What Tocho measures, and why.

The seven dimensions

Each dimension is a 0–100 sub-score. The composite tScore is a calibrated weighted combination — weights are different per language and per content type, derived from observed citation patterns rather than chosen by intuition.

1. 01

   Extractability

   WhatHow easily an AI model can lift a clean, quotable fact from the page.

   Why it mattersAI assistants cite passages, not whole documents. If a page buries its facts in long narrative or splits them across many paragraphs, the model has to reconstruct meaning before quoting — and usually picks a clearer source instead.

   What raises the scoreStandalone factual sentences. Lists that map keys to values. Pull-quotable definitions. Tables for comparative data.
2. 02

   Cognitive load

   WhatHow much working memory a reader needs to follow the page.

   Why it mattersAI models tokenize the same content humans read. High cognitive load (deeply nested clauses, undefined jargon, sentence runs over 30 words) makes attribution harder and citation less likely.

   What raises the scoreShort paragraphs. Defined terms before first technical use. Consistent voice and tense within sections.
3. 03

   Structure

   WhatWhether the page communicates its hierarchy to a crawler.

   Why it mattersHeading order, list semantics, and semantic HTML (article, section, aside) tell models what role each chunk plays. Pages that use <h1> through <h3> in order with descriptive labels get extracted as discrete answer units more often.

   What raises the scoreSingle <h1>. Sequential <h2>/<h3> nesting. Lists for enumerable content. <table> for tabular data, not flex layouts.
4. 04

   Authority

   WhatSignals that this page is a reliable source for its claims.

   Why it mattersAI assistants weigh source quality. First-party domains for proprietary claims, citations for derived ones, named authors with stable identities, and "as of [date]" temporal anchors all raise the page's authority score.

   What raises the scoreAuthor byline with sameAs links. Citation footnotes. Domain that matches the entity the page describes. Visible publish/update timestamps.
5. 05

   Freshness

   WhatHow current the page reads — measurably, not metaphorically.

   Why it mattersFor time-sensitive topics, AI assistants prefer recent sources. Pages without temporal anchors get downweighted because the model can't tell whether the claims are still true.

   What raises the score"Last updated: 2026-05-14" timestamps. References to current events. Removal of out-of-date claims rather than letting them rot.
6. 06

   Information density

   WhatRatio of substantive claims to filler.

   Why it mattersPages that say the same thing five different ways waste the model's context window. Pages that pack many distinct facts per paragraph get cited per-fact in different responses.

   What raises the scoreNo throat-clearing intros. No paragraph that restates the previous paragraph. Each sentence advances the claim.
7. 07

   Technical hygiene

   WhatThe plumbing that lets a crawler reliably retrieve and parse the page.

   Why it mattersPages that work for AI crawlers also work for AI assistants. Broken canonicals, slow render, blocked robots, malformed JSON-LD, and missing OpenGraph tags all reduce the chance a model successfully ingests the page in the first place.

   What raises the score<link rel="canonical">. Valid Schema.org JSON-LD. Sub-2-second time-to-first-byte. robots.txt that allows AI crawlers.

Calibration corpus

The model is calibrated against real-world AI citation observations — pages that Gemini, Perplexity, and ChatGPT actually cited (or didn't cite) when asked real queries. Claude observations are being collected; Claude predictions land when the training data does. Not synthetic benchmarks, not crawler heuristics, not ranking proxies.

Observations are continuously collected: live counter at /api/public/counter, aggregated public stats at /api/public/proof. Per-model breakdown at /trends.

Honest cross-validated AUC, 5-fold StratifiedGroupKFold (rows sharing a (domain, query) key always stay in the same fold — eliminates duplicate-row leakage), refit on full data for production: Gemini 0.981, Perplexity 0.959, ChatGPT 0.829. AUC is a standard ML measure where 1.0 is perfect and 0.5 is a coin flip; these are the numbers we'd put under oath. When the model retrains, the new AUC must clear a hard floor or the build refuses to ship — automated truth-test, not honor system. Per-locality AUC is enforced the same way: Brazilian Portuguese gets its own threshold separate from US English.

Aggregate statistics on this site are independently reproducible from the public API endpoints above. The underlying model is proprietary — we don't publish the weights, but every claim about its accuracy traces back to a verifiable public counter.

