AI Visibility Scorecard: Template and Measurement Framework

Quick answer: what should an AI visibility scorecard measure?

Track four evidence layers:

Layer	Core question	Primary evidence
Presence	Did the brand appear for the prompts that matter?	Repeatable prompt-panel observations
Citation quality	Was the brand supported by a useful, attributable source?	Cited URL, source owner, citation context
Answer accuracy	Was the answer about the brand materially correct?	Required answer elements and manual review
Business evidence	Did visibility create measurable discovery or demand?	Google generative-AI impressions/clicks, referral sessions, assisted conversions

Then connect those results to a change log. Without that final layer, the team knows visibility moved but cannot reliably explain why.

---

What this scorecard does not claim

This framework does not claim to measure total market-wide AI visibility.

It measures performance inside a defined observation set:

• selected prompts
• selected platforms
• selected countries and languages
• a documented testing cadence
• a defined set of owned target pages
• available first-party search and analytics data

The score is useful for comparison across your own periods. It should not be presented as an absolute share of every possible AI answer unless the underlying dataset genuinely supports that claim.

---

Step 1: build a stable prompt panel

The prompt panel is the denominator for your measurement.

Start with 20 to 50 prompts grouped by business intent. A small stable panel is more useful than hundreds of prompts that change every week.

Recommended prompt groups:

Prompt group	Example intent	What it tests
Category discovery	Find a tool or provider for a problem	Unbranded market visibility
Comparison	Compare two approaches or products	Competitive positioning
Problem diagnosis	Explain why an outcome is failing	Expertise and educational coverage
Implementation	How to complete a specific task	Practical authority
Alternatives	Alternatives to a known product or method	Consideration-stage inclusion
Brand validation	Is this company suitable for a use case?	Accuracy and trust

For each prompt, store:

• a permanent prompt ID
• cluster and buyer stage
• exact wording
• persona
• country and language
• platforms to test
• expected answer elements
• preferred target URL
• owner and review cadence

Prompt-panel rules

1. Keep at least 70% of the panel unchanged between measurement periods.
2. Record materially different follow-up prompts separately.
3. Do not rewrite a prompt after seeing an unfavorable result.
4. Mark blocked, failed, or unusable answers instead of silently deleting them.
5. Test the same market and language when comparing periods.

---

Step 2: log answer-level observations

For every valid answer, record the observable result rather than only the final score.

Minimum fields:

• platform
• prompt ID
• test date
• brand mentioned: yes/no
• brand cited: yes/no
• cited URL
• owned or third-party source
• competitor mentioned or cited
• answer materially accurate: yes/no/partial
• required answer elements present
• screenshot or evidence reference when your workflow permits it
• reviewer notes

Mention rate

Mention rate = answers mentioning the brand / valid answers tested

Citation rate

Citation rate = answers citing an attributable brand-related source / valid answers tested

Owned-source citation rate

Owned-source citation rate = answers citing an owned URL / valid answers tested

A mention and a citation are not equivalent. A brand may be named without evidence, and a third-party page may be cited instead of the brand's own site. Preserve that distinction.

---

Step 3: score answer accuracy

Visibility is not automatically positive when the answer is wrong.

Before testing a brand-validation or product prompt, define the answer elements that must be correct. Examples:

• current product category
• supported workflow
• availability status
• pricing status
• target customer
• integration status
• important limitation

Use three states:

State	Definition
Accurate	All material required elements are correct
Partial	Core answer is useful but one or more material details are missing or outdated
Inaccurate	A material claim is wrong, misleading, or attributed to the wrong entity

Accuracy rate

Accuracy rate = accurate answers / answers reviewed for accuracy

Keep partial and unverifiable answers visible in the raw data. Do not force uncertain observations into a positive score.

---

Step 4: use Google Search Console's generative-AI report when available

Google's dedicated Generative AI performance report includes impressions for AI Overviews and AI Mode. Google says the report is rolling out to a subset of properties and may not appear when access has not reached the property or the site lacks enough qualifying impressions.

Useful dimensions include:

• page
• country
• date
• device

Google also documents how clicks, impressions, and position are counted for AI Overviews and AI Mode. An AI Overview link generally needs to be scrolled or expanded into view to count as an impression, while an external-page click counts as a click.

Official references:

• Generative AI performance report
• How AI Overview and AI Mode impressions and clicks are counted
• Search Console data limitations and discrepancies

Record these fields by page and period

• generative-AI impressions
• generative-AI clicks
• click-through rate when meaningful
• country and device shifts
• newly appearing pages
• pages losing visibility

Do not treat a zero in an exported file as proof that no visibility occurred. Google notes that unavailable values displayed as ~ or - can export as zero, and Search Console reports have sampling, privacy, row, and processing limitations.

---

Step 5: connect AI referrals and conversions

Create a separate analytics segment for identifiable AI referrals. Preserve the raw source and medium before grouping domains into a reporting category.

Track:

• referral sessions
• engaged sessions
• landing pages
• signup or lead events
• assisted conversions
• revenue only when attribution is defensible

AI answers can influence discovery without producing a direct referral. That is why referral traffic belongs beside prompt observations and Search Console data, not in place of them.

---

The 20-point operational score

The downloadable scorecard uses four transparent 0-to-5 subscores. This is a triage index, not a universal market-share metric.

Subscore	0	1	3	5
Presence	No valid tests	Mentioned rarely	Mentioned in a meaningful share of priority prompts	Consistent presence across priority clusters and markets
Citation quality	No attributable citations	Weak or irrelevant citations	Useful citations with mixed owned/third-party sources	Consistent, relevant citations to strong owned and third-party sources
Answer accuracy	Materially wrong	Frequently incomplete or outdated	Mostly correct with identifiable gaps	Materially correct across priority prompts
Business evidence	No measurable evidence	Isolated unqualified visits	Repeatable impressions, referrals, or assisted outcomes	Sustained evidence connected to priority pages and conversion events

Operational AI visibility score = presence + citation quality + answer accuracy + business evidence

Maximum: 20.

