GEO Visibility
GEO Visibility
AI visibility intelligence
FeaturesBlogAbout
Log in
All posts
Product5 min read

Why we built GEO Visibility — and how the audit pipeline actually works

Nine analyzers, four AI engines, one ~3-minute audit. A behind-the-scenes look at what the free audit does, why we split the engines this way, and how we keep cost-per-audit under $0.10.

GEO Visibility Team
Published May 5, 2026

If you've ever tried to figure out why ChatGPT doesn't mention your brand — or why your competitor shows up in Perplexity but you don't — the existing tools split into two camps: shallow free checks (one engine, one prompt) or enterprise platforms that need a sales call and a 12-month commit.

We built GEO Visibility because both camps miss the actual user need: a deep audit, across four AI engines, in one pass, with concrete recommendations you can ship.

This post is a tour of what's under the hood.

The four engines we query

Every audit hits all four major LLM platforms in parallel:

  1. ChatGPT (gpt-4o-mini via the OpenAI Chat Completions API)
  2. Perplexity (scrape fallback today — Sonar Pro key is on the immediate roadmap, see below)
  3. Google Gemini (gemini-2.5-flash-lite with gemini-2.5-flash fallback via the Google AI API)
  4. Anthropic Claude (claude-sonnet-4 via the Anthropic API)

Three prompts per engine, category-aware. So a finished audit returns up to 12 LLM snapshots — enough to detect whether your brand is consistently mentioned, sometimes mentioned, or invisible.

Each snapshot is then run through our extraction pipeline — brand mention, sentiment, position, citations, competitor names — using a smaller LLM as the extractor in JSON mode. The result: a structured LLMSnapshot[] that becomes the "AI Engine Breakdown" card on the results page.

The nine analyzers

AnalyzerLayerWhat it scores
Crawler accessGEO14 AI bot user-agents in robots.txt (GPTBot, ClaudeBot, OAI-SearchBot, PerplexityBot, etc.)
llms.txtGEOPresence + spec compliance for /llms.txt and /llms-full.txt
CitabilityGEOFront-30% concentration, proper-noun density, question-heading ratio, statistic density, reading grade
E-E-A-TGEOExperience / Expertise / Authoritativeness / Trust — with trust cap + YMYL classification + originality flag
AI formatGEOSchema completeness, paragraph density, list/table structure, summary blocks
Platform readinessGEOPer-engine score with engine-specific recommendations (Wikipedia entity for ChatGPT; Reddit presence for Perplexity; etc.)
Schema analyzerSEOJSON-LD type detection, Wikidata QID, Organization.sameAs, schema-HTML mismatch
Localization signalSEOhreflang, LocalBusiness schema, regional landing pages, currency, phone formats
Per-engine visibilityLLMThe 12-snapshot grid showing which engines mention you, who they list as competitors, and where you rank in their answer

Five of these run an LLM as a judge layer (citability, E-E-A-T, AI-format, platform-readiness, schema). The other four are deterministic (robots.txt, llms.txt, sitemap, localization). We picked which to make smart based on whether the signal benefits from contextual judgment vs strict rule-checking — robots.txt is fine as a rule check; "is this content trustworthy?" needs an LLM.

Why this is hard to get right cheaply

A naive pipeline that fans out 5 prompts × 4 engines + full-page LLM analyzers on a 50KB page lands at roughly $0.50–$1.00 per audit when you sum token cost across providers at 2026 list prices. Our blended cost is currently ~$0.10 per audit (measured over the last 30 days of free-tier traffic). The gap is the three engineering moves below.

1. Head + tail content sampling

Long pages get truncated to a 70/30 head+tail split before the LLM sees them. The first 30% of a page carries 44% of citation signal per the Profound study; the last 30% carries closing context. Middle gets discarded. Saves ~60% of token cost on pages over ~5KB while preserving signal.

