Skip to content

/ Methodology

How we score AI visibility.

There's no Google Search Console for AI answers. No platform — OpenAI, Google, Anthropic, Perplexity — tells you why ChatGPT cited one page and skipped another. So our score isn't a measurement. It's a disciplined read of the signals research links to citation. It points you somewhere useful. It doesn't pretend to be the truth.

How to read thisDirectional, not definitiveBuilt from public researchChanges as the evidence does

The short version

There is no Search Console for AI answers. No platform, not OpenAI, Google, Anthropic or Perplexity, publishes why one page was cited and another skipped, so any score claiming to measure AI visibility is inferring. Ours is explicit about that: it reads the signals that peer-reviewed and causal research links to citation, entity resolution, content structure, E-E-A-T signals and crawler access, and returns a directional 0-100 with the evidence behind every check named inline. It tells you where to look. It does not tell you what ChatGPT will do next week, and a score that claimed otherwise would be selling certainty nobody has.

Read this first

This is a heuristic. Not a measurement.

SEO has ground truth: Google Search Console and Bing Webmaster Tools report what actually happened, straight from the source. AI visibility has none of that. The systems that decide citations don't expose an API, don't publish their factors, and change behavior every few weeks. Anyone selling you a definitive "AI visibility score" is selling certainty that doesn't exist. We won't.

No source of truth exists
Google and Bing hand SEOs real data. No AI platform hands you anything: no API, no documented factors, no framework. Everything past that point is inference from watching how the machines behave.
The number is directional
A higher score means you fixed things research links to getting cited. It does not mean you'll get cited. Correlation, measured on someone else's data, is the strongest claim anyone can honestly make.
It changes as the evidence does
The weights come from studies that are months old at best, on platforms that shift constantly. New research lands, we update the model. Today's score is today's best guess, not a constant.
We don't have the truth
Nobody does. We'd rather hand you an honest compass than a confident lie. So pair the score with the real test: run your queries in ChatGPT, Perplexity, and AI Overviews and see if you actually show up.

01

The framework behind the score — "The Stack"

The score sits on a four-layer model of how AI decides to surface a brand. We score the two layers you can measure from a site — L1 and L4. The other two live off-site or inside training data nobody can see, so we show them as context, not a number.

L1Entity Establishment
Scored — 40 pts
Can AI resolve you as a real entity before it retrieves anything? Schema, sameAs, Wikidata, directory and review presence, name consistency.
L2Entity Depth
Context only
What the model already knows about you from training. Opaque by definition — we surface proxy signals (press, author depth) and assign no score.
L3Category Citation
Manual guide
Whether AI recommends you for "best X" queries. It lives in third-party editorial sources, so we point you at the right ones instead of scoring it.
L4Informational Citation
Scored — 60 pts
Whether AI cites your content as a source: answer-shaped, the point up top, entity-dense, crawlable, fresh.

02

How we use it inside That SEO Agent

It's not a black box, and it's not a model guessing a number. It's a deterministic, rule-based read of what's on your page plus a couple of lookups. Here's exactly what runs.

One tool, on demand
Your AI calls ai_visibility_score on a URL. It fetches the page, checks Wikidata and (if configured) Google's Knowledge Graph, reads robots.txt and llms.txt, and returns L1 + L4 as a 0–100 score with every check shown.
Part of the audit, not a verdict of its own
It runs next to your real GSC, GA4, and PageSpeed data. The hard numbers stay the anchor. The visibility score is the directional layer on top — never a replacement for them.
Every check shows its work
Each signal is a pass/fail with points, a plain-English reason, and — where it matters — the study behind the threshold. You can see why the number is what it is, and disagree with it.
Actions, not a vanity number
The output ends with a ranked list of what to change next. The number is a starting line, not a trophy.

03

Where the numbers come from

We're upfront about the evidence because that's the whole point. Most weights trace to a research aggregation ("The Data Sieve" by Aaron Haynes / Loganix — 108 entries across 88 sources), anchored where we can by peer-reviewed work. Worth saying plainly: several headline figures rest on a single study each. Good signal. Not settled science.

Primary / peer-reviewed
Aggarwal et al. — GEO: Generative Engine Optimization (Princeton / Georgia Tech, KDD 2024)arxiv.org/abs/2311.09735
Ahrefs — We Tracked 1,885 Pages Adding Schema (causal study, May 2026)ahrefs.com/blog/schema-ai-citations
BehnamGhader et al. — LLM2Vec-Gen: answer-shaped embeddings (McGill NLP, Mar 2026)arxiv.org/abs/2603.10913
Wang & Sun — Retrieval interference across 35 LLMs (NYU / UVA, Jul 2025)arxiv.org/abs/2506.08184
Practitioner research
Kevin Indig / Gauge — The science of how AI picks its sources (1.2M ChatGPT citations)growth-memo.com
Search Engine Land — 44% of ChatGPT citations come from the first third of contentsearchengineland.com
Yext — 86% of AI citations come from brand-managed sources (6.8M citations, Oct 2025)yext.com
Aaron Haynes / Loganix — "The Data Sieve" (108 entries across 88 sources)loganix.com

We keep reconciling the weights against these sources. When a number is mis-stated or a better source shows up, we change the model. That's what "subject to change" actually means.

04

What this score is not

Not a guarantee
Improve the score and you improve factors tied to citation. You don't buy a citation. ChatGPT and AI Overviews don't take orders from us.
Not Google's or OpenAI's opinion
No platform supplies, reviews, or endorses this. It's our read of public research. Nothing more.
Not fixed
The weights move when the research and the platforms move. Re-run it over time. Don't tattoo one number on the wall.
Not a substitute for testing
The closest thing to ground truth you can get today is manual: ask the real questions in the real assistants and watch what gets cited.

Continue reading

← All researchGenerative Engine Optimization →Ahrefs schema/AI study →Google AI Optimization Guide →MCP setup →

Try it

Get a directional read on your AI visibility

Connect the MCP and ask your AI to run ai_visibility_score on any URL — every check and its reasoning, shown.

  • Never writes to your site
  • Diagnostic, not automation
  • Connects via MCP to Claude, ChatGPT & Cursor
See pricing

Read-only access. We never write to your site.