Citations and evidence grades
Every claim GeoHarness relies on, in its design, its GEO score and this documentation, is listed here with an evidence grade. The audit's findings cite these ids, and a test fails if any id is missing.
How are sources graded?
Grade A is peer-reviewed work or a primary platform specification; B a large-sample study with a published method; C a vendor log analysis or small sample; D commentary used for context only.
| Grade | Meaning |
|---|---|
| A | Peer-reviewed, or a primary platform specification/documentation |
| B | Large-sample industry study with published method |
| C | Vendor log analysis or small-sample measurement |
| D | Commentary or single observation; used for context only |
Research papers
| ID | Source | Grade | Supports |
|---|---|---|---|
aggarwal-2024 | Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024. https://arxiv.org/abs/2311.09735 (accessed 2026-11-19) | A | Original GEO lab result: quotations, statistics and citations raise visibility inside a fixed retrieved context; keyword stuffing hurts. Scored as advisory only. |
cseo-bench-2025 | "C-SEO Bench: Does Conversational SEO Work?", NeurIPS 2025 Datasets & Benchmarks. https://arxiv.org/abs/2506.11097 (accessed 2026-11-19) | A | Most content-rewrite methods are ineffective; classic SEO is more reliable. Why GeoHarness never rewrites prose. |
sageo-2026 | "SAGEO Arena", KDD 2026. https://arxiv.org/abs/2602.12187 (accessed 2026-11-19) | A | Structural information helps retrieval and reranking; several GEO methods hurt it. |
martinez-2026 | Martinez, survey of 45 GEO studies, arXiv 2607.14035 (2026). https://arxiv.org/abs/2607.14035 (accessed 2026-11-19) | B (preprint) | No technique shows a stable cross-platform causal effect; relevance and position are the reproducible levers. Basis for weighting access and extractability above content heuristics. |
what-gets-cited-2026 | "What Gets Cited", arXiv 2605.25517 (2026). https://arxiv.org/abs/2605.25517 (accessed 2026-11-19) | B (preprint) | Relevance and list position dominate; explicit price and a recent timestamp help; formatting-only edits barely matter. Basis for the visible-date check. |
htmlrag-2025 | Tan et al., "HtmlRAG", WWW 2025. https://arxiv.org/abs/2411.02959 (accessed 2026-11-19) | A | Keeping structure (headings, tables) beats flattening to plain text. Why markdown twins keep tables and headings. |
cormack-2009 | Cormack, Clarke and Büttcher, "Reciprocal Rank Fusion outperforms Condorcet and individual rank learning methods", SIGIR 2009. https://doi.org/10.1145/1571941.1572114 (accessed 2026-11-19) | A | RRF with k = 60, used for LatticeDB hybrid search. |
robertson-2009 | Robertson and Zaragoza, "The Probabilistic Relevance Framework: BM25 and Beyond", Foundations and Trends in IR, 2009. https://doi.org/10.1561/1500000019 (accessed 2026-11-19) | A | BM25 scoring and the k1/b constants used by LatticeDB's lexical index. |
Industry studies
| ID | Source | Grade | Supports |
|---|---|---|---|
ahrefs-schema-2026 | Ahrefs, "We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved", May 2026. https://ahrefs.com/blog/schema-ai-citations/ (accessed 2026-11-19) | B | JSON-LD gave no significant AI-citation lift. GeoHarness emits minimal JSON-LD for classic search only. |
ahrefs-llmstxt-2026 | Ahrefs, "We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read", June 2026. https://ahrefs.com/blog/llmstxt-study/ (accessed 2026-11-19) | B | llms.txt is rarely fetched, mostly by coding agents. Generated because it is free, scored as speculative. |
seranking-llmstxt-2026 | SE Ranking, "LLMs.txt: Why Brands Rely On It and Why It Doesn't Work", 2026. https://seranking.com/blog/llms-txt/ (accessed 2026-11-19) | B | No correlation between llms.txt and AI citations across ~300K domains. |
