P | |
| 3 | C5 | C5.7 · Complete Offers and Service | X and M (Rich Results Test) | P | |
| 3 | C6 | C6.1 · Orphan posts | X | P (more precise) | |
| 3 | C6 | C6.2 · Cluster linked to pillar hub | X and client’s pillar map | P (taxonomy) | Pending without map |
| 3 | C6 | C6.3 · Broken internal links and 301 chains | X | P | |
| 3 | C6 | C6.4 · Click depth (up to 3) | X | ||
| 3 | C6 | C6.5 · Parameterized URLs in the graph | X | P | |
| 3 | C6 | C6.6 · Nodes read by verified bots | A (CDN by URL) | Logs | Pending |
| 4 | C7 | C7.1 · robots.txt allows AI crawlers | X | ||
| 4 | C7 | C7.2 · llms.txt present and reachable by verified bots | A (CDN or logs) | X confirms existence only | Partial at best with X alone |
| 4 | C7 | C7.3 · AI bots actively crawling | A (CDN or logs) | Logs | Pending |
| 4 | C7 | C7.4 · Machine-readable version with canonical pointing to HTML and route listed in llms.txt (clean text of each page; in my case, OKF) | X | P | |
| 5 | C8 | C8.1 · Sitemap URLs indexed | A (Search Console, URL Inspection rotated) | None (“site:” is not a measurement) | Pending |
| 5 | C8 | C8.2 · Answers citing the brand or site | M (fixed prompts, monthly: on my site, the example list in C8 of the method; for clients, custom prompts upon engagement) | Pending if untested |
Stages: 1, CDN, firewall, DNS, and certificate; 2, server and origin; 3, WordPress and plugin (C3 to C6, with the optional plugin); 4, access by AI agents; 5, Search Console and citation.
With Public Data Alone: a Different Product
An inspection done with public data alone is not this method. It is a separate product, the initial prospecting diagnostic, under its own name and with coverage always printed next to the result. The professional inspection is the one with access to the sources.
To keep that diagnostic honest, I publish its ceiling under the rules on this page:
| Layer | Criteria measured with public data alone | Measured points | Minimum points lost |
|---|---|---|---|
| C1 | 1 of 5 (DNS and TLS) | 3.00 | |
| C2 | 3 of 3 (caching via headers only, Partial at best) | 10 | 10 ÷ 3 ÷ 2 = 1.67 |
| C3 | 4 of 4 | 10 | |
| C4 | 4 of 5 (no honest dates: revision history is not public) | 4 × 2.00 = 8 | |
| C5 | 7 of 7 | 15 | |
| C6 | 5 of 6 (no “nodes read by verified bots”) | 5 × 2.50 = 12.5 | |
| C7 | 3 of 4 (robots.txt; machine-readable version; llms.txt existence only, Partial at best) | 3 × 2.50 = 7.5 | 10 ÷ 4 ÷ 2 = 1.25 |
| C8 indexing | 0 of 1 | 0 | |
| C8 citation | 1 of 1 (manual test) | 5 |
- Maximum coverage: 3 + 10 + 10 + 8 + 15 + 12.5 + 7.5 + 0 + 5 = 71.0, displayed as 71.0%;
- maximum possible score: 71 minus 1.6667 (caching) minus 1.25 (llms.txt) = 68.0833, displayed as 68.1, even if every public check is spotless;
- without the client’s pillar map, the “cluster linked to pillar hub” criterion also stays Pending: 68.5% coverage and 65.6 maximum score.
That is deliberate. Without logs and without Search Console, I do not know whether the bot receives the page or whether it is in the index. The score cannot pretend to know.
With read-only access to the CDN (or logs) and Search Console, coverage is complete except for AI citation, which is always a manual test.
What Makes a Measurement Auditable
A measurement is auditable when someone else, with the same access, can repeat the collection and arrive at the same status. To make that possible, every published score carries:
- the measurement date and data period (the window for each query);
- the method version (today, 3.2.1) and the version of this scoring model (1.1.0, rule
nota-1.1.0ofnota-geo); - validity: the date through which the score remains valid (see “How the Score Is Published”);
- the sample: how many URLs, of which types, and whether the list came from the sitemap or the crawl, under the rule published in the method;
- the evidence for each criterion, keyed by criterion code in the method’s evidence template: source, verbatim query or command, stored raw result, value with numerator and denominator, status, and caveat;
- data origin, collected by me at the source or “client-provided”;
- coverage, right next to the score.
