AI visibility evidence checklist: verify a claim before you publish it
A defensible AI visibility claim tells the reader what was observed, how it was counted and what remains unknown. Before a number enters a blog, slide or product page, check the evidence behind that exact sentence.
This is a review guide, not an industry benchmark or a report of GetCited customer results. It is for the editor or analyst deciding whether a claim is ready to share. For collection basics, use the existing citation-tracking guide.
What exactly are you claiming?
Write the claim in one sentence before calculating a rate. “The answer names the brand” is different from “the answer recommends the brand,” and both differ from “the answer attributes a link to the brand’s website.”
A search result adds another measure: whether an owned domain ranks for a question. A third-party page naming the brand is a separate observation. Keep the evidence field and the wording aligned; a mention does not become a citation because it appears in an AI answer.
Which questions and methods does the claim cover?
State the question panel, observation window and collection method. Separate questions that supply the brand name from neutral questions that do not. Preserve the actual wording so a later comparison can use the same panel.
Record the model or method version, language and requested location. Distinguish requested geography from verified geography. An API method should be described as that API method; its output does not establish what a person saw in a consumer application. Google’s site-owner guidance for AI features describes AI Overviews and AI Mode in Search; it is not a validation of a separate API proxy.
What belongs in the denominator?
Use the eligible records defined for that metric, and report missing or unusable observations separately. A failed request is not evidence that the brand was absent. An answer that is complete and eligible but has no citation is a valid zero for an overall citation rate.
Do not quietly remove source-free answers to make citation performance look stronger. A rate conditional on source-bearing answers can answer a narrower question, but it needs that narrower label and its own denominator. If a comparison depends on three repeats, disclose when only one or two are usable rather than treating the cells as equivalent.
What would correct arithmetic look like?
Synthetic example — entirely invented. Suppose ten answer captures were planned. Eight returned complete, eligible answers; two requests failed. Three of the eight answers named the fictional brand Acme, and one attributed a link to Acme’s verified domain.
- Mention rate: 3 ÷ 8 = 37.5% of eligible answers.
- Own-site citation rate: 1 ÷ 8 = 12.5% of eligible answers.
- Unusable capture count: 2 of 10 planned captures.
The failed requests are unknown, not brand absences. Valid answers with no citation stay in the citation denominator. These arithmetic examples say nothing about an actual brand, method or market.
Does the source prove the exact statement?
Keep the answer or result-list reference, capture date and evidence span for every coded observation. Verify the entity and domain rather than relying on a model’s suggested website.
For a citation claim, inspect the returned attribution fields and the supporting text. The Gemini grounding documentation describes source and support metadata. A retrieved source, an attribution in an API response and a link displayed in a consumer interface should not be silently treated as the same event.
Manual review should also distinguish a genuine recommendation from a warning, comparison, integration reference or passing mention. Record unclear cases rather than forcing them into a positive or negative result.
Are you allowed to use the evidence this way?
Check permission for the intended purpose and audience. Operational monitoring, internal analysis, public aggregation, quotations, charts and screenshots may require different permissions. Removing a name does not resolve a missing right or prevent identification through a distinctive description.
A small cohort, an unusual score combination or a before-and-after exclusion can reveal a contributor. Review related assets together so one chart does not reveal what another suppresses. GetCited-own data and publicly accessible source material also need the relevant provider, source-use and editorial checks.
What must be recorded before publication?
| Check | Required review record |
|---|---|
| Measurement | Exact claim, question panel, method/version, window and metric definition. |
| Counting | Numerator, eligible denominator, missingness, exclusions and treatment of repeated answers. |
| Source | Evidence references, entity/domain verification and human coding or adjudication. |
| Permission | Purpose, audience, raw and derived use, display/excerpt rights, expiry and withdrawal conditions. |
| Review | Method and factual review, privacy/contribution review, exact-artifact editorial approval and accountable owner. |
| Reuse | Approved charts, schema, screenshots, exports and captions; correction contact and a way to find dependent copies. |
What should you do when a check fails?
Hold the unsupported number and preserve the useful explanation. A qualitative guide or a clearly labelled synthetic example can teach a method while evidence is being reviewed. Do not replace missing evidence with another unapproved sample or trigger a new collection merely to fill a content quota.
After a factual or permission correction, identify the affected article, structured data, image, export and social copy. Record what changed and when. An observation about mentions or citations does not, by itself, establish traffic, sales or causal improvement.