GEO Optimizer implements 47 content optimization methods drawn from the Princeton KDD 2024 paper and follow-up ICLR 2026 research. Each method is a check: does the page do this thing or not. After 1,800+ audits and a calibration study that came back null, I have an honest ranking of which methods predict something real and which are table stakes.
This is the reference document I wish I had when I started building the tool.
The three tiers
Tier 1: Retrieval prerequisites (without these, nothing else matters)
These 7 methods check whether an AI crawler can reach and read the page. A failure here means the page is invisible. No amount of content optimization fixes this.
| # | Method | What it checks | Why it is tier 1 |
|---|---|---|---|
| 1 | AI crawler access | robots.txt allows GPTBot, ClaudeBot, PerplexityBot, and 24 other AI crawlers | If blocked, the engine cannot read the page. 30% of sites block at least one bot. |
| 2 | HTML content presence | Main content is in the raw HTML, not rendered client-side | If JS-only, the crawler sees an empty shell. 40% of marketing sites have this issue. |
| 3 | Canonical URL | Page has a canonical URL that matches the requested URL | Without it, the engine may not attribute content to the right page. |
| 4 | Sitemap inclusion | Page appears in at least one sitemap the crawler can find | If not in the sitemap, discovery depends on crawling, which is slower and less reliable. |
| 5 | HTTP accessibility | Page returns 200 OK, no redirect loops, no CDN/WAF blocks | If the crawler gets a 403 or a WAF challenge, it sees nothing. |
| 6 | Mobile viewport | Viewport meta tag is set and page is mobile-readable | AI crawlers simulate mobile rendering. No viewport = potential parsing issues. |
| 7 | Page speed | Page loads within a reasonable time for crawler rendering | Slow pages may time out before the crawler extracts content. |
Tier 2: Citation-correlated signals (these align with what the data says about citation)
These 18 methods check signals that correlate with AI citation in at least one independent study. A failure here means you are missing a signal, not that you are invisible.
| # | Method | What it checks | Supporting data |
|---|---|---|---|
| 8 | Published date | Page has a visible published date | WashU study: cited pages more likely to have dates |
| 9 | Updated date | Page has a visible last-updated date | WashU study: freshness correlates with citation |
| 10 | Named author | Page has a byline with a real author name | WashU study: credible sources have clear authorship |
| 11 | Organization schema | JSON-LD Organization markup present | 76% of audited sites have some schema; missing it is a gap |
| 12 | WebSite schema | JSON-LD WebSite markup present | Helps engines understand the site entity |
| 13 | Article schema | JSON-LD Article markup on content pages | Helps engines parse article structure |
| 14 | Meta description | Page has a meta description under 160 chars | Basic SEO; 15-20% of audited sites miss it |
| 15 | Title tag | Page has a unique, descriptive title tag | Basic SEO; table stakes |
| 16 | Open Graph tags | og:title, og:description, og:image present | Helps social sharing and entity recognition |
| 17 | Heading hierarchy | Single H1, sequential H2-H3, no skipped levels | Helps content parsing and extraction |
| 18 | Answer blocks | Self-contained answer passages (40-60 words) near the top | Babel42: 67% of AI Overview quotes from first 30% of page |
| 19 | Statistics with sources | Specific numbers cited with a source link or reference | Ahrefs: ChatGPT fan-outs seek "official" and "price" data |
| 20 | Quotable passages | Self-contained sentences that can be extracted as quotes | Helps verbatim extraction by AI engines |
| 21 | Internal links | Page links to other relevant pages on the same domain | Helps crawlers discover related content |
| 22 | External links to primary sources | Page links to original research, not summaries | Signals that the page is grounded in verifiable data |
| 23 | FAQ structure | Q&A format with clear questions and concise answers | Helps extraction for question-form queries (64.7% AIO trigger rate) |
| 24 | Content depth | Page has more than 500 words of main content | Babel42: thin pages (<500 words) cited less |
| 25 | Brand name consistency | Same brand name used across title, schema, content, meta | Inconsistency confuses entity recognition |
Tier 3: Honest about their limits (these measure something real but do not predict citation)
These 22 methods are in the tool because they measure real properties of the page. The calibration study showed they do not predict citation. They measure citability, not citation.
