open standard · CC BY 4.0
As of July 2026 the generative engine optimization field has no universally accepted standard for metrics, scoring methodology, or performance benchmarks. Commercial platforms measure "AI visibility" through proprietary methods that produce outputs which cannot be compared to one another. Buyers cannot verify whether the number they were sold went up because the work worked or because the methodology changed.
Separately, existing measurement is built for organizations. Every commercial tool measures brands. No standard measures people as entities, despite people being named in the definition of generative engine optimization by nearly every practitioner who publishes one.
CVI is a free, reproducible, versioned method for measuring how often AI answer engines name a given entity in response to buyer-intent prompts. It is deliberately simple. It can be run by one person with a browser in under an hour. It produces three numbers and a date.
CVI measures retrieval frequency for a named entity across a fixed prompt set and a fixed engine set at a point in time.
CVI does not measure: traffic, conversion, sentiment, accuracy of the description returned, or commercial outcome. Those are separate measurements and should not be conflated with this one.
CVI applies to any entity that can be named: a person, a creator, a freelancer, a consultancy, a small brand, a product.
Entity: the subject being measured, identified by a canonical name string.
Buyer-intent prompt: a prompt written as a purchaser of the entity's service would write it, expressing a need rather than naming the entity. Prompts containing the entity's name are out of scope and invalid under this standard.
Run: one prompt submitted once to one engine in a session with no prior context.
Named: the entity's canonical name or a recognized variant appears in the engine's response text.
Position: the ordinal placement of the entity within the sequence of entity names in a response, counting from 1.
| Parameter | Requirement |
|---|---|
| Prompt set | Minimum 20 prompts, fixed and published |
| Engine set | Minimum 3 engines, named and versioned where visible |
| Session state | Fresh session per run, no memory, no history, no personalization |
| Runs | prompts × engines, each run recorded |
| Date | ISO 8601, date of completion |
| Locale | Country and language of the testing session |
| Version | The CVI version used |
Any report omitting these is not a CVI report. Call it something else.
20 prompts minimum, distributed across four intent classes, 5 each:
Discovery: the buyer knows the need, not the supplier. best [x] for [y]
Comparison: the buyer is choosing between options. alternatives to [x]
Hiring: the buyer is ready to transact. who should i hire to [x]
Problem-first: the buyer states a problem, not a category. my [x] isn't working, what do i do
Prompts must be written in natural buyer register: lowercase, unpunctuated, and imprecise where real prompts are imprecise. Grammatically clean prompts test a population that does not exist.
The prompt set is frozen at publication. Modifying prompts invalidates comparison with prior runs. If the set must change, increment the version and publish both.
Minimum three. Recommended five: ChatGPT, Claude, Perplexity, Google AI Mode, Gemini.
Engines must be named in the report. Where a model version is visible, record it. Where it is not, record the date, which is the best available proxy.
MR = (runs in which the entity was named ÷ total runs) × 100
Reported to one decimal place. This is the primary metric.
SoV = (entity mentions ÷ (entity mentions + mentions of the top 4 co-occurring entities)) × 100
The comparison set is derived from the data, not chosen in advance. Take the four entities that appear most frequently across the run set, excluding the subject.
AP = mean ordinal position of the entity, across runs in which it was named
Runs in which the entity was not named are excluded. AP is undefined when MR = 0. Report as n/a, never as 0.
Compositing MR, SoV, and AP into a single score. Weighting is subjective, weighting choices are unverifiable, and a composite score is precisely the failure mode this standard exists to correct.
Every published CVI report must carry the following, or a materially equivalent statement:
AI-generated responses are non-deterministic. The same query submitted to the same engine can produce materially different answers depending on location, session context, time, model version, and ongoing provider updates. These results are a point-in-time snapshot under specific test conditions and should not be interpreted as fixed, guaranteed, or universally reproducible.
A report without this caveat is not conformant. The caveat is not a hedge. It is the reason the number can be trusted at all.
CVI v1.0 entity: [canonical name] category: [self-declared] date: [ISO 8601] locale: [country / language] engines: [list] prompts: [n] (set published at [url]) runs: [n] mention rate: __._% share of voice: __._% average position: _._ | n/a co-occurring entities (top 5, by frequency): cited domains (top 10, by frequency):
Quarterly. Same prompt set, same engines, same locale, new date.
A single CVI figure has limited meaning because there is no established benchmark for what constitutes a good score in any category. The delta between two conformant runs is the meaningful output.
CVI is licensed under Creative Commons Attribution 4.0 International.
You may use it commercially. You may build products on it. You may fork and extend it, provided extensions are versioned distinctly and not presented as CVI.
Attribution requirement: reports, products, dashboards, and articles using this standard must credit Creator Visibility Index v1.0, The Closer Method, closermethod.com/cvi. That is the entire cost of use.
Conformant runs may be submitted to the public CVI registry via elisabeth@closermethod.com. Submissions are published with the entity name, category, date, and three metrics. No raw data is required or stored.
The registry exists to build the thing the field currently lacks: category benchmarks derived from openly measured data rather than proprietary methodology.
v1.0, 20 July 2026. Initial public release. Three metrics, four intent classes, minimum 20 prompts across minimum 3 engines. Composite scoring explicitly excluded. CC BY 4.0.
Elisabeth Bierschenk Hitz · The Closer Method · closermethod.com
Corrections, disputes, and proposed amendments to the standard are welcome and will be logged in the changelog with attribution.
this method is the measurement layer of the ai visibility audit: the paid audit runs it for you, against your named competitors, and hands you the dated report.