peec-reading-budget-2026 | Peec AI, "Deep Research reading budget", June 2026. https://docs.peec.ai/research/deep-research-reading-budget.md (accessed 2026-11-19) | C | First read covers ~78% of content on pages with fewer than 20 links and ~33% with 60 or more; ~5.7K-character first window. Basis for the link-density and front-loading checks (advisory). |
peec-how-deep-research-reads-2026 | Peec AI, "How Deep Research reads your site", 2026. https://docs.peec.ai/research/how-deep-research-reads-your-site.md (accessed 2026-11-19) | C | Agents read text only and re-read after an in-page match. Basis for the empty-SPA-shell check. |
indig-2026 | Kevin Indig, "The science of how AI picks its sources", Growth Memo, Feb 2026. https://www.growth-memo.com/p/the-science-of-how-ai-picks-its-sources (accessed 2026-11-19, paywalled) | C | 44.2% of ChatGPT citations come from the first 30% of a page. Sometimes misattributed to SparkToro. |
checkly-2026 | Checkly, "State of AI agent content negotiation", Feb 2026. https://www.checklyhq.com/blog/state-of-ai-agent-content-negotation/ (accessed 2026-11-19) | C | Which agents send Accept: text/markdown (Claude Code, Cursor, OpenCode) and which do not. |
evilmartians-2026 | Evil Martians, "Which AI actually reads your site: two months of LLM traffic measured", 2026. https://evilmartians.com/chronicles/which-ai-actually-reads-your-site-two-months-of-llm-traffic-measured (accessed 2026-11-19) | C | Claude Code received markdown 88% of the time; ChatGPT-User 0.7%. |
Platform documentation and specifications
| ID | Source | Grade | Supports |
|---|---|---|---|
google-ai-guide-2026 | Google Search Central, "AI features and your website" optimization guide, May 2026. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide (accessed 2026-11-19) | A | Generative features rest on core Search eligibility; markdown/AI text files are neither required nor used by Google Search. |
google-spam-policies | Google Search Central, "Spam policies: cloaking". https://developers.google.com/search/docs/essentials/spam-policies (accessed 2026-11-19) | A | Why GeoHarness negotiates on Accept, never on search-crawler user agents, and checks markdown/HTML equivalence. |
llmstxt-spec | llms.txt proposal. https://llmstxt.org/ (accessed 2026-11-19) | A (spec) | llms.txt format; per-page .md convention; rel="alternate" type="text/markdown". |
cf-content-signals | Cloudflare, "Content Signals Policy". https://blog.cloudflare.com/content-signals-policy/ (accessed 2026-11-19) | A (spec) | Content-Signal: search, ai-input, ai-train in robots.txt and headers. |
cf-markdown-for-agents | Cloudflare docs, "Markdown for Agents". https://developers.cloudflare.com/fundamentals/reference/markdown-for-agents/ (accessed 2026-11-19) | A | Plan availability (Pro+), headers (Vary: Accept, Content-Signal), 6 MiB limit; GeoHarness reproduces this for free. |
cf-transform-rules | Cloudflare docs, Transform Rules availability and URL Rewrite parameters. https://developers.cloudflare.com/rules/transform/ and https://developers.cloudflare.com/rules/transform/url-rewrite/reference/parameters/ (accessed 2026-11-19) | A | 10 free rules, concat and wildcard_replace on all plans, no regex on Free. |
cf-rules-functions | Cloudflare docs, Rules language functions and header fields. https://developers.cloudflare.com/ruleset-engine/rules-language/functions/ (accessed 2026-11-19) | A | Matching http.request.headers["accept"]. |
cf-phases | Cloudflare docs, Ruleset Engine phases list. https://developers.cloudflare.com/ruleset-engine/reference/phases-list/ (accessed 2026-11-19) | A | Transform runs before cache. |
