AI Search Measurement and Audit Workbook

Use these worksheets to document a small project from its first question through a follow-up report. Copy the tables into your working document, or print the blank worksheets. Keep full answers and evidence files alongside the completed records.

Use your browser’s Print command to print or save a PDF. This file does not collect information or save entries online.

Completed audit example · Scoring definitions · Guide resources

Version: September 10, 2026. These are working templates. The completed example below is fictional. Replace it with your own measured results.

1. Define the project

Record the business, customer question, products to test, search conditions, dates, and decision this work should support.

Business and domainCustomer needProducts / conditionsDecision and owner

2. Record the prompts

Use one row per prompt version. Identify whether the question came from customer evidence or was constructed for the test. Keep branded verification questions separate from unprompted discovery.

Prompt ID / versionExact wordingSource of questionGroup / requirements

3. Record every attempt

Give retries new IDs and link them to the earlier attempt. Save the full answer and available tool records separately. Do not put failed attempts into the denominator of completed-answer rates.

Attempt ID / datePrompt version / product / modelSearch setting / actual useStatus / answer record

4. Score completed answers

Use one row per brand per completed answer. Mention = 1 when the brand appears; recommendation = 1 when the answer presents it as a suitable choice; site citation = 1 when the answer links to its domain. Otherwise use 0. Use “unresolved” when scoring is ambiguous and resolve it before final totals.

Attempt ID / brandMention 0 or 1Recommend 0 or 1Site citation 0 or 1Claim errors / evidence

5. Keep source and fact records

A returned search result, an opened page, and an answer citation are different records. State which was observed. Check each disputed fact against the responsible business source.

Attempt ID / URLObserved operationClaim checkedApproved fact / source / date

6. Assign corrections

A task is complete when its stated check passes. Retain old and new page content for material corrections.

Issue / evidenceAffected page or recordCorrection / ownerPublished / verification

7. Report the comparison

Compare the same prompt versions and test conditions. Retain failures and known changes. A before-and-after difference alone does not establish causation.

Period / conditionAttempted / completedMention / recommendation / citation countsRates / factual errorsKnown changes / next action

Completed example: Example Cooling

All business details and observations here are invented. The exact test question is: “Which HVAC companies serve central Phoenix and offer weekend appointments? Compare service area, weekend availability, and any published diagnostic fee.”

MeasureInitial batchFollow-up batch
Attempted2020
Completed1819
Mentions912
Recommendations69
Site citations37
Mention rate9 / 18 = 50.0%12 / 19 = 63.2%
Recommendation rate6 / 18 = 33.3%9 / 19 = 47.4%
Site citation rate3 / 18 = 16.7%7 / 19 = 36.8%
Answers containing old fee41

Observed problem: an old FAQ lists a $79 diagnostic fee, while the current approved fee is $99. Two initial answers cite the old FAQ.

Completed correction: the editor updates the FAQ and makes weekend appointment conditions visible in text. The reviewer checks the maintained pages and supported markup against the approved business facts.

Report: the factual correction is verified. The later answer sample contains fewer stale fees and a higher mention rate. The small before-and-after samples do not establish that the correction caused the visibility change.

Next action: inspect the remaining stale answer and its sources; repeat a comparable batch if the decision requires stronger evidence.

Calculation and quality checks