GEO score methodology and the Lighthouse gate
GeoHarness scores how well generative engines can reach, index and read a site. Most of the score rests on access and delivery checks backed by platform documentation; content heuristics are advisory because the research does not show they reliably move citations.
Why are the weights what they are?
A survey of 45 GEO studies finds no content technique with a stable cross-platform effect; relevance and position are the reproducible levers, and a page that cannot be fetched, indexed or extracted has neither. Rewrite-style tactics mostly fail under benchmark conditions. So 80 of 100 points sit on access, indexability and markdown delivery, and those checks are blocking: a failure stops the run whatever the score. The remaining 20 points are advisory.
| Category | Id | Weight |
|---|---|---|
| AI access | access | 30 |
| Indexability | indexability | 25 |
| Markdown delivery | delivery | 25 |
| Extractability (advisory) | extractability | 12 |
| Optional signals (advisory) | signals | 8 |
Which checks run, and what evidence backs each?
Blocking checks carry grade A or B evidence. Checks with live in their id run after deploy against the real site, with the Accept matrix and eight AI user agents.
| Check | What it verifies | Category | Severity | Evidence grade | Citations |
|---|---|---|---|---|---|
access.spa_shell | Page has server-rendered text | access | blocking | A, C | peec-how-deep-research-reads-2026, google-ai-guide-2026 |
access.robots_bots | robots.txt admits search and user agents | access | blocking | A | openai-bots, anthropic-bots, perplexity-bots, google-ai-guide-2026 |
access.live_ai_agents | AI user agents get HTTP 200, not a challenge | access | blocking | A | cf-bot-fight-mode, openai-bots, anthropic-bots, perplexity-bots |
index.title | Page has a <title> | indexability | blocking | A | google-ai-guide-2026 |
index.description | Page has a meta description | indexability | blocking | A | google-ai-guide-2026 |
index.canonical | Page has a canonical link | indexability | blocking | A | google-ai-guide-2026 |
index.broken_links | Internal links resolve | indexability | blocking | A | google-ai-guide-2026 |
index.live_build | Live site serves this build | indexability | blocking | A | cf-workers-static-assets |
delivery.twin | Markdown twin exists and matches the page | delivery | blocking | A, A (spec) | google-spam-policies, llmstxt-spec, htmlrag-2025 |
delivery.live_negotiation | Accept: text/markdown gets markdown with Vary: Accept | delivery | blocking | A, C | cf-markdown-for-agents, cf-transform-rules, checkly-2026 |
delivery.live_files | HTML pages, robots.txt, llms.txt and sitemap.xml are served correctly | delivery | blocking | A (spec) | cf-content-signals, llmstxt-spec |
extract.link_count | Few links (<20) so the first read covers the page | extractability | advisory | C | peec-reading-budget-2026 |
extract.front_loaded | Main content starts within the first ~5,700 characters | extractability | advisory | C | peec-reading-budget-2026 |
extract.entity_early | Entity named in the first 30% of the content | extractability | advisory | C | indig-2026 |
extract.question_headings | A heading is phrased as a question | extractability | advisory | A, B (preprint) | what-gets-cited-2026, sageo-2026 |
extract.visible_date | A date is visible in the content | extractability | advisory | B (preprint) | what-gets-cited-2026 |
extract.tables_kept | Tables survive into the twin | extractability | advisory | A | htmlrag-2025 |
signals.llms_txt | llms.txt is present and well-formed | signals | advisory | A (spec), B | llmstxt-spec, ahrefs-llmstxt-2026, seranking-llmstxt-2026 |
signals.json_ld | JSON-LD is present and valid | signals | advisory | A, B | ahrefs-schema-2026, google-ai-guide-2026 |
How is the score computed?
Each finding scores 0 to 1; a category's score is the mean of its findings, and the GEO score is the weighted mean over the categories that had findings. The build audit runs on the processed output; the live audit adds the post-deploy observations. geo-min-score sets the minimum for both.
What does the Lighthouse gate require?
- A score of 100 in performance, accessibility, best practices and SEO.
- The median of three runs per URL, because single Lighthouse runs vary by several points on the same page.
- URLs are the first
lighthouse-max-urlsentries of the generated sitemap, served from the build output before anything is deployed. - Reports are uploaded to Lighthouse CI's temporary public storage and linked from the job summary; do not run the gate on content that must stay private.
lighthouse: falseskips the gate and says so in the summary.