How do you know whether assistants mentioned your business more often this month? Record the same prompts, repeat the runs, preserve the evidence, and make only the claim your sample supports.
This article gives you a completed prompt-panel register—not a substitute for analytics. A monthly AI search mention log records what selected assistants visibly returned for a fixed set of buyer questions at a particular time. It does not establish how often all buyers see your business, what the market recommends, or whether a website will receive citations.
The fictional owner, Maya, chooses four recurring buyer prompts for the same service area. She tests two assistant environments, labeled Platform A and Platform B, with three independent runs of every prompt in each monthly window. That is 4 prompts × 2 environments × 3 runs = 24 observations per month. The prompts, order, target business, competitor watchlist, geography and evaluation rules are fixed before collecting results.
Business: Cedarline Home Services (invented); market context: fictional North Valley service area; watchlist: Maple Ridge Repairs and Harbor Home Care (both invented). Prompt owner: Maya, marketing lead (invented). Collection windows: October 15 and November 15, 2026, 10:00–11:00 local time (fictional records). Environments: Platform A web assistant and Platform B search assistant; specific product/model version not captured in this demonstration. In real use, record product, mode, model/version if shown, account settings, location, personalization and browsing/search availability.
Fixed prompts: P1 “Which local companies repair dry-rotted window trim in North Valley?”; P2 “Recommend three North Valley contractors for small exterior wood repairs”; P3 “Where can I find a local exterior repair contractor with clear contact details?”; P4 “What sources explain how to choose a contractor for damaged exterior wood?”
The sample intentionally includes discovery, explicit recommendation, contact and source-seeking intents. Never replace a poorly performing prompt mid-series without starting a new panel version.
Not observed (N): the business name is absent from the saved answer. Mentioned (M): the name appears, but the answer does not explicitly recommend the business or cite its website. Recommended (R): the answer explicitly presents the business as an option or recommendation; a recommendation can exist without a website citation. Cited (C): the answer includes a visible link/source citation to the business website; a citation alone is not assumed to be an endorsement.
These are mutually exclusive primary labels for compact reporting, with a separate evidence field preserving overlaps. If an answer both recommends and cites the business, primary label R and evidence flags recommended=yes, cited=yes. For aggregate reporting, compute mention presence, explicit recommendation and site citation as separate binary measures from the evidence flags; never sum category totals as though they were the same outcome.
The register below compresses the three independent runs for each prompt/environment into an ordered three-character sequence. Each character is one recorded run, not an estimated frequency. All observations and evidence IDs are fictional.
| Prompt / environment | Oct 15 runs 1–3 | Nov 15 runs 1–3 | Evidence IDs | Change note |
|---|---|---|---|---|
| P1 · Platform A | N / M / N | M / M / N | O-A-P1-01…03; N-A-P1-01…03 | One additional name mention; inconsistent |
| P2 · Platform A | N / N / N | R / N / R | O-A-P2-01…03; N-A-P2-01…03 | Two explicit recommendations; not stable across runs |
| P3 · Platform A | M / M / N | M / M / M | O-A-P3-01…03; N-A-P3-01…03 | Name present in all November runs |
| P4 · Platform A | N / N / N | N / C / N | O-A-P4-01…03; N-A-P4-01…03 | One source citation; no recommendation |
| P1 · Platform B | N / N / M | M / N / N | O-B-P1-01…03; N-B-P1-01…03 | Same one-in-three presence |
| P2 · Platform B | N / N / N | N / R / N | O-B-P2-01…03; N-B-P2-01…03 | One recommendation; unstable |
| P3 · Platform B | M / N / N | M / M / N | O-B-P3-01…03; N-B-P3-01…03 | One additional mention |
| P4 · Platform B | N / N / N | N / N / N | O-B-P4-01…03; N-B-P4-01…03 | No observed brand citation |
How to read one row: For P2 on Platform A, October was N/N/N; November was R/N/R. The saved answer excerpts in the fictional evidence record show Cedarline explicitly offered as a contractor in runs 1 and 3, without a visible Cedarline URL. The middle run omitted Cedarline. This is not evidence of a consistent recommendation, and it says nothing about all users of that assistant.
