Measure AI-search work without confusing citations and visits

An AI-search project needs more than a screenshot of an answer mentioning your site. Define what you want to learn: whether the site appears for a chosen set of questions, whether recorded sessions have relevant source attribution, or whether those sessions contain a useful action. Each question needs its own evidence and denominator.

This guide includes an editable measurement plan and a small fictional dataset you can work through by hand. It is practical preparation for growth roles, content roles and GTM engineering roles that involve evaluating AI-search work. No model was run and no campaign result is claimed.

Start with four separate observations

Observation	Unit to record	What it can answer
Fixed prompt panel	One scheduled prompt trial, including its status	Did a completed answer contain the target link under these recorded conditions?
Search platform report	The platform’s defined impression or click metric	How often did the site appear or receive clicks within that report’s scope?
Site acquisition	A recorded session with session-scoped source dimensions	Which sources received attribution for the recorded sessions?
Useful outcome	An event count or sessions containing the event	Did recorded visitors take the defined action?

These are not interchangeable stages of one measured funnel. A prompt-panel answer is not a search impression from the audience. A citation does not establish a click. A click does not necessarily produce a recorded session, and an outbound apply click does not establish a completed application.

Use the worksheet before collecting results

Download the complete exercise as a ZIP, or inspect the five files individually. Open the CSVs in a spreadsheet app or text editor:

File	Use
Measurement plan	Fill the blank column with your decision, conditions, filters and outcome definition.
Prompt observations	Inspect 12 scheduled trials, including an error and a missed observation.
Session observations	Filter 13 synthetic sessions and count actions at the right level.
Rules and questions	Fix the citation rule, reporting window, exclusions and source bucket before calculating.
Worked readout	Check the arithmetic and compare your interpretation.

The CSVs contain values, not live formulas or external connections. You do not need a paid analytics tool or API access to complete the exercise. For real work, preserve the source export and record its extraction time and filters separately from your calculations.

Keep failed observations out of the citation denominator

Freeze the questions before the first round. Record the engine, displayed version or unknown, mode, date, exact prompt, trial number, language/location and account context. Preserve failed attempts and missed slots. If you change the panel or conditions, label the change instead of presenting the series as directly comparable.

In this exercise, a citation means a link to the target domain in a completed answer. A name-only mention does not count. A completed answer with no link is zero; an error or missed observation is unknown.

Both fictional waves complete five of six planned trials, or 83.3% coverage. The first has two cited answers out of five; the second has four out of five. That gives 40.0% and 80.0% among completed answers.

But the missing slots differ. Only four prompt/trial pairs complete in both waves. On that matched subset, the result is 2/4 versus 4/4, or 50.0% versus 100.0%. Report both the matched subset and its exclusions. Neither calculation is market share or evidence that the fictional content change caused an improvement. A small selected panel can be useful for diagnosis while remaining a poor basis for forecasting traffic.

Use Google’s current reports at their actual scope

As checked on October 25, 2026, Search Console has a dedicated generative AI performance report for impressions in AI Overviews and AI Mode. Its help documents impressions, not an AI-specific click metric; insufficient impressions can explain an unavailable report. This is an overlapping view of data included in overall Web performance. Do not add its impressions to the overall total.

Google’s AI features guide still describes inclusion in overall Web search traffic. Read that together with the dedicated report documentation, rather than concluding that no separate AI visibility view exists.

The exercise supplies 30 GenAI impressions, 120 overall Web impressions and eight overall Web clicks. Dividing eight by 30 would mix scopes and invent an AI click-through rate. The prompt panel cannot supply the missing click count either.

Check availability before treating an export zero as observed absence: the GenAI report’s documentation warns that unavailable ~ or - values become zeros in downloads. Preserve the original state when recording the export.

Count recorded sessions with consistent source rules

Use session-scoped dimensions for a session question. Google distinguishes Traffic acquisition from User acquisition: the latter describes first-user acquisition, which can answer a different question. Session attribution is also not proof of the raw referrer on every arrival or of the exact AI answer a visitor saw.

Write down the source rule. The exercise uses exact chatgpt.com or perplexity.ai values with medium referral; it is a teaching bucket, not an exhaustive list. After date, hostname and QA exclusions, three of ten included sessions match it. Keep Google organic and direct traffic separate. Google explains that (direct) / (none) lacks a clear referral source and can have several causes. It is not an AI attribution category.

For links you actually distribute, consistent campaign parameters help identify the placement in acquisition reporting. Use the real source, medium and campaign; distinguish variants with content labels. Tagging a social post about AI measures that distribution placement, not an assistant’s organic citation. Do not label a link as an AI referral merely because the content discusses AI.

Define useful actions at the session level

The fictional AI-source bucket has three apply-click events across three sessions. Two clicks occurred in one session, so only two sessions contain an apply click. The session action rate is 2/3 = 66.7%, not 100%.

Across all included sources, six apply-click events occur in five of ten sessions: a 50.0% session action rate. Neither number counts applications. The worked readout also shows why you cannot construct a guide-to-application funnel from unordered session totals: some apply-click sessions never have the guide-to-jobs event.

Choose an action that matters to the decision and confirm what it actually records. If your decision requires completed applications, obtain suitable completion evidence instead of renaming link clicks.

Close with the limits and the next decision

The exercise excludes marked QA, staging traffic and starts outside the declared New York reporting window. Its boundary row shows why a UTC date alone can misclassify a local reporting day. Search Console’s standard date-range views use Pacific Time, while Analytics uses the property’s timezone. Google also documents click/session differences and consent-related coverage limits; equal totals should not be assumed.

Record whether the real report includes observed or modeled data and retain its coverage limitations. Preserve consent choices. Google’s freshness documentation says processing can take 24–48 hours and recent data can change. Mark a fresh readout provisional and record later revisions rather than silently replacing it.

A useful recommendation names what you observed, what remains unknown and which next observation could change your decision. For this packet, repeat a predefined panel, investigate incomplete slots and validate outcome measurement with excluded QA traffic. More before/after observations alone do not establish causal lift.

In a portfolio, credit the supplied GTM Vibes dataset and explain your own calculations or additions. Use the AI-assisted take-home decision log when assistance is permitted, and the workflow failure-test exercise if you also need to evaluate the evidence and action boundaries of an AI workflow.