AI citation evidence: rates and uncertainty

Treat an AI citation as a placement you can observe: earn a reason to be cited, verify what the answer references, and monitor repeated answers. Keep the prompt, engine, date and evidence beside every result. A citation in one answer does not promise a place in the next answer.

This glossary explains AgentLinkOps citation reports. The citation monitoring overview covers setup, and the source research guide covers how to study cited pages. None of these records establishes a ranking gain or a visit to your site.

Mentioned and cited are separate facts

Mentioned means the answer text matches your brand or a supplied alias. Cited means a source URL matches your configured target. One answer can do both, either, or neither. A mention without a citation does not add to the citation count. A citation can count even when the answer never names your brand.

A target is the domain or page you chose to watch. Domain matching and exact-page matching answer different questions. Record the target rule with the result; do not compare a domain-wide count with an exact-page count as though they measured the same thing.

Verified, where a report uses that label, adds a successful reachability check of a cited URL. It does not certify that the page supports every statement in the answer. Read the answer and source together before making that judgment. A separate backlink check is needed to establish whether that source page links to your site.

Known answers and unknown attempts

Not cited means the answer was readable and its extracted citation list did not match the target. Unknown means the attempt could not establish that result. A blocked browser, failed source extraction or missing answer can produce an unknown.

Unknown attempts stay outside the citation-rate denominator and remain visible as a separate count. This prevents a failed measurement from becoming a miss. It does not remove bias: the answers you could read may differ from those you could not.

If every attempt is unknown, the rate and interval are unavailable. Report the failure reasons and try to resolve them before interpreting performance.

Panel, cell and epoch

A panel is your saved set of prompts, targets and engine settings. Prompt suggestions are editable research inputs. Their presence in a panel does not establish search demand.

A cell keeps one prompt, target and engine context together. Keep browser and API surfaces separate. Preserve model, language, country and available account context; an unknown account tier remains unknown.

An epoch is a batch with its planned sample count set before collection. A budget stop or an engine failure can leave the batch incomplete. Report the planned count, known answers and unknown attempts separately. Keep the same panel when comparing dates, or disclose what changed.

Citation rate and its denominator

Let k count known answers that cite the target and n count all known answers. The observed citation proportion is k/n. An answer with several matching URLs still contributes one success for that target.

AgentLinkOps has two rate conventions that must be named when sharing a report:

• The local epoch field uses a smoothed estimate, (k+1)/(n+2). It does not equal the observed proportion.
• The hosted aggregate uses k/n, derived from the retained counts during sync. The local evidence remains unchanged.

With no citations in one known answer, the observed count is 0/1. The local smoothed estimate is 0.33, and the Wilson 95% interval is approximately 0.00 to 0.79, n=1. These are calculations for an illustrative count, not a study result. One answer cannot support a trend claim.

Always include k, n, unknowns, the estimator name and the interval. The local CLI guide explains where your repository records live.

Declined, grown and insufficient data

Declined and grown describe a rate comparison using the method named in the report. Not distinguishable means that method cannot separate the samples. Insufficient data means the report refuses to interpret the available evidence. Neither refusal establishes stability.

The local epoch classification compares Wilson intervals descriptively. Its additional statistical comparisons have their own assumptions and multiple-comparison limits. A Wilson interval alone does not guarantee correct alerts across unlimited prompts or future runs. Keep the method and both samples with any change claim.

Write “citation rate declined” with the dates, counts and intervals. Avoid “citation removed”: a different answer is not evidence that a publisher removed an earned link. Published research on measurement uncertainty supports repeated sampling and uncertainty reporting; it does not validate every panel or guarantee a universal sample size.

Estimated cost is not a supplier charge

A budget check reserves an estimated amount before a call. Supplier-reported charges, when available, are a separate fact. An overrun can be recorded after a response and stop later calls; an estimate cannot guarantee the supplier's final bill.

Browser access can carry account, server or proxy costs even when an answer has no reported per-call charge. Missing cost evidence stays unknown. A historical estimate should never be relabelled as a verified charge.

Sources and review

Reviewed September 20, 2026. This is a product glossary, not an original-data study.

• AgentLinkOps citation statistics, AgentLinkOps, maintained source reviewed September 20, 2026: statistics source. Source definition: “Laplace-smoothed rate.” Publication date is not asserted for this changing source file.
• AgentLinkOps citation matching, AgentLinkOps, maintained source reviewed September 20, 2026: matching source. Source rule: “a successful re-fetch of the cited URL upgrades cited to verified.”
• Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement, Ronald Sielinski, arXiv, revised August 26, 2026: version 3. The abstract states: “citation visibility must be reported with uncertainty estimates.”

These definitions describe the reviewed code and the limits of its evidence. Consult the citation command reference for the installed version's inputs and outputs.