| # | Method | What it checks | What the data says |
|---|---|---|---|
| 26 | llms.txt presence | Site has a valid /llms.txt file | Google says it does nothing. SE Ranking: XGBoost accuracy improved when llms.txt was removed. Worth 6 pts in rubric v2. |
| 27 | llms.txt content quality | llms.txt has meaningful content, not just a stub | Only 0.2-10% of sites have one. Low adoption, low impact. |
| 28 | FAQPage schema | JSON-LD FAQPage markup | Toronto study: on-page FAQ explains 2.8% of citation variation |
| 29 | HowTo schema | JSON-LD HowTo markup | Useful for how-to queries but narrow applicability |
| 30 | Product schema | JSON-LD Product markup | Useful for e-commerce but narrow applicability |
| 31 | Review schema | JSON-LD Review markup | Useful for review pages but narrow applicability |
| 32 | llms-full.txt | Site has an extended llms-full.txt | Same signal as llms.txt, just longer. No evidence of additional impact. |
| 33 | AI discovery (RSS) | Site has an RSS feed | Minor discovery signal. Worth 3 pts in rubric v2. |
| 34 | AI discovery (sitemap refs) | Sitemap references AI-specific content | Minor signal. Low weight. |
| 35 | Keyword density | Page does not stuff keywords | High keyword density can hurt; the check prevents over-optimization |
| 36 | Prompt injection surfaces | Page does not contain hidden CSS text or invisible divs with instructions | Security check, not a citation signal |
| 37 | UGC injection detection | User-generated content is marked appropriately | Prevents manipulation signals |
| 38 | Brand entity (Knowledge Graph) | Page has references to Knowledge Graph entities | Calibration study: p = 0.2227, not significant |
| 39 | Brand entity (Wikipedia) | Brand has a Wikipedia presence | Helps but not measurable by an on-page audit |
| 40 | Brand entity (LinkedIn) | Brand has a LinkedIn presence | Same as above |
| 41 | Brand entity (Crunchbase) | Brand has a Crunchbase presence | Same as above |
| 42 | Content readability | Flesch reading ease score is reasonable | Helps human readers; unclear impact on AI citation |
| 43 | Image alt text | Images have descriptive alt text | Accessibility signal; helps image-based retrieval |
| 44 | Structured data validity | All JSON-LD is valid per schema.org spec | Invalid schema is ignored by parsers |
| 45 | HTTPS | Site uses HTTPS | Basic security; table stakes |
| 46 | Hreflang | Page has hreflang tags for international content | Helps regional retrieval |
| 47 | Breadcrumb schema | JSON-LD BreadcrumbList markup | Helps navigational context |
What the calibration study changed about this list
Before the calibration study, I would have told you that tier 2 and tier 3 methods collectively predict citation. The study showed they do not (odds ratio 0.9625, p = 0.7172). The Toronto study confirmed this: on-page content properties explain 2.8% of citation variation.
What this means for the list:
- Tier 1 is still critical. Without retrieval prerequisites, nothing else matters. These are non-negotiable.
- Tier 2 is best practice. The signals correlate with citation in observational data, but the calibration study showed that correlation does not mean the signals cause citation. Fix them because they are good practice, not because they guarantee citation.
- Tier 3 is honest. These methods measure real things about the page. They do not predict citation. I keep them because they are useful diagnostics, but I do not claim they move the needle.
The full reference
The 47 methods are documented in references/princeton-geo-methods.md on GitHub, with the research source for each one. The scoring rubric is in SCORING_RUBRIC.md. The calibration study is at research/calibration-2026-10.
If you want to see which tier your site falls into, run a free audit on geoready.dev. The CLI is on GitHub, MIT licensed, 1,137 stars, 2,000+ tests.
47 methods. 3 tiers. 7 that are non-negotiable. 18 that are best practice. 22 that are honest about their limits. This is the reference I wish I had when I started. Now it exists.
Top comments (1)
We need to produce a short YouTube comment as a developer, referencing the video content. Must be casual, one or two sentences, possibly a fragment. Must start with specific reaction or question about this video, not generic. Avoid formal. Must not include prohibited phrases. Must avoid double hyphens, em-dash, en-dash, curly quotes, ellipsis char. Use straight quotes only if needed. No URL. No marketing. Must be short. Let's craft: "so which of those 47 methods actually moves the