How the RulerIQ Score is measured.

1. What we query (APIs with web search, not consumer apps)

RulerIQ queries each engine's official API with live web search enabled: OpenAI's Responses API (web_search), Perplexity's Agent API (/v1/agent), Google Gemini with Search Grounding (google_search), and Anthropic Claude with Web Search (web_search_20250305).

Consumer chat apps add user memory, location, and UI layers that vary per person. Querying the APIs provides a clean, unpersonalized baseline suitable for week-over-week tracking and competitor comparison.

Every check has a spending cap. A free check stops asking the engines when it reaches $0.90 (it can finish a call already in flight), and a daily limit on what we spend on our own keys pauses new checks when it is reached, so spend can't run away.

2. The four components (0–100)

On the site and in reports these are called Named, Cited, Facts right and Site readability. The formal names below are the ones used in machine-readable files.

• Named (Mention, 40%): Across non-branded discovery and comparison prompts, weighted by first-appearance rank (1st = 1.0, 2nd = 0.7, 3rd = 0.5, lower mention = 0.2, absent = 0.0), averaged over engines and repeat runs.
• Cited (Citation Authority, 30%): Of the pages the engines cite for your category's questions, the share (weighted by how often each is cited) that appear in answers that name your brand. We judge this from the engine's answer, not by reading the cited page.
• Facts right (Fact Accuracy, 20%): We ask each engine one branded question about your price and compare the dollar figures it states with those on your live pages. Correct = 1.0, unknown = 0.5, wrong = 0.0. Features are not compared yet. If your pricing is not in your HTML, an engine's stated price counts as wrong, so check the site check first.
• Site readability (Agent Readiness, 10%): Deterministic technical checklist (0–100): AI search crawlers allowed in robots.txt (35 pts), server-rendered HTML text (25 pts for 300 or more words, partial credit from 40), JSON-LD entity schema (15 pts) and FAQ schema (5 pts), valid /llms.txt (10 pts), and title/description/sitemap basics (10 pts). Every one of these is checked with plain code, not an AI: it reads your robots.txt rules for AI crawlers, the raw HTML your server sends, your structured data, your llms.txt, and your title, description and sitemap. Same input, same result.

3. Why every question runs several times (the margin of error)

Generative answers are probabilistic. One run is one draw from a changing answer. Several runs show how far the answers move, and that is the ±. Running a prompt once and graphing the result mistakes sampling noise for movement.

In the full check, RulerIQ runs every question 2 times per engine and computes a two-level 90% bootstrap confidence interval across question×engine cells and individual runs. Planned for paid plans: 3 runs per question.

How many times we ask Live

The free check asks each question 2 times on each engine. Paid plans ask 3 times, when your plan opens. Where runs disagree on whether an engine names you, we ask that question on that engine again, up to 5 times in all, inside the same cost cap. Every result shows how many runs agreed, like 3/5.

Why 4 engines?

We check ChatGPT, Perplexity, Gemini and Claude through each provider's official API. We ask each question several times on every engine, so we can state a margin of error. That matters more than counting more engines that are each asked once. Google AI Overviews is next, and it is planned, not live. Some trackers list engines that have no public API, which means scraping consumer apps. We use official APIs only.

4. How we report change over time With paid plans

• A week-over-week move counts as a real change only when it is bigger than the two weeks' combined margin of error (sqrt(band_prev² + band_curr²)). Smaller moves are labelled No real change.
• The trend says rising or falling only when the data supports it. Otherwise it says steady.
• The explanations are written by code from the numbers, not by an AI.

Verified fixesWith paid plans

Mark a fix as done. Later checks then confirm it only when the change is bigger than run-to-run noise, for example you are named in most runs on 2 checks in a row, or a wrong price is gone from every run on 2 checks. Otherwise the fix stays “improving”. We don't mark a fix as working off one lucky answer.

Fix list order: items seen consistently across runs and engines rank above one-off sightings.

5. What we use AI for

The AI we measure is the 4 engines: ChatGPT, Perplexity, Gemini and Claude. The Score, the margin of error, the fixes and every number are made by code from their answers.

With paid plans, each weekly check can carry a two- or three-sentence summary of what changed, for example which engine started naming you on which question. Our code works out the changes first; Claude Haiku only puts that list into words. The summary is labelled AI-written, and code rejects any text that contains a number, mentions the Score or the fixes, or gives advice, so those always come from code.

6. Why /llms.txt is only 1% of the Score, and what moves the rest

We have no evidence that publishing /llms.txt alone makes ChatGPT, Perplexity, Gemini, or Claude recommend a brand, so we weight it lightly. It is worth only 10 points inside our 10% Agent Readiness check (1% of the total Score).

The rest of the RulerIQ Score, and our paste-ready fixes, focus on four things that affect whether engines can find and accurately describe your brand:

• Search bots vs. training bots in robots.txt: Many teams enable Cloudflare's one-click “Block AI Bots” rule to block model training (GPTBot, CCBot) and accidentally block real-time search grounding crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot). RulerIQ audits search bots separately from training bots.
• Pricing and fact tables in server-rendered HTML: Retrieval systems can read plain server-rendered HTML <table> specifications and <dl> pricing definitions directly, and may miss content that only appears after JavaScript runs, such as pricing toggles.
• Structured data that states your facts (JSON-LD): Gives engines a clear source for your pricing and plans, which can reduce wrong answers about your brand.
• Pages to get onto (Reddit, G2, industry guides): Lists the pages engines cite for your category, in answers that don't name you, tagged by where they sit (review site, Reddit, publisher, your own site).

Those fixes (FAQ, fact table, JSON-LD and llms.txt) are built from facts on your own site, not written by an AI, so an AI can't make things up in them.

7. Connect your data With paid plans

None of this is live yet. Each source is read-only context next to the Score, and we report links between them as correlation, never as cause.

• Google Search Console: Questions built from your real searches, pages that rank on Google but are missing from AI answers, and clicks vs. impressions over 16 months.
• Google AI Overviews: Tracked next to the 4 engines and joined to your Search Console queries, so you see where an AI Overview shows and you aren't named.
• Core Web Vitals: Page speed as site-health context. Not part of the Score.
• AI referral visits: Visits that reach you from AI assistants, from Google Analytics.
• AI crawler fetches: Which of your pages AI crawlers actually fetch, from your server logs.
• Exports: CSV and Google Sheets or BigQuery, so it drops into Looker Studio.

8. What we do not claim

• We do not estimate “AI search volume” or impression counts; engines do not publish query volume.
• We do not guarantee that a fix will change an engine's answer on a specific date; you check again and we report what moved.
• We never report a site check alone as a full RulerIQ Score; it measures only Agent Readiness (10% of the Score).
• We do not claim our Score matches what any one person sees in the ChatGPT or Perplexity apps. We have not yet validated API answers against the consumer apps.