How every number is computed

How prompts are chosen

A prompt is one buyer question, run as written against every engine on your plan. The prompt set is the sample; everything else is measured on it.

Suggested set

When a brand is added, AI Rankly proposes a starting set of prompts from the brand's category, its products and the competitors it names. You review, edit, add and remove; nothing is tracked that you have not seen.

Unbranded and branded

Each prompt is classified as unbranded (the buyer has not named anyone: "best CRM for a small sales team") or branded (it names you or a rival). Share of voice and the headline visibility score are computed on unbranded prompts only; branded prompts feed reputation and comparison views.

Allowance

The allowance counts distinct prompts, not runs: 100 on Starter, 300 on Professional, custom on Enterprise. Prompts are shared across brands rather than allocated per brand.

Not a statistical sample

The prompt set is curated, not drawn at random from real query logs, and we do not claim it represents everything buyers ask. It is a fixed panel: the same questions, asked the same way, every day, so movement over time is comparable.

How engines are queried

AI Rankly queries 7 answer engines: ChatGPT, Perplexity, Gemini, Google AI Overviews, Google AI Mode, Claude, Grok. Which of them run for you depends on the plan.

Daily cadence

Every tracked prompt re-runs on every engine the plan covers once a day. A Professional plan across five engines produces 1,500 captured answers a day from 300 prompts.

Fresh sessions

Each run is a new, unauthenticated session with no conversation history, so one prompt's answer cannot colour the next. Where an engine offers a live web-search mode, that is the mode we query, because it is the one that cites sources.

Full answer stored

The complete answer text is stored word for word, with the engine, the prompt, the timestamp and every URL the engine cited. Every number on the dashboard is computed from that stored text and can be opened back to it.

Gaps are gaps

If an engine is unavailable or rate-limits us on a given day, that result is recorded as missing for that engine. It is never estimated, interpolated, or carried forward from the previous run.

Markets

Prompts run for a country and language pair. Plans without multi-market tracking run the primary market only.

How each metric is computed

All metrics are computed per prompt-and-engine pair, then aggregated over the look-back window you are viewing (30, 60 or 90 days by default).

Mention

Our own models read each stored answer and decide whether it genuinely refers to your brand rather than a similarly named company, product or person. A mention is a yes on that judgement. The same check runs for every tracked competitor.

Mention rate

The percentage of prompt-and-engine runs in the window in which you were mentioned. Reported overall, per engine, per prompt, and split by unbranded and branded prompts. The headline figure is the unbranded rate.

Average position

When an answer lists or ranks options, your position among the brands named, averaged over the runs where you appear. First is 1. An answer that names you without ordering anything contributes a mention but no position.

Sentiment

Each mention is scored positive, neutral or negative on how the answer describes you, from the passage itself. The reputation score is the balance of those across the window.

Share of voice

On unbranded prompts only: the share of brand mentions in the category's answers that are yours, against the competitors you track. If an answer names three of the four brands you track and you are one of them, you take a third of that answer's voice.

Citations

Every URL an engine cites is recorded and grouped by domain and by source type (your own site, a competitor's site, social and community, earned media, institutions, PR wires, and everything else). Citation share of voice is your domain's share of the citations that went to any tracked vendor.

Visibility score

A single 0 to 100 composite for the window, built from unbranded mention rate (how often you appear), average position (how prominently) and sentiment (how favourably). Mention rate carries the most weight; an unmentioned brand scores 0 regardless of the other two. It is reported per engine and overall.

How to read the numbers

AI answers vary between runs. The method is designed to make trends reliable, not to make any single day precise.

Daily runs are a series, not a verdict

One day's movement on one prompt is usually noise. The reads worth acting on are a trend held over two or more weeks, a shift visible across several prompts at once, or a change on more than one engine.

No significance claims

We do not claim statistical significance for a single daily run or for a panel of 100 or 300 prompts. What the cadence buys you is a longer series and a consistent method.

Benchmarks are anonymised

Category benchmarks compare your score with other brands measured in the same category. They are shown only where enough brands exist to meet a k-anonymity threshold; in a thin category you see no benchmark rather than a fragile one.

Corrections take time

When you fix a wrong fact at its source, an engine has to re-crawl or re-index before its answers change. Days to weeks is realistic for a well-cited correction. Daily tracking is how you find out when it has landed.

How our published research is produced

Reports on the research page use exactly this pipeline, run on a category AI Rankly tracks for itself, with the numbers frozen at publication.

Same pipeline

Published figures come from the same daily runs, stored answers and citation extraction described above. There is no separate research crawler.

Frozen at publication

A report states its look-back window and the day it was exported. The numbers on that page do not change afterwards; a later snapshot is published as a new report with its own URL.

Our own brand is included

Where AI Rankly appears in a category we report on, its figures are shown alongside everyone else's, including when they are poor.

Frequently asked questions