How GEOall measures AI visibility

How often AI answer engines recommend a company when a buyer asks them a commercial question in the company's category (for example “What is the best CRM for a small B2B company?”). The public number is the AI Visibility Score (0–100). It is an automated measurement of AI answers at a point in time, not an endorsement, a quality rating or a professional ranking.

Every answer comes from the provider's official API, asked with a neutral system prompt (“You are a helpful assistant.”) and no other instructions. API answers can differ from the consumer apps (see Limitations).

Each category is scored on at least 5 buyer questions, written the way a buyer would ask an AI assistant, never naming a brand, and tagged with the buyer segment they represent so every kind of buyer in the category is covered. Local categories always name the city. Years are filled in when the question is asked, so questions never go stale. The full list:

In every weekly run, each question is asked 3 times per engine at the engine's default temperature, and answers are never served from a cache, so the score reflects how often a brand is recommended across varied answers rather than one snapshot. Answers are capped at 2,000 output tokens; an answer cut off at the cap is stored but never scored.

The free website scan is different: it asks one question once per engine (temperature 0, reused for 24 hours) and is labelled as a single snapshot.

A separate extractor model (gpt-4o-mini, strict JSON schema, temperature 0, extraction version x2) reads each plain-language answer and lists the brands it presents, with their sentiment. Deterministic checks then decide what counts: a brand must appear in the answer text, a link counts only if that URL is in the answer or the engine's citations, and position is the order in which brands first appear. The same extraction is used for leaderboards, Pro tracking, free scans and paid audits.

Brands are matched on their name, known aliases (for example Kit = ConvertKit) and their own domains. Descriptors such as “CRM” or “Sales Cloud” are ignored, but a different product never matches (Zoho Desk is not Zoho CRM).

Per engine, over the valid samples of the last 30 days under the current methodology version:

• Mention frequency × 60 — the share of answers that recommend the brand.
• Position, up to 25 — averaged over answers that recommend the brand (25 at #1, minus 3 per place, never below 0), then multiplied by mention frequency.
• Citation frequency × 10 — only when the engine returned web-search citations; otherwise these points are redistributed proportionally over the other parts.
• Sentiment — averaged over answers that recommend the brand (+5 positive, 0 neutral, -20 when the answer warns against it), then multiplied by mention frequency.

Each engine score is clamped to 0–100. The overall score is the average of the engines with at least one valid sample; an engine that was not tested or failed is left out, never counted as 0.

Coverage rule: a company gets a public score only if at least 2 engines produced valid answers for every active question in its category. Otherwise it shows “Insufficient data” and is not ranked. Rankings across categories only compare companies scored on every engine.

Ties: companies are ordered by displayed score, then mention frequency, citation frequency, average position and name. Equal displayed scores share a rank (shown as “T-3”).

A brand is cited when its own website appears in the engine's actual citation or grounding metadata: Perplexity's citations, the url_citation annotations of ChatGPT's web search, and Gemini's Google Search grounding sources. A website merely written into an answer without web search is not a citation; answers without web search have no citation data and are excluded from citation rates.

The free scan and the $299 audit report a GEO Readiness Score: 60% AI visibility from the live questions (the formula above) and 40% technical readiness. If no live answer is available, it is technical readiness only. Technical readiness (0–100):

Only crawlers that fetch pages for live answers and search affect the score: OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, Claude-SearchBot, Claude-User, Bingbot, Googlebot. Blocking model-training crawlers (GPTBot, Google-Extended, CCBot, ClaudeBot, anthropic-ai, Applebot-Extended) does not change live answers and is reported for information only. robots.txt is read per Google's specification. llms.txt is an emerging standard that no major engine has confirmed using, so it carries little weight.

Public leaderboards are re-scored every Monday at 05:00 UTC; the “Scored” date on a page is the date of that run, and rank changes compare one weekly run with the previous one. Extra runs started by our team collect data but publish nothing until the next weekly run. Pro tracking runs every Monday at 07:00 UTC.

• AI answers vary by user, location, account, conversation history and time; we measure a neutral, logged-out API request.
• API answers can differ from the ChatGPT, Gemini and Perplexity apps, which use their own system prompts, models and personalization.
• Samples are small: a few questions × a few answers per week. Treat small score differences as noise.
• Models and search indexes change without notice; the engines table records our configuration, not the providers' internals.

If you think a score, ranking or company detail is wrong, send it through Support → “Dispute or correct a ranking” with the company page URL. We review every request, reply within 5 business days, and record methodology corrections in the changelog below.

1. v2026.09.10 · 2026-09-27

   Position and sentiment points are weighted by mention frequency, so a brand recommended once in many answers no longer receives the full position points.

2. v2026.09.9 · 2026-09-27

   Matching also reads a firm given in parentheses after a person's name ("Robert A. Clifford (Clifford Law Offices)").

3. v2026.09.8 · 2026-09-27

   Extraction v x2: a recommended person is recorded with their firm ("Lawyer / Firm") so the firm gets credit; "Trial" (Trial Lawyers) and "Smart" (Smart CRM) are treated as descriptors in matching.

4. v2026.09.7 · 2026-09-27

   Matching reads names given as person / firm pairs ("Robert Clifford / Clifford Law Offices") part by part, found in the pre-launch spot check of stored answers; Power Rogers & Smith added as an alias of Power Rogers.

5. v2026.09.6 · 2026-09-26

   Audit technical score reweighted toward proven signals (AI search crawler access, indexability, server-rendered content, Organization schema with sameAs); training-crawler blocks are informational; robots.txt read per Google's spec (wildcards, longest match); sitemaps must be real XML; llms.txt treated as an emerging standard.

6. v2026.09.5 · 2026-09-26

   Brand matching uses each company's aliases and other domains (e.g. Kit = ConvertKit, Brevo = Sendinblue) and compares distinctive words only, so descriptors like CRM or Sales Cloud are ignored but other products (Zoho Desk vs Zoho CRM) never match; 5+ questions per category tagged by buyer segment.

7. v2026.09.4 · 2026-09-26

   Leaderboards sample every question 3 times per engine at the engine's default temperature (no answer cache) and score mention frequency; new score formula (mention 60, position 25, citation 10, sentiment +5/−20); coverage rule; tiebreaks; engines search the web; citations only from grounding metadata.

8. v2026.09.3 · 2026-09-26

   Engines answer in plain language; a separate extractor lists the brands; only brands and links really in the answer count.

9. v2026.09.2 · 2026-09-26

   Answers cut off at the token cap are stored but never scored.

10. v2026.09.1 · 2026-09-24

    Launch.