Ingested
Technical eligibility
Can engines fetch the page, extract usable text, and get past crawler blocks or paywalls? This is a gate. If it fails, the overall score is capped at 25 because nothing downstream matters.
Methodology
AEO Audit is a Dunware product powered by aiseo-audit, a deterministic CLI that scores how reusable a page is for generative engines. It does not call ChatGPT or any ranking API. Same URL, same HTML, same score.
POST /api/audit).aiseo-audit <url> --json.No API keys. No live citation check against ChatGPT. The measurement is local and repeatable. Engine behavior is not.
Traditional SEO optimizes for a ranked list of links. Answer Engine Optimization (also called GEO) optimizes for being cited inside a generated answer. Pages that lose the organic ranking game still show up in AI Overviews: measurement of deployed engines found that 53% of domains Google AI Overview consults sit outside the organic top 10, and 27% outside the top 100 (Kirsten et al., Findings of ACL 2026).
SAGEO Arena (KDD 2026) is the strongest recent case that retrieval, reranking, and generation must be scored separately. Body-only “GEO rewrites” often hurt retrieval. Structural fields help a page get found; the body supplies evidence that gets cited.
Ingested
Technical eligibility
Can engines fetch the page, extract usable text, and get past crawler blocks or paywalls? This is a gate. If it fails, the overall score is capped at 25 because nothing downstream matters.
Understood
Retrieval alignment
Do title, meta description, headings, entities, and JSON-LD carry the terms that get a page retrieved into an engine’s context?
Reused
Citation fitness
Once in context, does the page lead with an answer, cover the query, ground claims, and avoid hedged or off-topic content so it can be cited?
Provenance
Provenance
Authorship, organization identity, and attribution hygiene. Authority is a real engine objective; page-level proxies are still weakly validated, so this is scored cautiously.
The overall score is a research-informed heuristic. Weights and thresholds are expert-set, not fitted probabilities. aiseo-audit is explicit that stacking fixes does not stack citation odds, which is why reports rank the top items and tell you to re-measure.
Deployed engines are unstable. Kirsten et al. found 9–27% of some answer decisions changing within five minutes at temperature 0, and only 18% of Google AI Overview cited pages recurring across two months (versus 45% for organic search). Same URL, same audit score; no tool can promise the same citation.
Every factor in aiseo-audit 2.0 is labeled by how strong the evidence is. Diagnostics and experimental checks appear in the report but do not enter the score or the denominator.
Supported
Direct evidence in the factor’s stage, or a prerequisite (a page that cannot be fetched cannot be cited).
Conditional
Real evidence, but domain-, engine-, or quality-dependent. Scored with caveats or caps.
Heuristic
Plausible mechanism or adjacent evidence. Low weight, labeled honestly.
Diagnostic
Reported, not scored. Isolated causal tests were null, or the evidence is only observational.
Experimental
No outcome evidence. Informational only — for example llms.txt.
| Category | When | Measures |
|---|---|---|
| Content extractability | Always | Fetch, extraction, boilerplate, crawler access. |
| Structural alignment | Always | Whether title, meta, headings, and JSON-LD carry the page’s own salient terms. |
| Content structure for reuse | Always | Heading hierarchy. Most formatting checks are unscored diagnostics in 2.0. |
| Answerability | Always | Lead summary, definitions, direct answers, and steps. |
| Query alignment | With target queries | Term and aspect coverage. Scored by the weakest query, not the average. |
| Entity clarity | Always | Named entities, topic consistency, term-repetition balance. |
| Grounding signals | Always | Sources, attribution, numbers, hedged language. |
| Authority context | Always | Author, organization, structured data, freshness. |
| Product fit | Product pages | Price, specifications, comparison content. |
| Readability for compression | Always | Floors only. Polishing already-readable text is not rewarded. |
aiseo-audit 2.0 rescored its factors against ten peer-reviewed GEO papers (2025–2026), each reviewed from the primary source, plus the original Princeton GEO paper (KDD 2024). Marketing blogs, vendor benchmarks, and the llms.txt spec are not treated as outcome evidence.
KDD 2024
GEO: Generative Engine OptimizationFoundational result: once a page is already in an engine’s context, adding citations, quotations, or statistics can change how much of it appears in the answer. It does not prove those edits get a page crawled, retrieved, or cited in the wild. The “up to 40%” figure is a conditional visibility maximum, not a citation probability.
NeurIPS 2025
C-SEO Bench: Does Conversational SEO Work?Most fixed rewriting tricks failed as citation-rank methods. Statistics addition hurt ranking in 19 of 24 settings. Moving a document earlier in context beat every content rewrite. Qualifies GEO: word-count visibility is not the same as being cited first.
ICLR 2026
What Generative Search Engines Like (AutoGEO)Learned engine- and domain-specific preference rules beat a single global checklist. Supports generation-stage optimization in RAG setups where documents are already retrieved. A universal weighting scheme is unlikely to stay valid across engines.
KDD 2026
SAGEO ArenaEnd-to-end retrieval → rerank → generate. Structural fields (title, meta, headings, JSON-LD) raised retrieval hit rate about 22% vs body-only edits. Body text still supplies most cited evidence. Body-only GEO often hurt retrieval. Strongest case for scoring pipeline stages separately.
SIGIR 2026
What Gets Cited: Competitive GEO in AI Answer Engines252k causal pairwise trials. Strongest effects: on-topic content, earlier context position, explicit price on product pages, recent vs stale dates, specs, query-term coverage, evidence next to claims. Formatting-only changes (lists, tables, scatter vs organize) were weak or null.
ACL 2026
Think Before Writing (FeatGEO)The same rewrite can help a rough human page and hurt an already-fluent one. Supports floors instead of “more readability is always better,” and adaptive features instead of hard universal thresholds.
ACL 2026
Mind ReaderCovering latent user demands in the query outperforms cosmetic FAQ formatting. Question-shaped H2s are not a substitute for actually answering the query.
Findings of ACL 2026
IF-GEOEdits that help one query can hurt another. That is why this tool scores query alignment by the weakest query, not the average, when you supply target queries.
Findings of ACL 2026
From Experience to Skill (MAGEO)Engine-specific strategies and citation fidelity matter. Visibility that distorts or misattributes the source is not a win.
Findings of ACL 2026
Characterizing Web Search in the Age of Generative AIGoogle AI Overview consulted many domains outside the organic top 10 (53%) and top 100 (27%). Only 18% of AI Overview citations recurred across two months vs 45% for classic search. At temperature 0, 9–27% of some answer decisions flipped within five minutes. A one-shot score is not a citation probability.
ACL 2026 Industry
From Relevance to AuthorityProduction search A/B tests show authority is a real retrieval objective. It does not prove that a byline, About page, or Open Graph tag causes citations. Those remain cautious provenance checks.
These may still be good accessibility or classic SEO habits. They are not treated as universal citation predictors, and several are unscored diagnostics in 2.0:
Factor-to-paper mapping lives in aiseo-audit’s EVIDENCE.md and paper reviews. Last literature review cited there: 28 August 2026.
Audit a URL