What we tested
This is a controlled experiment using four purpose-built HTML snapshots, not a survey of customer websites. The first snapshot is a small guide with visible text, a labelled form and image alt text. The second adds noindex. The third removes the form label and image alt attribute. The fourth puts the guide inside a JavaScript string instead of the initial page body.
All runs use scoring version 2026-10-06.2, with AI enrichment disabled. Network and auxiliary inputs are held fixed. The source files and full checks are available in the public sample report.
| Snapshot | Overall | SEO | GEO | Accessibility |
|---|---|---|---|---|
| baseline | 84 | 72 | 65 | 100 |
| noindex | 81 | 55 | 65 | 100 |
| unlabelled | 77 | 68 | 65 | 55 |
| script-only | 61 | 40 | 19 | 76 |
Finding 1: an overall grade can hide noindex
Adding noindex changed SEO from 72 to 55, but the weighted overall score moved only from 84 to 81. GEO stayed at 65. The indexable check correctly reported a failure; the GEO content checks did not measure indexing.
Google requires a supporting page to be indexed and eligible for a snippet, and the site must be included in Search generative AI features. A content diagnostic cannot establish those account and indexing conditions. Read the noindex case study before treating a headline score as evidence that a page can appear in Search.
Finding 2: small markup changes can materially affect one dimension
Removing one form label and one image alt attribute reduced accessibility from 100 to 55. The formLabels check identified one unlabelled field; imgAlt identified one image missing an alt attribute. The overall score moved from 84 to 77.
These observations concern markup coverage. They do not establish complete WCAG compliance, screen-reader usability or a conversion lift. The accessibility case study includes the exact edit and recheck.
Finding 3: a static scan sees the response, not every rendered state
Moving the guide into a client script changed the static GEO score from 65 to 19. The initial body contained only the skip link. The scan could not see the article, image or form as page elements.
This does not prove that Google or every AI provider cannot render the page. We did not run a browser or measure citation behaviour. Compare initial HTML with rendered content before attributing a real access problem to JavaScript. A scanner also cannot report missing labels on a form it has never seen.
Method, controls and limits
Four constructed HTML snapshots scored by the production scoring functions in static mode, with the evaluation clock fixed to October 13, 2026 at 12:00 UTC. The baseline, noindex and unlabelled variants differ only in the stated HTML edits. The script-only variant moves the article into a client script. Page identity, response time (200 ms), compression, robots access and auxiliary context are fixed experiment inputs, not measurements of a live host. No browser, DNS, sitemap or real-user performance tests were run. No customer data or citation measurements are included. example.org is a reserved example domain, not the site tested.
The four variants are one experiment, not four independent website owners. Percentages from this sample would be misleading. For aggregate observations from real submissions, use the website quality benchmark, which documents its own cohort and exclusions. The benchmark and this controlled experiment answer different questions.
Reproduce the experiment from the repository with node scripts/build-public-audit-evidence.mjs. Check results.json and compare the source HTML. The number of checks varies with page features: 229 were emitted for the first three variants and 226 for the script-only variant. The engine has additional conditional checks; each run does not execute every possible check.
How to use these findings on your own page
Start with the intended purpose of the URL. Check indexability for public pages you want found, inspect the precise HTML evidence, fix confirmed defects and rerun the same URL. Use the website grader for the overview, the AI readiness checker for access and HTML checks, and the GEO checker for content and attribution diagnostics.
Measure actual results separately: Search Console impressions and clicks, observed AI citations and referrals, and completed product actions. A higher score alone does not establish any of these outcomes.