AI Mentions, Citations & Recommendations

TANTU AI Research & Insights · 16 September 2026 · Practical guide

A mention identifies the brand

A sampled answer can name a company in a list, comparison or warning. Mention rate says the agreed name appeared. It does not determine whether the context was favorable or commercially useful.

A citation points to a source

An answer can link to a brand-owned page without recommending the product. Conversely, it can recommend a brand while citing another source. Define which domains count as owned sources and retain the source URLs.

A recommendation requires context

Review the full answer against a documented rule. An explicit favorable recommendation for the relevant use case is different from an incidental reference. Ambiguous cases need review, not a convenient positive label.

Use the correct denominator

Report the count of qualifying valid answers alongside the total valid sample. Failed collections are recorded separately. If a platform does not trigger an AI feature, record that condition rather than silently converting it into brand absence.

Connect to decisions carefully

Review changes by market, language and interface. Ask whether the cited page supports the buyer’s next question and whether the customer can act. Visibility measures observed answers; sales performance needs additional evidence.

Decision summary

A citation identifies a source used in an answer; a recommendation presents an option as suitable for the buyer’s stated need. They can occur together or separately. Classify the full answer using documented rules, record the cited domain independently, and preserve uncertain cases for review. A link to your website does not mean the product was endorsed, and a recommendation can rely on another publisher’s page. Report these signals separately because they reveal different information problems and cannot be added into a single count of positive outcomes.

Test the classification with difficult examples

Use short, fictional examples to align reviewers before collecting a trend. “Company A publishes a guide to deployment” names the brand and may cite its page, but does not recommend its product. “Company A may suit a single-site operator, but it does not meet your regional requirement” is not a positive recommendation for that stated requirement. The constraint in the buyer’s question determines the classification.

By contrast, “Company A is a suitable option if the documented integration meets your requirements” can qualify under a rule that permits conditional recommendations. Preserve the condition in the record. Do not delete qualifying language to produce a stronger result. When reviewers disagree, retain the answer, record the competing interpretation and resolve the rule before applying it across the sample.

Match the signal to a useful investigation

A citation without a recommendation may mean the page is useful as a factual source but does not address supplier selection. A recommendation without an owned citation may lead you to inspect the independent evidence being used. An inaccurate recommendation needs a factual investigation even when the brand appears favorably. These are investigation paths, not automatic conclusions about why a platform produced an answer.

Hypothetical example: of 12 completed eligible answers, four cite the owned domain and three recommend the brand; two do both. Report 4/12 citations and 3/12 recommendations. Do not report seven distinct positive answers: the categories overlap. Keep the answer-level classification so a buyer can inspect which two records overlap and whether the underlying statements are accurate and relevant.

Turn the insight into a focused plan.

Apply these principles to your market, evidence and customer journey.

Discuss your priorities

A citation is not the same as a recommendation