Measurement methodology

How We Measure AI Visibility

Every number in a Seenu Tech report comes from a documented, repeatable process. This page publishes that process in full: which engines we query, what each metric means, how scores are weighted, and the honest limits of what GEO measurement can claim.

See a Sample Audit Run a Free Scan

Monitored engines

Four engines, search-grounded only

We only count answers where the engine searched the live web. Each run records the exact model version used, because answers shift when models update.

ChatGPT Search

OpenAI Responses API with the web_search tool and your market as the search location. Answers without live search are rejected.

Perplexity

Sonar models, which ground every answer in live web results. Responses without citations are rejected as non-evidence.

Gemini

Google Search grounding. Only grounded responses with source metadata count toward scores.

Claude

Anthropic server-side web search. Answers that skipped search are rejected as non-evidence.

Answer metrics

What we count in every AI answer

Being named, being used as a source, and receiving an inline citation are three different outcomes. We track each one separately, per engine and per question.

Mention rate

Of all evidence-backed answer checks, how often the brand name appears in the answer text.

Source rate

How often the brand's own domain appears in the engine's cited-source list for the answer.

Citation rate

How often the brand's domain is written inline in the answer body itself - the strongest signal.

Average mention rank

When the brand appears, its position among all tracked brands in the answer (1 = recommended first).

Share of voice

The brand's share of every tracked-brand appearance across the full answer set, versus named competitors.

Journey coverage

Tracked questions mapped to six buyer-journey stages (understand, explore, compare, solution, recommend, decide) - with mention rate per stage.

Information accuracy

Brand-mentioning answers are checked claim-by-claim against the verified business profile. Contradictions are flagged for human review - never auto-published.

Site audit

The multi-page technical audit

Up to ten pages discovered from your sitemap and internal links, checked page by page. These are the actual items - not a vague 'schema review'.

robots.txt policy per AI crawler (GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Google-Extended, Googlebot, Bingbot, CCBot)
sitemap.xml presence, URL count, and lastmod freshness coverage
llms.txt, favicon, and site-name (WebSite schema / og:site_name) configuration
Per-page title and meta description presence and length (25-70 / 70-170 characters)
H1/H2 heading hierarchy, canonical URLs, noindex flags, duplicate titles
JSON-LD schema types per page, parse validity, and required-field completeness (name, address, dateModified, mainEntity...)
Schema expectations by page type - Organization and WebSite on the homepage, FAQPage on FAQ pages, Article on posts
Rendering mode per page (server-rendered, hybrid, client-side) and an estimated bot-readable text percentage
Visible word count, image alt-text coverage, FAQ and question-heading signals
Every issue classified critical / warning / notice, with new and resolved issues diffed against the previous round

Results roll up into five SEO axes (bot access, structured data, content quality, meta completeness, technical trust) and five GEO axes (AI bot access, render readability, answer structure, entity clarity, freshness). The GEO readiness score blends AI answer visibility (40%), evidence coverage (20%), website readiness (25%), and open-action readiness (15%).

Measurement principles

The rules every report follows

Evidence-backed only

An answer counts toward your scores only when the engine actually searched the live web and returned sources. Model-memory answers, mock data, and failed checks are excluded and reported as excluded.

Ranges, not single numbers

AI answers vary run to run. We aggregate recent rounds into a mean and range - "mention 22-25% across 3 runs" - and record the exact model versions used, so you can tell real change from measurement noise.

Your own baseline is the benchmark

There is no universal "good score." Every report compares your site against your own previous round: score deltas, new issues, resolved issues. Progress is the metric.

Human review before client delivery

AI-assisted findings - including accuracy flags - are drafts until a person verifies them. Nothing generated ships to a client without review.

Honest limits

What we deliberately do not claim

GEO is full of recycled statistics and overclaimed mechanics. These are the lines we hold.

We do not claim that ranking high in Google causes AI citations. Search position and AI visibility often diverge - that gap is exactly what we measure.
We treat structured data as hygiene, not a lever. Google states no special schema is required for AI features; schema helps engines understand your entity, and we audit it for that reason.
Bot-readable percentage is a heuristic based on raw-HTML text density - a strong signal, not a claim about any specific crawler's internals.
A single run is one sample. We never draw conclusions from one check, and neither should any vendor.

FAQ

Methodology questions

Which AI engines do you monitor?

Four engines, all in search-grounded mode: ChatGPT Search, Perplexity, Gemini, and Claude. We record the exact model version used on every run. We do not currently monitor Google AI Overviews or Bing Copilot, and we say so rather than counting them.

How often do you measure?

Monitoring runs weekly on managed plans. Each run stores every answer, source list, and model version, and the report shows movement against your previous rounds as a mean and range rather than a single snapshot.

What does the GEO readiness score mean?

It is a weighted blend of AI answer visibility (40%), evidence coverage (20%), website readiness from the live crawl (25%), and open-action readiness (15%). The score exists to track your own progress round over round - not to compare against other businesses.

Why were some checks excluded from my report?

If an engine answered from model memory instead of searching the live web, that answer cannot be verified and is excluded from customer-facing scores. The report shows how many checks were excluded and why.