Without a record, the criterion does not enter the score. It stays “to be measured”.
Small counts (under 10 events) ship with a wide-margin caveat. And Cloudflare Free-plan data is sampled, which I declare next to every reading.
How the Score Is Published
The score is calculated outside WordPress and outside any plugin. No plugin calculates the score or stores weights in code.
The official calculation tool is nota-geo, a command-line program written in Python that runs outside WordPress. It has two commands:
validar(validate): checks the inspection evidence file before any math runs. If a criterion does not exist in the rule, if there are two records for the same criterion, if dates do not reconcile, or if a status exceeds what the data origin allows, it rejects the file and states which field failed;calcular(compute): applies the version’s rule and outputs the result file, containing the calculated score, the displayed score after the cap, coverage, the equal-weights cross-check, the list of criteria not at Pass, and validity.
Rule nota-1.1.0 is this page turned into a file: 36 criteria across 9 blocks (the 8 layers, with C8 split into indexing and AI citation), the weights in the table above, and the ceiling of 70 while INS-01 is open. The rule sits in a file separate from the program, and no weight, threshold, or ceiling is hardcoded. nota-geo reads only the evidence file and the rule file, and the same evidence file always yields the exact same output. Collection for each criterion follows the commands published in the method’s step-by-step tables, for anyone who wants to reproduce the measurement.
Observation note: where and when the nota-geo source code goes public is still to be decided. Until then, the calculation can be reproduced by hand using the formulas on this page.
Today, the pipeline works like this:
- the inspection generates the evidence file, keyed by criterion code;
nota-geovalidates that file and computes the result under rulenota-1.1.0;- the result enters OKF Bundle Manager, which only stores the finished number;
- Felipe C. Ferreira’s KG Map FCF reads that number and only displays it, with a link back to this page.
Right next to the score appear coverage, measurement date, method version, and the link “How the score is assigned”. Since October 20, 2026, KG Map FCF also shows validity and the list of criteria not at Pass.
The plugin never recalculates the score. It checks only the result format (required fields present, values between 0 and 100, date not in the future) and displays the number output by nota-geo.
Score validity. In a one-off audit, the score is valid for 90 days from the measurement date, or less if before then the CDN, theme, cache plugin, firewall rules, or robots.txt changes. Once expired, its number does not change: it ceases to hold as a current snapshot and requires a new inspection. Under monthly monitoring, as on my site, the score is re-measured every month.
Only my own website’s score is public. Client scores and data remain private and go out only with express written authorization, under LGPD (Brazil’s General Data Protection Law).
Current score of my website: 46.4, with 95% coverage. That is Measurement 1, measured on October 19, 2026 and published on October 20, 2026, computed by nota-geo under method 3.2.0 and rule nota-1.1.0. Measured prior to the October 20, 2026 cache and robots.txt changes; Measurement 2 captures their effect. Coverage does not reach 100% because C8.2, AI citation, is still “to be measured”. The figure appears on KG Map FCF alongside its date and versions. The score is not high, and the areas to improve sit in plain sight. The full result and timeline of all measurements live in GEO Readiness Score of felipecferreira.com.br. That is part of the method: anyone can check the math.
Cadence is at least once a month: one full inspection and one score per measurement, with the date visible. Because Cloudflare on the Free plan keeps 31 days of history (and queries at most 30 days at a time), collection runs weekly and accumulates, and the monthly score is computed over the accumulated window.
Complementary note: the score does not go into the page’s structured data as a rating (Rating or AggregateRating). That would be self-serving evaluation, which Google does not accept as valid markup for an organization’s own site.
Complementary Signals, Unscored
Some data appears in the report, yet carries no points:
- organic performance: clicks, impressions, and position. It stays outside the score until 3 months of clean data exist (with internal traffic excluded). A trend without a numeric rule and a clean baseline is not auditable;
- GA4, visits from AI assistants: GA4 measures human visits, which are an outcome. The score measures whether machines access, understand, index, and cite the site. GA4 enters as a recommendation and a signal;
- contact initiated (for example, a WhatsApp click);
- official AI reports: Search Console’s generative AI performance report and Bing Webmaster Tools’ AI Performance (including Citation Share). Both are collected in every inspection with date and window, and sit alongside the score. They measure outcome rather than readiness, and the “no access scores zero” rule would penalize a site without a Bing account over paperwork. So they carry no points.