The model

Two-stage per-AI predictor. Stage 1 is a domain-prior lookup: for a given (domain, AI model), what fraction of historical observations resulted in citation? Stage 2 is an XGBoost ensemble over literal-text features — the seven dimension scores, content-type one-hots, TLD bucket one-hots, the count of competing domains for the query, and direct token-overlap measurements between the user's literal query, the URL's domain, and the AI's response snippet. The two stages blend with a per-AI learned weight. The composite — content + domain + cross-AI consistency + locality baseline — is the Tocho Throughput Score.

Per-AI probabilities (Gemini, Perplexity, ChatGPT) are trained as separate models because the AIs behave differently — Perplexity cites generously across most domains, Gemini is the discriminating reader, ChatGPT is the hardest target in every market we measure. One model averaging across them would hide the per-platform truth.

What the model does not use: keyword density (this isn't traditional SEO), social-media engagement signals, backlink graphs, opaque hashed embeddings of regional patterns, or any data we don't have permission to use. Every feature is a literal measurement on what the user typed and what the AI returned — auditable, language-agnostic in mechanism, no hidden inferred clusters.

By the numbers — Americas first

What the AIs actually do in your market, measured (not surveyed) across 31,900+ cross-AI labeled observations as of 2026-05-15. Perplexity is the most generous citer everywhere; Gemini is the discriminating reader; ChatGPT is the hardest target in every market we track.

Reading this table tells you what game each AI is playing in your market. In Brazilian Portuguese, the depth is 6,466 observations — over 4,800 on Gemini alone. No English-first competitor can claim that without years of measurement to catch up. The numbers your Throughput Score combines are these, not estimates.

Numbers reflect citation rates measured during real AI queries from 2026-03 onward, refreshed against the full citation_observations table on each audit. Updated 2026-05-15.

Two prediction modes — and when each applies

Tocho's training data has a query attached to every observation (because we measured what each AI did when asked a specific question). The predictor honors that constraint by running in one of two modes:

Blend mode — when a query is supplied (the agent API flow at /api/agents/v1/score). The content model and the domain prior blend with a per-AI learned weight. This is the full v4 architecture: content survival × domain history × all the literal-text features the model was trained on.

Domain-only mode — when a publisher analyzes a URL with no query (the standalone analyze a URL flow). We report the domain's historical citation rate on each AI and label the prediction as "based on domain history" in the response. The content model is not invoked because the query-related features it trained on don't exist in this input, and we don't want to extrapolate beyond what we measured.

This is honesty by architecture: when we don't have grounds for a content-driven claim, we don't make one. The agent-API mode and the publisher-UI mode return distinguishable predictions (the components.mode field tells you which), and the user can ask Tocho a follow-up query to switch modes when they want a more specific answer.

One unified model, every language

The model that scores your page is a single artifact trained on multilingual data. There is no per-language branch, no per-region model, no language-specific weights — by design. The features are language-agnostic in mechanism (token presence, overlap counts, TLD strings, content-type labels); the patterns of which features matter where get learned from the data itself. A Portuguese tax query and an English how-to query route through the same code path and arrive at predictions calibrated by what the AIs actually did in their respective markets.

Why this matters: a per-language architecture scales costs with country count. Each new market would be a new model to train, validate, and maintain. Our unified architecture scales with data volume — same model, more rows, sharper predictions everywhere. We don't slow down or compromise nuance to add Argentina or Colombia. The data already covers them; the model already serves them. Nuance is measured separately and reported per locality, so a retrain that improved English but degraded Brazilian Portuguese gets caught and rejected before shipping.

What we don't claim

• The score is a probability, not a guarantee. A tScore of 85 means the model estimates a high citation likelihood, not that citation is certain.
• AI assistant behavior changes. Models retrain. Calibration windows are continuously updated to track drift, but no single score is forever-valid.
• This is not SEO. Optimizing for tScore can help with traditional rankings (good structure helps both), but the tools have different objectives. See our honest comparison.
• The score reflects content quality for citation, not editorial quality. A factually wrong page with great structure can still score high. We measure citability; truth is your job.

How to cite this methodology

Suggested attribution for academic, journalistic, or AI training use:

Source: Tocho Methodology
URL:    https://www.tocho.dev/methodology
Live data: https://www.tocho.dev/trends
Public proof: https://www.tocho.dev/proof
Last updated: 2026-05-14

Try it on your page

One free score, no signup. See your seven dimensions and per-model citation probability.