ANSWENA
// SCORING MODEL

Methodology

Three layers, in order: hard filters, ten weighted signals, penalty multiplier.

Final score = (signals × weights) × penalty. Same signals re-weighted for ChatGPT, Claude, Gemini, Perplexity, Google AIO.

Pipeline

  1. Hard filters. Pass/fail gates. Word count ≥80, robots not blocked, headings present, content visible without JS.
  2. Ten signals. Each scored 0–100. Weighted sum produces composite.
  3. Penalty. Logistic gate on detected issues. Multiple small penalties compound.

Signals

Same ten cards as the home page.

  • Topic alignment. Page content vs. target queries.
  • Intent coverage. Definition, mechanics, alternatives, follow-up.
  • Readability. Sentence length, paragraph rhythm. Per-language formula.
  • Trust signals. Sourced numeric claims. Confident phrasing.
  • Brand recognition. Schema, Wikipedia, external profiles.
  • Crawlability. Clean HTML, schema blocks, sitemap, AI bot access.
  • Quotability. Standalone lines, stats, tables, FAQ pairs.
  • Freshness. Last-modified date, current-year references.
  • External mentions. Reddit, GitHub, Wikipedia, YouTube, G2.
  • Action surface. Pricing, sign-up, API docs, server-rendered.

Languages

Per-language readability formulas. English, Turkish, German, Russian, Arabic.

Live probe

70% from signals. 30% from live API calls to ChatGPT, Claude, Gemini, Perplexity. Median across three temperatures. Cached 6h.

Calibration sources

Signal weights derived from published GEO research:

  • Sourced claims — +40% citation rate.
  • Standalone quotable lines — +36%.
  • Confident tone — +23%.
  • Keyword stuffing — −22%.

Source: Aggarwal et al., KDD 2024.

Per-platform vectors

Same signals, re-weighted per platform:

  • ChatGPT — Wikipedia, brand reputation.
  • Perplexity — freshness (82% under 30 days).
  • Google AIO — Reddit, schema.
  • Gemini — structure, schema.
  • Claude — depth, sources.

All five scores in every report.

Calibration

Ground truth corpus: 37 URLs. Re-runs every 6h. Spearman ρ, Pearson r, AUC-ROC, NDCG@10. Live numbers: /validation.

Auto-retrain

L2 logistic regression on ground truth. New weights accepted only if Spearman and AUC hold. Old versions retained.

Out of scope

  • Domain authority. PageRank-style metrics not used.
  • Keyword density. Stuffing is a penalty signal.
  • Backlink schemes. Off-page gaming not measured.
  • Hidden multipliers. All scores trace to a specific signal.

FAQ

Same as classic SEO?

Classic SEO targets ranked search results. This targets answer text inside LLM replies.

API keys required?

Signal scoring runs without keys. Keys enable the 30% live probe.

Guarantee #1 on ChatGPT?

No tool can. This identifies weak signals and ranks fixes by expected impact.