Citation probability
The informal name for the chance an AI answer cites a page. No tool computes it; it is only estimated from an observed citation rate over a fixed prompt set.
Karl-Gustav Kallasmaa, Founder & CEOLast updated Citation probability is the informal name for the chance that an AI system will cite a particular page when it answers a particular question. No product computes it directly. What is measurable is a citation rate — the share of completed runs, over a prompt set you chose, in which a source was named — and everything else is inference from that rate.
The distinction is the whole subject. A rate is an observation with a denominator you can state. A probability is a claim about runs nobody made, and it is conditional on the prompt wording, the model, the model version, the state of the retrieval index that day, and the user's own context. Treating the second as if it were the first is how a dashboard number becomes a story.
What actually determines whether a page is cited
In a retrieval-augmented answer the sequence is fixed, and each step can end the page's chances on its own:
- Fetch. A named agent requests the URL. A page it may not fetch, or cannot render, is not a candidate at any quality level. See robots.txt and crawling and indexing.
- Index and chunk. The page is split. A section covering four subjects produces a chunk that represents none of them well.
- Retrieve. The query is matched against the index, usually by vector similarity over embeddings, often blended with lexical matching. Paraphrase matches well; exact identifiers match badly.
- Select. Only the highest-scoring passages fit the budget of the context window.
- Generate and attribute. The model writes, and the surface attaches links — sometimes to everything retrieved, sometimes only to what it judges it used.
A citation appears at step five, but is lost at steps one through four far more often. This is why "improve authority" is a weak instruction: it acts on the corpus, slowly, while a robots directive or a client-rendered page that times out acts immediately and totally.
Why the number you can honestly report is a rate
Models are non-deterministic. The same prompt put to the same model twice can return two different answers citing two different sets of sources. A single draw therefore gives you a value of zero or one, and neither is a rate.
Sampling is the only response. Attensira's limitations page describes one draw per prompt per model per reading on the entry plan and three on higher plans, and states that every reported change is tested with a two-proportion z-test at 95% significance — a movement that fails that test is reported as unproven, not as zero. That is the shape of an honest citation measurement: a value, a sample size, and a refusal to call noise a trend.
Two failure modes follow directly. Reading a rate without its n turns one lucky answer into a percentage. Comparing rates across models without noting which models were even queried turns an untracked surface into an apparent failure.
What Attensira does and does not report
Attensira records the verbatim answer text for every successful run and derives mention rates from substring matching of a brand name or domain. Its documentation is explicit about the limits of that: ShareOfVoice is the brand's own mention rate under an older name, not a share of a fixed total; AvgPosition is the mean index of a URL within a model's own citation list, computed per source domain, and is not a rank of your brand against competitors. There is no visibility score and no sentiment metric, and there is no predicted citation probability.
Failure modes to avoid
- Reporting a probability you measured as a rate. Say "named in 4 of 12 runs on this prompt set" and the claim survives scrutiny. Say "a one-in-five chance of being cited" and it does not.
- Optimising the answer instead of the passage. The retrieved unit is a section. A qualifier two paragraphs above its claim does not travel with the claim.
- Assuming a citation means the source supported the claim. It records what was retrieved, not that it was right.
- Chasing a guarantee. Google's own AI-features documentation states there are no additional requirements or special optimizations for appearing in AI Overviews or AI Mode, and that indexing and serving are not guaranteed even when every requirement is met.
Frequently asked questions
Is it a field I can query?
No. Tools report an observed citation or mention rate over a prompt set; the probability is your inference from it.
Why is a single run not enough?
Because models are non-deterministic, so one draw yields zero or one, never a rate.
What moves it most?
Whatever is upstream of retrieval — fetchability first, then whether a self-contained passage answers the question.
Does a citation prove the page was read correctly?
No. It records what was retrieved. Retrieval and generation fail separately.
Terms related to Citation probability
The share of completed AI answer runs, over a defined prompt set, in which a brand or its domain appeared — an observed frequency with a denominator, not a score and not a prediction.
How an AI answer attributes what it says to the pages it read, and why a citation is a distinct outcome from a click or a mention.
The architecture that retrieves documents at query time and has a model write from them, and the reason your page can be quoted without ever being trained on.
Search where a model composes the answer and fetches pages through named crawlers, rather than returning a ranked list of links for you to read.
The practice of getting a source reproduced inside an AI-generated answer, introduced as a named paradigm in a 2023 research paper.