Citation position (AvgPosition)

When an AI assistant answers with a list of sources attached, those sources have an order. Citation position is where a URL sits in that order. Attensira reports it as AvgPosition: the mean index of a URL within a model's own citation list, computed per source domain.

It answers one narrow question — when this domain is cited, how far down the list does it sit — and it is routinely asked to answer a much bigger one it cannot.

What it is not

Attensira's published limitations are explicit on three points, and each one closes off a tempting misreading:

• It is not a leaderboard position. Nothing in the number compares you to another brand.
• It is not a ranking of your brand. The unit is a URL on a domain, not an organisation.
• It is not comparable to a search engine result position. A blue-link position orders the whole index against a query. A citation position orders the two to ten sources one answer already decided to attach.

The last one is the difference that matters. In search, position is the outcome of selection: the ranking is the competition. In a generated answer, selection happened first — the model picked its sources — and the position merely describes how the surviving handful were arranged. Improving from fourth to second inside a list of four is not the same kind of event as moving from fourth to second on a results page, and treating them as the same is how a chart becomes a story that the data never told.

Why it is per domain

A model cites URLs. Aggregating those URLs into a brand requires somebody to decide which domains belong to which brand — including subdomains, documentation hosts, help centres, and pages a brand does not control that talk about it. That decision is a judgement.

Computing the average per source domain avoids making it. The number stays a fact about a domain, and any brand-level reading is left to the reader, who can see which domains went into it.

Reading it next to frequency

Citation position is close to meaningless on its own, because it is conditioned on being cited at all. Two examples make the point:

• A domain cited once, at the top of a two-source list, has an average position of 1.0.
• A domain cited two hundred times, usually third in a list of eight, has an average position near 3.

The first number looks better and describes a domain that is essentially absent from AI answers. Always read the position beside how often the domain was cited, and beside the sample the whole window rests on. A rate in Attensira is reported as {value, n} for exactly this reason: the sample travels with the figure.

The noise underneath it

Models are not deterministic. The same prompt to the same model can return a different answer minutes later, with a different set of sources in a different order. A position computed from a single draw per prompt is one sample presented as a measurement.

Two things in the product respond to that. Sampling depth puts a prompt to a model more than once per reading where the plan allows it. And every delta is tested with a two-proportion z-test at 95% significance, returning {real: false, value: null} when the movement does not clear the bar — we cannot prove a change, which is a different statement from nothing changed.

Freshness cuts the same way: each prompt × model × country combination runs at most once per workspace-local calendar day, so a partially elapsed day produces a partial and unevenly shaped sample. Compare whole days against whole days.

What to do with it

Citation position is most useful as a diagnostic on a domain you already know is cited. If your documentation host is consistently cited late in the list while a competitor's is cited early for the same topic, that is a signal about how the answer is being assembled — worth reading the stored answers for, since the verbatim text is kept for every successful run.

It is not a target to optimise in isolation, and it does not belong inside a composite score. Blending an index with a rate produces a number whose units are nothing; see AI visibility score. For what counts as a citation in the first place, see source citation; for the frequency half of the picture, see share of voice.

Finally, the honest ceiling on all of it: Attensira cannot tell you that a citation produced a visit or a signup. There is no identity join between an answer a model gave someone and a person who later arrived at your site. This is a measure of presence in answers, and connecting it to revenue is an inference you make with your own analytics.