Measurement

Complete one measurement path first

Choose the decision that is active now: answer visibility, website outcomes, or incremental influence. Complete the smallest path that can answer it, and defer the other programs unless they are needed for that decision. The path is complete when its population, conditions, counting rules, and result can be inspected and used for the stated decision.

Choose the question before the metric

State the decision in one sentence. “Which high-risk product facts need correction?” needs preserved answer text and a fact review. “Do recognized answer-engine links produce useful visits?” needs a maintained source rule and website events. “Did answer visibility create additional demand?” needs an incremental study design; it cannot be answered by referral traffic alone.

Then name the unit: a prompt run, an eligible answer, a visible citation, a session, a recognized user, an order, or a claim under review. Name the resulting unit as well. Answers with a mention divided by eligible answers is a percentage of answers; orders divided by sessions is orders per session. Both can be useful, but they cannot be interpreted as the same kind of rate.

Use the evidence layer that matches the question

Evidence layer	It can answer	It cannot establish by itself
Answer observation	How a named product represented an entity or claim under stated conditions	How all users see the topic, or why that answer appeared
Visible citations	Which displayed sources or owned URLs appeared in sampled answers	Complete source use, endorsement, visits, or ranking
Search-platform reporting	Provider-defined impressions and clicks for the supported surface	The text, citations, or customer journey behind each impression
Website analytics	Recorded sessions, on-site events, and conversions under the source and identity rules	No-click exposure or a complete causal path
Incremental study	Whether an exposure or intervention changed an outcome beyond an appropriate comparison	Every individual mechanism that produced the effect

Google Search Console’s current Generative AI performance report covers impressions from AI Overviews and AI Mode in Search. It groups results by page, country, date, and device, and it applies its own aggregation rules. Use it for that provider-defined visibility question, not to reconstruct individual answers, citations, or referrals. (Google Search Console documentation)

Bing Webmaster Tools AI Performance can separately report cited pages, citation activity, and its provider-defined grounding queries. Those are aggregated retrieval-theme phrases associated with citations, not the individual prompts people entered or a complete record of every answer. Use them to identify themes worth inspecting alongside cited pages; do not treat them as keyword rankings or a causal explanation for a citation. (Bing AI Performance documentation)

Route the work

If you need to…	Start here	Result
Build a defined prompt sample and repeat it consistently	Prompt-set evaluation and sampling	A versioned frame, allocation, pilot, and run protocol
Calculate presence, citation, accuracy, or recommendation metrics	AI answer visibility metrics	Reproducible counts, rates, and a clearly bounded result
Preserve one answer as reviewable evidence	Answer-capture log	An immutable observation record with prompt, output, citations, and conditions
Measure recorded referral sessions and known customer journeys	Website analytics for AI answer-engine journeys	Separate session, acquisition, and assist views
Configure and inspect recognized referrals in GA4	Measure AI referrals in GA4	A documented channel rule and verified report view
Evaluate possible no-click demand or a site change	AI influence and incrementality	An honest comparison design and decision rule
Audit an entire program	AI SEO measurement checklist	A modular review record

Follow one learning path to a finished result

For answer visibility, build the sample with the sampling guide, preserve a run in the capture log, then classify the small batch in the metrics guide before calculating rates. You are ready to report when another reviewer can trace a numerator back to the included answers and explain the excluded ones.

For website outcomes, work through the analytics guide’s fictional session table and reproduce its separate session, user, order, and assist totals. Use the GA4 tutorial to configure the reporting view when GA4 is your tool. You are ready when the source rule passes its inclusion and exclusion checks and overlapping revenue reconciles to distinct orders.

For incremental influence, write the study brief and follow the counterfactual calculation in the influence guide. You are ready to commission or interpret a study when you can explain what was changed, why the comparison is credible, and which evidence would leave the decision inconclusive. A positive before-and-after number is not that result by itself.

Preserve the conditions behind every observation

For answer measurement, keep the exact prompt and complete response, product and displayed mode, market, language, time, account state when known, and any visible search or browsing setting. A refusal, unavailable feature, interrupted run, and an eligible answer without the entity are different outcomes. They must not silently become zeroes in the same denominator.

For website measurement, retain the maintained source rule, raw source or referrer where policy permits, landing page, event definition, identity rule, lookback window, consent limits, and reporting period. A session with no referrer is unknown. It may be a genuine answer-engine visit with missing attribution, an ordinary direct visit, or an implementation problem. Do not recategorize it as AI traffic.

For any comparison, version the prompt set, classification rubric, source grouping, and tracking changes. Annotate site releases, campaigns, outages, provider changes, and material fact changes. A before-and-after difference is an observation; it becomes evidence of an effect only when the design addresses credible alternatives.

Keep website outcomes as separate views

A recognized answer-engine referral can be relevant in more than one way. Report the views separately:

• a referral session began with the recognized source;
• an answer-engine-acquired user first reached the site through that source;
• an assisted outcome followed an earlier recognized referral within a stated window; and
• a same-session outcome occurred during the recognized referral session.

The same order can appear in several views. Those views answer different questions and must not be added into one revenue total. The website-analytics guide shows a completed fictional journey and the required counting rules.

Treat possible influence as a study, not a relabeling rule

A person can see a brand in an answer and later visit through search, direct navigation, or another channel. First-party analytics normally cannot join that unclicked exposure to the later visit. Trends in answer visibility, branded search, or direct demand can motivate investigation, but they do not identify individual influence.

Use an exposed-versus-comparison design, a randomized or phased rollout where feasible, customer research, or a documented external study. Define the eligible population, exposure rule, outcome, comparison, timeframe, and competing changes before reading the result. The incrementality guide explains when a useful inference is possible and when the result should remain descriptive.

A compact operating cycle

1. Choose one decision and its evidence layer.
2. Define the population, unit, inclusion rule, and owner.
3. Capture the underlying records before summarizing them.
4. Calculate only metrics whose numerator, denominator, and exclusions are explicit.
5. Review the result with its conditions and competing explanations.
6. Send a factual issue to case intake, a content gap to its accountable owner, and an unresolved attribution question to a better study design.

The checklist is useful once this model is clear. It keeps answer observation, website outcomes, tool review, and influence evaluation in separate modules so a broad review does not erase the evidence boundaries.