Skip to main content
Everything on your dashboard is built from one kind of record: a run.
A run is one prompt, sent to one engine, in one daily slot. That is the unit rekkal counts, the unit it bills, and the unit almost every figure is divided by.
Twelve prompts on three engines, refreshed daily, is 36 runs a day. Over a four-week window that is 1,008 runs — and 1,008 is the n you will see printed under your visibility figure.

​
The three measures

Each is a fraction, and the interesting half is the denominator.

​
Visibility

The headline number. If you ran 40 prompts last month and 9 of the answers named you, your visibility is 22.5% from 40 runs. It is denominated on every answer collected, including the ones that named nobody. That is what makes it comparable week to week: the denominator is your measurement programme, not whatever the engines happened to feel chatty about. An answer that names you five times counts once. Visibility measures how many answers mention you, not how loudly one of them does.

​
Share of voice

Your mentions as a fraction of the mentions across your tracked set — you plus the competitors you track. It answers a narrower question than visibility: not “am I visible” but “when this category comes up, how much of it is me”. Both halves count answers, not name-drops, for the same reason as above: an answer that repeats a brand five times is one observation, and counting it as five would make every confidence interval falsely narrow.
Share of voice and the competitor ranking are included on every plan, and the tracked set is unlimited on both. Those brands constrain the whole product: they are the denominator here, and the sources view calls a cited domain a competitor’s site only if it belongs to one of them. See Plans and limits.

​
Average position

When an answer presents a ranked list, rekkal records where you appear in it, and averages that over the answers that ranked you. Average position never gets a direction arrow, on purpose. rekkal shows this period’s mean and the previous period’s mean side by side, and says Change not testable between them. It never says “up 0.4 places”. Two reasons. First, the arithmetic: an average carries no confidence interval in rekkal’s data, so there is nothing to test one period against the other with, and a difference of means with no interval is not evidence of anything. Second, the evidence: rank order in AI answers is even less stable between runs than list membership. Treat average position as context, never as a KPI. The research is on the Methodology page.

​
Every figure carries its sample size

Under each figure, in the layout — not in a tooltip — is the number of answers it was measured over.
If you take one habit from this page: read the n before the percentage. A figure measured over 12 answers and one measured over 1,200 are not the same kind of fact, and only one of them is worth acting on this week.
Each measure also carries a reporting threshold, counted in its own unit. Below it, the figure is still shown — hiding your own data is not honest either — but it is labelled:
Below this measure’s reporting threshold of 30 runs — read the range, not the number.
The threshold defaults to 30 runs. For share of voice it is converted into answer-mentions and for average position into ranked answers, so the number quoted is always in the unit that measure is actually denominated in.

​
Every estimate carries a range

The percentage on the card is a point estimate from a sample. The bar underneath it is the confidence interval — the range the true value is plausibly in, given how many answers you actually collected. It is an exact binomial (Clopper–Pearson) interval, and it behaves the way intuition says it should: few answers, wide range; many answers, narrow range. The per-engine breakdown draws it the same way: the bar is the interval, and the tick inside it is the point estimate. A wide bar is not a bad result, it is a small sample.

​
Where the numbers come from

rekkal reads the answer a person without an account is shown: No account is used, and none exists — every question is asked signed out, in a fresh session that carries no cookies and no history from the previous one. Two of the three are read in a real browser session driven by rekkal. The Google surface is retrieved through SerpApi, a vendor that sells search-results retrieval and carries the legal indemnity for it.
These platforms do not tell anyone which model answered. A chat interface routes a signed-out session to a model of its choosing and publishes nothing about the choice, so there is no model version for rekkal to record. Every answer says “Not disclosed” in those words rather than showing a dash or a guess. One consequence is worth knowing: a week-over-week move can be a routing change rather than a visibility change, and no tool reading these surfaces can tell you which.They also decide for themselves whether to search the web. rekkal used to turn grounding on and could assert it; now it records what it observed — whether the answer cited sources — per answer, and reports the fraction per slice.And the same question asked twice returns different answers anyway. So if your CEO types one of your prompts into ChatGPT and does not see your brand, that is not evidence your report is wrong — it is one draw from a distribution rekkal has sampled dozens of times. This is exactly the problem Methodology exists to address, and the reason every figure here is printed with its n.

​
What rekkal does not claim

Position is reported as an average over the answers that ranked you, with no direction and no trend line. A single “your rank in ChatGPT is 4” would be a fiction.
rekkal measures answers and citations. It does not connect them to sessions, conversions or revenue, and does not read your analytics or server logs.
Not measured. A mention is counted whether the answer is flattering or not.