cf-workers-static-assets | Cloudflare docs, Workers static assets billing, routing, headers. https://developers.cloudflare.com/workers/static-assets/ (accessed 2026-11-19) | A | Free unlimited asset requests, run_worker_first, _headers (100 rules). |
cf-workers-limits | Cloudflare docs, Workers platform limits. https://developers.cloudflare.com/workers/platform/limits/ (accessed 2026-11-19) | A | 100k requests/day, 10 ms CPU, 128 MB memory on Free. |
cf-pages-limits | Cloudflare docs, Pages limits and Functions pricing. https://developers.cloudflare.com/pages/platform/limits/ (accessed 2026-11-19) | A | Pages Functions share the Workers quota; fail-open option. |
cf-workers-ai | Cloudflare docs, Workers AI pricing and OpenAI-compatible embeddings. https://developers.cloudflare.com/workers-ai/ (accessed 2026-11-19) | A | 10k free neurons/day; embedding models and dimensions. |
cf-bot-fight-mode | Cloudflare docs, Bot Fight Mode and verified bots. https://developers.cloudflare.com/bots/get-started/bot-fight-mode/ (accessed 2026-11-19) | A | Why GeoHarness turns it off and verifies AI bots get HTTP 200. |
cf-managed-robots | Cloudflare docs, managed robots.txt and AI crawler policies. https://developers.cloudflare.com/bots/additional-configurations/managed-robots-txt/ (accessed 2026-11-19) | A | AI crawler policy defaults. |
cf-speed | Cloudflare docs, Early Hints, Speed Brain, Tiered Cache, Compression Rules. https://developers.cloudflare.com/speed/ (accessed 2026-11-19) | A | Free speed settings GeoHarness enables. |
cf-api-permissions | Cloudflare docs, API token permissions. https://developers.cloudflare.com/fundamentals/api/reference/permissions/ (accessed 2026-11-19) | A | cloudflare-permissions.json. |
indexnow | IndexNow documentation. https://www.indexnow.org/documentation (accessed 2026-11-19) | A (spec) | Key file and submission endpoint. |
openai-bots | OpenAI, "Overview of OpenAI crawlers". https://developers.openai.com/api/docs/bots (accessed 2026-11-19) | A | GPTBot, OAI-SearchBot, ChatGPT-User tokens. |
anthropic-bots | Anthropic, "Does Anthropic crawl data from the web". https://support.claude.com/en/articles/8896518-what-is-claudebot (accessed 2026-11-19) | A | ClaudeBot, Claude-SearchBot, Claude-User tokens. |
perplexity-bots | Perplexity, "Perplexity crawlers". https://docs.perplexity.ai/guides/bots (accessed 2026-11-19) | A | PerplexityBot, Perplexity-User tokens. |
qdrant-query-api | Qdrant docs, hybrid queries, sparse vectors, inference. https://qdrant.tech/documentation/concepts/hybrid-queries/ (accessed 2026-11-19) | A | Request shapes LatticeDB's text, sparse and fusion queries follow. |
Tools and prior art
| ID | Source | Grade | Supports |
|---|---|---|---|
markdown-new | markdown.new. https://markdown.new/ (accessed 2026-11-19) | D | Prior art: three-tier HTML-to-markdown conversion, x-markdown-tokens, vary: accept. |
geo-new | GEO.new. https://geo.new/ (accessed 2026-11-19) | D | Prior art: category weights for a GEO audit (AI Citability 22%, Content/E-E-A-T 22%, Brand/Entity 20%, AI Access 15%, Schema 13%, Platform 8%). |
lighthouse-ci | Google, Lighthouse CI. https://github.com/GoogleChrome/lighthouse-ci (accessed 2026-11-19) | A | The Lighthouse gate. |
htmd | htmd crate. https://crates.io/crates/htmd (accessed 2026-11-19) | A | HTML-to-markdown engine. |
lol-html | Cloudflare lol_html. https://github.com/cloudflare/lol-html (accessed 2026-11-19) | A | Streaming head injection and link rewriting. |
Rejected sources
Cited in research GeoHarness was given, checked, and not relied on.
| Claim | Why rejected |
|---|---|
| "Bajemon and Rochet (Sept 2026)" re-evaluation of GEO tactics | No such paper could be found. |
| "SparkToro, January 2026: 44.2% of citations from the first 30% of a page" | The figure is Kevin Indig's (see indig-2026), not SparkToro's. |