Methodology

Every number Answena produces is reproducible from the signals on your page, with formulas published below. No black box, no hidden weights. This is the full spec of the scoring engine — what we measure, how we combine it, and which peer-reviewed papers each signal is grounded in.

1. The three-layer scorer

1. Hard filters — kill-switches that cap the final score: empty body, thin content (< 150 words), blocked robots, no schema at all, no headings, soft-404 pattern, JavaScript-only rendering.
2. 10 weighted clusters — each a 0..1 number produced by engine/components.js. Weights sum to 1. Current baseline + learned blend is shown at /api/learn/weights.
3. Penalty multiplier — logistic gate 1 / (1 + exp(1.5 · (totalSeverity − 1.2))). Multiple small penalties compound; a single severe penalty (e.g. broken schema + thin content) clamps the score hard.

2. The 10 clusters

3. Readability (multi-language)

We dispatch the published per-language formula — coefficients do not transfer across languages.

4. Princeton GEO tactics

We audit your page against the 9 content interventions measured by Aggarwal et al., KDD 2024. Each tactic has a published citation-visibility uplift. Your audit card shows the tactics you have applied, the tactics you haven't, and a priority ranking uplift × gap — i.e. the highest-leverage unclaimed wins first.

• +40.6% Cite authoritative sources (research, standards, press)
• +36.2% Add quotable standalone sentences with statistics
• +30.0% Add numeric data / tables
• +24.0% Target plain-English fluency (Flesch ≥ 60)
• +22.7% Authoritative / decisive voice (no hedging)
• +22.2% Easy-to-understand grade level
• +15.5% Lexical diversity (Yule's K < 200)
• +12.3% Named methodology or technical terminology
• −22.0% Keyword stuffing — penalty, not a bonus

5. Per-platform weights

After the global composite, we re-weight the 10 clusters to produce 5 platform-specific scores — because ChatGPT, Claude, Gemini, Perplexity and Google AIO retrieve differently. Example: Perplexity weights citationPotential and trustReliability higher (sources shown inline); Google AIO weights structureReadability and entityAuthority higher (Knowledge-Graph-tight retrieval).

Full weight matrices: engine/platformScores.js.

6. Self-consistency probing

With ANTHROPIC_API_KEY set, we actually ask the LLMs. The probe runs each query at 3 temperatures ([0, 0.3, 0.7]) and takes the median outcome — a pattern from Wang et al.'s self-consistency decoding (ICLR 2023) and Google AI Mode's 2025 Query Fan-Out architecture. This cuts single-call variance (which can be ±15%) to roughly ±3% on the final citation rate.

Probes are cached for 6 hours by {domain, query, temperature} hash, so a re-scan doesn't re-spend API credits.

7. Ground truth & calibration

Every 6 hours a background sweep runs a controlled seed corpus (~37 curated URLs across tiers and sectors) through the full scorer and records the real citation rate from LLM probes. The corpus is appended to data/groundtruth-YYYY-MM.jsonl.

Against that corpus we report: Spearman ρ (rank correlation between our score and citation rate), Pearson r (linear fit), AUC-ROC (ability to separate cited from non-cited at ≥30% threshold), and NDCG@10 (top-10 ranking match). Live numbers: /validation.html.

8. Meta-learner (auto-retrain)

The 10 cluster weights aren't fixed. An L2-regularised logistic regression continuously learns from the ground-truth corpus (engine/metaLearner.js). New weights are only accepted if Spearman and AUC hold on the validation set — regressions trigger an automatic rollback to the previous snapshot.

Each accepted version is stored in data/weights/weights_YYYY_Qn_<timestamp>.json, so historical calibration can be diffed and any version can be restored via POST /api/admin/retrain/rollback.

9. Schema depth (JSON-LD completeness)

Presence-only scoring can't tell a stub {"@type":"FAQPage"} from one with 10 Q&A items and full acceptedAnswer.text. Our schemaDepth module measures completeness per entity (filled / expected) with type-specific bonuses (FAQPage needs answer text on ≥80% of items, HowTo needs step text on ≥80% of steps, BreadcrumbList needs ≥2 levels), and a graph-connectivity bonus when @id / @graph stitches entities.

10. What we deliberately don't do

• No Domain Authority / DR emulation. LLM retrieval ignores PageRank-style metrics; we don't include them.
• No keyword density targets. The keyword-stuffing signal is a penalty, not a goal.
• No backlink farms or off-page gaming. We score what the page actually contains.
• No hidden weights. If a number appears in your report, it traces back to a signal in engine/pageAnalyzer.js, a formula in engine/components.js, and a weight in engine/metaLearner.js.

11. References

• Aggarwal, P. et al. GEO: Generative Engine Optimization. KDD 2024. arXiv:2311.09735
• Wang, X. et al. Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR 2023.
• Guo, C. et al. On Calibration of Modern Neural Networks. ICML 2017.
• Björnsson, C.H. Readability in Different Languages. 1968. (LIX)
• Amstad, T. Wie verständlich sind unsere Zeitungen? Universität Zürich 1978.
• Bamberger, R., Vanecek, E. Lesen–Verstehen–Lernen–Schreiben. 1984. (Wiener Sachtextformel)
• Oborneva, I.V. Автоматизированная оценка сложности учебных текстов. 2006.
• Al-Tamimi, A. et al. OSMAN — A Novel Arabic Readability Metric. LREC 2014.
• Ateşman, E. Türkçede okunabilirliğin ölçülmesi. Ankara Üniversitesi TÖMER Dil Dergisi 1997.
• Bezirci, B., Yılmaz, A.E. A software library for measurement of readability of texts and a new readability metric for Turkish. 2010.

Source: github.com/siyahtonu/gec · Last updated with every deploy · Calibration refreshed every 6h