| Evidence field | Filled fictional entry | Why it matters |
|---|---|---|
| Capture ID | N-A-P2-01 | Links the log row to one run |
| Run timestamp | 2026-11-15 10:17 local | Distinguishes repeat observations |
| Environment | Platform A · web mode · version not displayed | Preserves known configuration and unknowns |
| Prompt | P2 exact wording, panel v1 | Prevents silent prompt drift |
| Answer excerpt | “Consider Cedarline Home Services…” (invented) | Supports R classification |
| URL citation | None observed | Recommendation ≠ citation |
| Capture artifact | fictional://captures/N-A-P2-01.txt | Illustrative filename only; no actual artifact exists |
| Competitors | Maple Ridge Repairs mentioned; Harbor Home Care not observed | Tracks the fixed watchlist without implying market share |
| Reviewer note | R; recommended=yes; cited=no; second reviewer agrees | Shows how classification was checked |
For a real study, store screenshots or exported answer text, capture date/time, exact prompt, account and environment settings, observed URLs, and reviewer decisions in an access-controlled folder. The fictional evidence IDs above illustrate the register structure; they do not point to real screenshots. Keep the same two invented competitors on the watchlist in both months. Competitor appearance is context, not a denominator for a defensible market-share claim.
Use presence as a binary observation: M, R or C counts as present; N counts as absent. For this small illustrative panel, call a month-to-month change “meaningful within the tested panel” only if (a) the same prompts and environments are used, (b) all 24 observations are captured in both months, (c) presence changes by at least 8 of 24 runs (33.3 percentage points), and (d) at least three of the eight prompt/environment cells change in the same direction by two or more of their three repeated runs. If settings or panel wording change, mark not comparable rather than scoring a trend. These thresholds are conservative editorial decision rules for this fictional exercise, not validated statistical significance or an industry benchmark.
Repeat-run variability is visible when identical prompts return different labels within the same month. With only three runs per cell, do not treat a single new mention or citation as a durable gain. If the rule fails, the action is to continue sampling and investigate, not declare improved AI visibility.
| Measure | October | November | Interpretation |
|---|---|---|---|
| Any brand presence | 5 / 24 (20.8%) | 12 / 24 (50.0%) | +7 observations; threshold (c) not met |
| Explicit recommendation | 0 / 24 | 3 / 24 | Observed in two P2 cells, inconsistent |
| Website citation | 0 / 24 | 1 / 24 | One observed source link, not a ranking |
| Cells gaining ≥2 presences | — | 1 / 8 | Rule (d) fails; no meaningful change declared |
Decision: INCONCLUSIVE / CONTINUE. The observed brand presence rises from five to twelve of the 24 fictional runs, but only one prompt/environment cell gains two or more presences. The predeclared stability condition therefore fails. The owner does not report “AI visibility improved.” Instead, she retains panel v1, adds a second reviewer for ambiguous recommendation language, checks whether any environment setting drifted, and repeats the same panel next month.
A prompt-panel answer is a sampled output. It is not a Search Console impression, a Bing citation report, an analytics referral, or a qualified lead. To connect this observational log to actual website evidence, use the first-party AI search measurement and 30-day baseline guide. For practical tradeoffs between manual sampling and third-party monitoring, read the AI visibility tools guide. Any vendor-specific platform behavior, reporting features or model/version labels should be checked against the platform’s current documentation before adopting a real workflow.
Copy the columns: month, panel version, prompt ID and exact text, assistant environment, model/version if exposed, run timestamp, run number, primary label, recommended flag, cited flag, named competitors, evidence file, reviewer, settings changes and action decision. Keep the raw evidence separate from summary counts. Publish the measurement rule before running tests, then document exceptions rather than rewriting the rule after seeing results.
StartLab can help separate sampled AI-answer observations from first-party search performance, referral traffic, site clarity and real inquiry outcomes. Begin with the actual business and website evidence—not a promise of citations or assistant recommendations.
Explore SEO & AI Search GrowthRun the Free Business Check
For a scoped discussion, book a Strategic Session.