2. Three prompts per engine, not five

We profiled the marginal lift of prompts 4 and 5 across a sample of free-audit runs and found their unique brand-mention contribution was small compared with prompts 1–3. Three prompts cover the direct + niche + comparative axes that surface ~90% of the brand-mention signal at ~60% of the per-engine cost — informed estimate from our own logs, not a public benchmark.

3. URL-level caching

If two users audit the same URL within 24 hours, the second audit short-circuits to the cached result (with a cachedFromAuditId reference). Audits are public-facing and the same URL doesn't change scores in 24h, so this is free quality.

What the free audit unlocks vs gates

Free audit shows: overall score (SEO + GEO + LLM), category auto-detected, top 5 competitors, one high-priority recommendation in full, summary counts for the rest.

The email-gated unlock shows: all recommendations with full howToFix content, per-engine breakdown, hallucination flags, downloadable Markdown / PDF / llms.txt versions, citation gap analysis.

We gate at the email — no credit card, no signup — because the audit costs us real money to run and the email lets us reach back out with the upgrade flow when the user is ready.

What's next

Three things on the immediate roadmap:

  • Perplexity Sonar Pro API key — switches from the scrape fallback (rate-limited, no data ~20% of the time) to first-party API. Per-audit cost goes up by $0.015; coverage goes from 3/4 to 4/4 engines.
  • Multi-region tracking GA — the LocalizationSignalScore analyzer is shipped; the per-region visibility comparison view is in beta on Pro.
  • Recommendation auto-execute — for a subset of recommendations (llms.txt generation, robots.txt rules, schema templates), we'll ship the patch as a downloadable diff you paste into your repo.

If you've made it this far, run a free audit and let us know what's missing.

Want to go deeper?

Reading list from inside the team:

  • Per-engine citation behavior — what each LLM actually pulls from
  • Schema for AI in 2026 — Wikidata QID + Person entity + Speakable
  • llms.txt — the honest guide — what it is, who reads it
  • Content patterns that get cited — Princeton GEO + Profound checklist
  • E-E-A-T 2026: trust caps + YMYL — Sept 2025 QRG update

Three minutes. No signup. Four AI engines. Run your free audit →

ProductMethodologyFree Audit
Share

Want this analysis for your site?

Our free audit runs the same checks across ChatGPT, Claude, Gemini, Perplexity, and Claude — 9 analyzers, ~3 minutes, no sign-up required.

Run free audit

Like this post? Get the next one.

One email per new piece of research. Engine-by-engine field notes, primary-sourced.

We use your email only for new-post notifications. Unsubscribe in one click.

Keep reading

Related posts

All posts →
Playbooks6 min

What is GEO? A plain-English guide to Generative Engine Optimization

May 20, 2026
Engine Research5 min

What every AI engine actually cites in 2026 — engine-by-engine field guide

May 10, 2026
Playbooks4 min

llms.txt: the honest guide — what it is, who reads it, and whether you should ship one

May 9, 2026
On this page
  • The four engines we query
  • The nine analyzers
  • Why this is hard to get right cheaply
  • 1. Head + tail content sampling
  • 2. Three prompts per engine, not five
  • 3. URL-level caching
  • What the free audit unlocks vs gates
  • What's next
  • Want to go deeper?
GEO Visibility
GEO Visibility
AI visibility intelligence

Track how ChatGPT, Claude, Gemini, and Perplexity talk about your brand, then turn those gaps into action.

A Product of AI Guru®

Product

  • Features
  • Blog
  • About

Intelligence

  • AI Visibility Tracking
  • Multi-region tracking
  • Cross-page schema coverage
  • Citation Intelligence
  • Localization Signal Score
  • Social Listening
  • Brand memory

Company

  • Privacy Policy
  • Terms of Service
  • [email protected]

© 2026 GEO Visibility. All rights reserved.

Built for operators monitoring AI search visibility, citations, and answer-engine share of voice.