Always publish the four components beside the total. A score of 14 caused by strong mentions and weak accuracy is not equivalent to a score of 14 caused by modest presence and strong business outcomes.

---

How to set the 0-to-5 thresholds

Do not copy another company's percentage thresholds.

Set internal thresholds after collecting at least two valid baseline periods. Recommended process:

1. Run the same panel weekly for four weeks.
2. Calculate the median and range for each metric.
3. Define 1 as weak but observable performance.
4. Define 3 as repeatable performance in priority clusters.
5. Define 5 as the target state your current market and data can realistically support.
6. Document every threshold in the scorecard notes.

This keeps the index stable while allowing the raw metrics to remain comparable.

---

Step 6: connect every optimization to a measurement window

A content update without a change log creates attribution debt.

For every shipped action, record:

• target URL
• change type
• hypothesis
• supporting evidence
• exact ship date
• baseline period
• measurement period
• primary metric
• guardrail metric
• owner
• rollback reference

Recommended windows

Change type	Initial observation	More stable review
Factual correction	Next prompt-panel run	Two to four consistent runs
Definition or answer block	One to two weeks	Four to eight weeks
Internal-link improvement	After recrawl and reprocessing	Four to eight weeks
New supporting source or original evidence	Two to four weeks	Eight to twelve weeks
Title or snippet change	After reprocessing	Two to six weeks

These are observation windows, not guaranteed result timelines. Record external events and major platform changes that may affect interpretation.

---

Weekly operating cadence

Monday: collect

• export available Google generative-AI data
• run the stable prompt panel
• capture citations and answer accuracy
• export AI referral and conversion segments

Tuesday: diagnose

• identify prompt clusters losing presence
• identify inaccurate or outdated answers
• identify cited third-party sources that could be strengthened
• identify owned pages receiving impressions but weak clicks or conversions

Wednesday: prioritize

Use a simple action score:

Priority = expected impact × confidence / effort

Keep the inputs visible. Avoid presenting the result as scientific precision.

Thursday: ship

• update one answer block
• add or improve one source
• fix one factual inconsistency
• strengthen one internal-link path
• record the exact change

Friday: annotate

• add the ship date to the change log
• set the baseline and review windows
• note any platform or measurement anomalies

---

Interpretation matrix

Observation	Likely diagnosis	First action
Mentions rising, citations flat	Brand is known but sources are weak or not selected	Improve attributable evidence and source clarity
Citations rising, accuracy weak	Sources are found but facts are inconsistent or outdated	Reconcile product, pricing, entity, and availability statements
Google generative impressions rising, clicks flat	Visibility increased without stronger click motivation	Review query/page intent, titles, and destination usefulness
Referrals rising, conversions flat	Landing page does not match the AI answer's promise	Align page message, proof, and CTA with referring intent
One platform improves, others do not	Platform-specific retrieval or source behavior differs	Diagnose by platform instead of averaging everything together
Score improves after many simultaneous changes	Attribution is unclear	Reduce batch size and use the change log rigorously

---

Common measurement mistakes

Changing the prompt panel every week

This destroys the denominator and makes period comparisons unreliable.

Counting every mention as a citation

A mention has lower evidentiary value than a visible attributable source.

Ignoring wrong answers

An inaccurate recommendation can be more damaging than no recommendation.

Averaging countries and languages together

Visibility can differ materially by market. Keep dimensions separate before creating a global summary.

Claiming causality from timing alone

A change followed by an improvement is evidence worth investigating, not automatic proof that the change caused the result.

Publishing a score without the method

Always disclose the prompt count, platform set, market, period, formulas, and unavailable data.

---

Minimum viable scorecard for a small team

A founder or lean team can begin with:

• 20 stable prompts
• three platforms
• one country and language
• weekly observations
• one scorecard row per platform
• one shared change log
• Google generative-AI data when available
• an analytics segment for identifiable AI referrals

That is enough to create an operating baseline without pretending to measure the entire AI-search market.

---

FAQ

What is an AI visibility scorecard?

An AI visibility scorecard is a repeatable measurement framework for tracking whether a brand appears in priority AI-answer prompts, whether it is cited by attributable sources, whether the answer is accurate, and whether visibility creates measurable search or business evidence.

How many prompts should I track?

Start with 20 to 50 stable prompts grouped by buyer intent. Consistency matters more than volume. Keep most prompts unchanged between periods and record new prompts separately.

Is AI referral traffic enough to measure visibility?

No. Referral traffic measures identifiable visits, while many AI answers create mentions, citations, or discovery without a click. Combine referrals with controlled prompt observations and available first-party search data.

Can Search Console measure AI Overview and AI Mode visibility?

Google has introduced a dedicated Generative AI performance report for a subset of Search Console properties. It includes impressions from AI Overviews and AI Mode and supports page, country, date, and device analysis. Access and qualifying data may vary by property.

Should I publish one AI visibility percentage?

Only when the denominator and scoring method are disclosed. For operational use, publish raw mention, citation, accuracy, impression, referral, and conversion metrics beside any composite score.

---

Final takeaway

The purpose of an AI visibility scorecard is not to create a more impressive dashboard. It is to make the observation loop reproducible:

1. define the prompts
2. collect the evidence
3. separate mentions, citations, accuracy, and outcomes
4. ship one attributable change
5. measure the next window

That is how AI-search visibility becomes an operating system rather than a collection of screenshots.