How Agencies Should Measure Mentions, Recommendations and AI Citations

Build a repeatable prompt panel, preserve the evidence and report what each AI engine actually showed without confusing mentions with endorsements.

Share
Measuring tapes and rulers with different scales.
Illustrative image. Photo by William Warby / Unsplash

A client sends a screenshot: their brand appeared in an AI answer. The next question should be, “What exactly appeared?”

A name in a list, an explicit recommendation and a link to the client's website are different observations. Reporting them as one win makes the next client conversation harder. You cannot explain what improved if the original evidence was never classified.

Use a measurement procedure that another person on the account can repeat. The answers may change. Your definitions and collection method should not change quietly with them.

Define three separate outcomes

Before collecting responses, agree on a short coding guide:

  • Mention: the answer names the correct brand. Record misspellings and ambiguous names for manual review.
  • Recommendation: the answer positively presents the brand as suitable for the stated need. A neutral directory, warning or rejection does not qualify.
  • Linked citation: the answer includes a source link to the client's domain. Save the destination URL, not just the displayed domain label.

These outcomes can overlap. Record each independently. Also distinguish a link to the client's own page from a link to a third-party article that mentions the client. Both may be useful, but they represent different paths to visibility.

Add a separate accuracy check. A recommendation based on a service the client does not offer is a correction opportunity, even if the brand appears prominently.

Build a fixed panel around buying questions

Begin with questions the client can substantiate through sales enquiries, support conversations and approved search data. Include discovery, comparison, use-case and constraint-based questions. Keep branded questions separate from unbranded discovery questions.

For example, a hypothetical panel could contain 24 questions: six in each of four intent groups. That is an operating choice, not a statistically representative sample of all demand. Record why each question belongs and which page should help answer it.

Freeze the core wording for the reporting period. Keep new question ideas in an exploratory panel so that adding easier prompts cannot silently improve the headline result.

Choose a repeat rule in advance, such as three separate observations per question in each weekly collection window. Do not keep rerunning a question until a favourable answer appears. Report the sample size and missing runs alongside results.

Record the conditions, not just the screenshot

For every observation, retain:

  • Exact prompt and collection timestamp, including timezone.
  • Product surface, visible model or mode, language and intended location.
  • Whether search was used, where that is observable.
  • Account and session conditions, including whether prior conversation context was present.
  • Full answer, source links and the three outcome labels.
  • Errors, unavailable responses and ambiguous cases.

Use a fresh conversation for each observation. Keep personalisation settings consistent where the product allows it. OpenAI documents that memory and location can influence search queries, which is a reason to record those conditions rather than assume every user sees the same answer. OpenAI search guidance

Keep Google AI Overviews, Google AI Mode and Gemini observations in separate rows. Likewise, label API-based collection distinctly from consumer-interface checks. A shared provider name is insufficient evidence that two collection methods measure the same experience.

Keep the denominator visible

Report mentions, recommendations and citations as separate counts over valid observations, broken down by engine and intent. Show collection failures separately rather than treating them as brand absences.

For Google searches, distinguish “no AI Overview shown” from “AI Overview shown without the brand.” Report the trigger rate alongside brand presence within observed Overviews. Otherwise, a change in how often the feature appeared can distort the apparent brand trend.

Do not describe your panel result as market share. You selected the questions, and repeated responses are not necessarily independent samples of customer behaviour.

Reconcile with native reporting and business evidence

Google's Generative AI performance report covers impressions from AI Overviews and AI Mode. Google says its worldwide rollout completed on August 31, 2026, but the report may not appear for sites without enough impressions. Those impressions are a different measure from your hand-labelled prompt observations. Google report documentation

Track referral sessions and qualified enquiries separately. Neither a citation nor an impression establishes that a prospect visited, enquired or purchased. Preserve the original observation so someone can audit a surprising change.

Kloovy describes engine rechecks after approved fixes. See the workflow. Bring this measurement protocol to that process: define the question, retain the evidence and compare like with like.

Get started with Kloovy and evaluate your next approved change against a panel your agency can defend.

Sources

  1. OpenAI: Searching the web with ChatGPT
  2. Google: Generative AI performance report (Search)
  3. Kloovy product workflow