How RankX AI measures visibility

RankX AI measures AI visibility by asking. It sends each tracked prompt to each enabled AI assistant on a cadence, reads the answer, and records whether your brand was named, where it appeared, and which sources the answer cited. One prompt asked on one assistant on one run is one check, and every figure in the product is built from checks with a known verdict.

This page is the method. It exists because a visibility number with no stated method is a claim rather than a measurement, and because knowing the method is what lets you tell a real movement from noise.

The unit: one prompt, one assistant, one run

The atom of AI visibility in RankX AI is the (prompt x assistant x run) triple. Nothing is averaged before that point and nothing is inferred after it.

Three consequences follow immediately:

• A Website with 25 prompts on 5 assistants performs 125 checks per run, not 25. The same multiplication is what your credit spend follows, which is why credits and metering opens with it.
• Per-assistant figures are the real figures. An "overall" rate is a summary of five different measurements taken on five different systems that disagree with each other by design.
• A single check is not a measurement. Assistants are non-deterministic: asking the same question twice returns different brand lists more often than not. The reliable unit is a rate across a panel of prompts over several runs.

What RankX AI asks

A prompt is a natural-language question, written the way a buyer would ask it, attached to a topic cluster and tagged with a buyer stage: researching, comparing, or ready to buy.

RankX AI proposes an initial set during onboarding from your brand profile, competitor set and clusters, and you edit it. After that the set is yours to manage, and managing it matters: every active prompt is checked on every enabled assistant on every run, so the prompt list is both your measurement panel and your recurring cost.

A prompt has three states. Active runs in the scheduled check and is billable when checked. Inactive is paused, reversible, and costs nothing. Suggested is a proposal you have not accepted. Pausing is how you stop a prompt costing credits without losing its history, and nothing is ever deleted to achieve it.

Where RankX AI asks

Five AI assistants, on every plan, with no engine add-ons:

ChatGPT is the one every reader recognises and covers everyday consumer and business questions. Gemini is integrated with Google Search and Workspace, which makes it the one that matters most to brands whose buyers live in Google. Claude skews to nuanced research and document-heavy work, and is strong in B2B contexts. Perplexity is AI-native search with real-time citations, so it is the clearest view of which sources are being surfaced beside you. Grok has live web and X access, so it picks up trending and real-time mentions the others miss.

Google AI Overviews is deliberately not in that list. It is a different surface measured a different way, and folding it in would overstate your assistant coverage by a fifth. See citations, two kinds.

When RankX AI asks

Nightly is when RankX AI looks, not how often your Website runs. A Website runs on the cadence you choose, raised to a floor if your choice is faster than the floor allows. There are two floors and the higher one wins:

• Your plan's floor, which is 1 day on most plans. See plans and limits.
• A 7-day floor while you are on trial, on every plan and every check family. That is deliberate: a daily sweep of a full prompt list would consume more than the entire trial grant in the first few days and leave you with a single measurement and no credits. A 7-day floor buys a re-check, which is the thing that actually shows you a movement. See the free trial.

The scheduled run is designed to survive being interrupted. It works through prompts It works through prompts in bounded batches with a deadline, and if it runs out of time it records what it completed, marks the run partial, and dispatches the remainder rather than starting again from the top. A partial run is visible as partial, so a coverage gap is queryable rather than silent.

If the wallet runs dry mid-run, the run stops and says so. It does not report the prompts it never reached as absences.

What RankX AI records from an answer

For each check, RankX AI stores the answer and derives:

• Whether your brand was named. A tri-state: mentioned, not mentioned, or unknown. Unknown is used when the check could not produce a verdict, and it is never collapsed into "not mentioned".
• Where your brand appeared, when the answer contains a ranked list.
• Which competitors were named, from the competitor set you defined.
• Which sources the answer cited, which becomes AI Answer Citations.
• The answer text itself, which is what the AI Chat Feed shows you.

The four things RankX AI refuses to do

These are enforced in the product, not aspirations, and each exists because the opposite was a real defect:

1. Never report an absence of measurement as a zero. A rate divides by checks with a known verdict. If none has one, the rate is unknown, and unknown is what renders. Null is not zero is the whole page on this.
2. Never merge "we did not check" with "we checked and found nothing." Not in the top 20 and never checked are separate states. An AI Overview detected but not captured is "unknown", not "you were not mentioned".
3. Never show a score without its coverage. AI Readiness renders a score beside the fraction of checks that were actually assessed, and below a coverage floor it emits no score at all rather than a misleading one.
4. Never name the upstream vendor. Where a check fails, RankX AI reports which assistant failed and refunds it. It does not surface raw upstream error text, on any surface.

How to read a movement

Given the above, a defensible reading looks like this:

• Compare rates, not counts. A count moves when your prompt list moves.
• Compare per assistant. A drop on one assistant and a rise on another is two facts, and the average hides both.
• Carry the denominator. "Named in 12 of 40 checks" survives being quoted. "30%" does not.
• Give it several runs. Single-run movement on a single prompt is noise, and treating it as signal is the most common analytical error on this kind of data.
• Read the AI Chat Feed when a number surprises you. The raw answers usually explain the number in one read, because they show you which brands were named instead and what the answer was actually about.

Where to go next

• Metrics defines each figure the product shows.
• Null is not zero is the honesty contract in full.
• Citations, two kinds settles the AI Answer Citations against AI Overview Citations confusion.
• Credits and metering is the same arithmetic seen from the bill.