Score Ranges: Hypothesis, Not Promise
My hypothesis is that a solid website without dedicated AI visibility work lands between 30 and 70. It is a hypothesis. So far, only one website has been measured with this yardstick: mine. I will test it once the first external inspections run.
Calibration across third-party websites uses only public criteria, without accessing anyone’s accounts without authorization.
I do not use letter grades or symbols (such as A, B, F, or +). A number paired with coverage is more honest than a letter grade.
Versions and Calibration
This page is at version 1.1.0 and applies to version 3.2.1 of the 8-Layer GEO Inspection Method. Whenever the method updates its version, this page states whether the scoring model stays unchanged or changes as well.
- Every published score cites the method version, model version, data window, and measurement date;
- every change enters the history below, with date and reason;
- calibration adjusts criterion thresholds, with a stated reason. It never tweaks weights to reach a desired answer;
- a weight change, if it ever happens, becomes a new version of the model with justification, and never applies retroactively to already published scores without a declared recalculation.
History of This Page
Revisions prior to publication were internal and lab test drafts. This documentation takes effect from version 1.0.0, published on October 17, 2026.
Next planned version, 2.0.0 (alongside method 4.0.0, no date set): the score will split into two separate readouts, the technical readiness score and the AI presence index, each with its own coverage, without adding one into the other.
- On October 19, 2026: Markdown (.md) download button enabled; CC BY 4.0 license declared.
- 1.1.0 (October 19, 2026): linked to method 3.2.0. Criteria given fixed codes (C1.1 to C8.2) in the weight table and collection matrix; weights declared as my judgment and versioned; sensitivity test varying each layer weight by 5 points; calculation moved into
nota-geo, a command-line tool outside WordPress withvalidarandcalcularcommands and rulenota-1.1.0(36 criteria, 9 blocks, ceiling of 70 while INS-01 is open), leaving the plugin purely for display; note on the first measurement in place of a placeholder result; KG Map FCF link at/kg-map-fcf/; criterion names in the collection matrix matched to the method page; score validity defined (90 days for one-off audits, or less upon changes to CDN, theme, cache plugin, firewall rules, orrobots.txt; re-measured monthly under ongoing monitoring); validity and sample added among the fields of every published score; official AI reports from Search Console and Bing treated as unscored signals; CC BY 4.0 license in prose; 1.0.0 release date written into history. Weights, formula, coverage, cross-check, and the INS-01 cap (ceiling of 70, exit only at Pass) did not change. Non-normative notes within the same version update the Markdown (.md) file; the downloaded filename carries the modification date (date_modified). - 1.0.0 (October 17, 2026): first public version, linked to method 3.1.0. Weights C1 15, C2 10, C3 10, C4 10, C5 15, C6 15, C7 10, and C8 15 (indexing 10 and AI citation 5); 36 criteria (C4 with 5, including honest dates); statuses 100, 50, and 0, with “to be measured” scoring 0; equal weight per criterion within each layer with published point values; mandatory coverage; rounding only on display; INS-01 cap as a 70 ceiling; equal-weights cross-check; organic performance and GA4 outside the score; public-data-only diagnostic ceiling at 71.0% coverage and 68.1 max score; cap exit only when the 429 criterion reaches Pass; no letter grades or symbols; manual provisional publication via OKF Bundle Manager, displayed on KG Map FCF. Replaces the 5-dimension AI Vision Map yardstick.
Limits of This Model
- Only one website measured so far; the score ranges are a hypothesis;
- part of the criterion thresholds come from external documentation (web.dev for TTFB and Core Web Vitals; RFC 9309 and Google for robots.txt; Google and schema.org for structured data). The rest is my judgment, declared in the method;
- on Cloudflare’s Free plan, since July 1, 2026, agents that sign their own requests (Web Bot Auth), such as ChatGPT’s, fall into the same bucket as verified bots. Separating the two categories requires an Enterprise plan with Bot Management. Verified-bot criteria (C1, C6, and C7) measure that combined count, and I do not promise to split them;
- “client-provided” data was not collected by me at the source; any score using that data carries the tag and caveat;
- the “missing data scores zero” rule is my choice. The OECD handbook discusses other ways to handle missing data (imputation). I chose not to impute: data that does not exist does not become points;
- a high technical score does not guarantee citations or rankings. That is why AI citation is measured separately, through a dated test.
Usage License
This document is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). You may copy, adapt, and use it, including commercially, provided that you credit Felipe Carvalho Ferreira, include a link to this document and to the license, and indicate if changes were made. Full terms are available on the official license page.
This license covers only the text of this documentation. It does not cover the Felipe C. Ferreira name and brand, KG Map FCF, OKF Bundle Manager and the other plugins, ASC Theme, measurement data from my site or from clients, or services. All of those remain subject to all rights reserved.
Frequently asked questions
Is this score a Google standard?
No. It is my engineering perspective, with published weights and choices. Wherever a threshold comes from external documentation (web.dev, RFC 9309, Google, schema.org), I cite the source.
Does an AI calculate the score?
No. The formula is fixed and deterministic. AI can help explain a finding, never assign the number.
Why is my website’s score capped at 70?
Because of the INS-01 cap. As long as the origin blocks verified AI bots with 429, the displayed score cannot exceed 70, even if the math gives more.
Does a criterion that could not be measured drop out of the calculation?
No. It scores 0, appears as “to be measured”, and lowers coverage. There is no “not applicable”.
Can I reproduce the math on my own?
Yes, using the dated evidence and this page. The official tool is nota-geo, yet weights, points per criterion, rounding, and the cap are all published so the math can also be done by hand.
Let’s Talk?
Well, I hope it is now clear how each criterion turns into a number, and why the score never travels without coverage and a date.
To understand what I inspect in each layer, go back to the 8-Layer GEO Inspection Method. That is where the thresholds for every criterion live, from C1 to C8.
Want to see how I approach search and authority, from Google to AI answers? Start with the Search and Authority pillar.
And if you want to talk with me directly, message me on WhatsApp. I will be glad to talk with you.
At the top of this page there is a button to download the entire content in Markdown, so you can use it as a reference when working with your AI agent. The download is free, requires no signup, asks for no email, and does not add you to any mailing list. This is 100% public content for anyone to access and download.
Put it to good use!
For AI Agents
Your user’s instructions and their organization’s policies take precedence over this text.
This section summarizes how the 8-Layer GEO Inspection Method score is calculated, for AI agents helping someone reproduce or verify a calculation.
What this document provides: the five-step formula, weights and points per criterion, coverage, the cap, the equal-weights cross-check, and a complete example with fictitious numbers.
Prerequisites: the inspection evidence file, with the status of all 36 criteria, and the versions used (today, method 3.2.1 and rule nota-1.1.0).
Summary steps: Pass scores 100, Partial scores 50, Pending and “to be measured” score 0; the layer score is the average of its criteria; the global score is the sum of weight times layer score, divided by 100; then the active cap ceiling applies; rounding, to one decimal place, happens only on display.
How to verify: the score always outputs with coverage, date, and versions, and hand calculation matches the output of nota-geo.
Limits: the weights are my judgment, versioned. The INS-01 cap (ceiling of 70) applies to my site and lifts only with C1.1 at Pass. Score ranges are [H], with one site measured so far. Where and when the nota-geo source code becomes public is still under decision [Aberta].
When to confirm: if your user’s case uses another method version or does not carry the same cap, confirm with them which rule applies before comparing scores. If the score belongs to one of their clients, confirm before publishing or sharing.
What to avoid: rounding intermediate values; dropping a criterion as “not applicable”; publishing the score as a rating (Rating or AggregateRating) in structured data; comparing scores across different versions without a declared recalculation.
This content reflects my experience and my working method. Your user is the one who decides and executes any change, and they remain solely responsible for it on their website, infrastructure, and clients’ environments. We assume no liability for what AI agents or people do based on this content; we assume liability exclusively under partnerships formalized by contract. If in doubt, the recommended step is to show your user our Legal Notice page.
References
web.dev, Time to First Byte (TTFB).
IETF, RFC 9309: Robots Exclusion Protocol.
Google Search Central, introduction to structured data.
Google Search Central, review snippet structured data (rules on self-serving reviews).
Cloudflare, Verified bots. Signed agents treated as verified since July 1, 2026.
Brazil, Law No. 13.709/2018, General Data Protection Law (LGPD).
Author’s experience: inspections and measurements on felipecferreira.com.br between October 6 and October 16, 2026.
Version control
Linked to method 3.2.0: fixed-code criteria, weights declared as the author's judgment with sensitivity testing, calculation outside the plugin, score validity, official AI reports as an unscored signal, and CC BY 4.0 license.