What are AI visibility metrics measuring?
AI visibility metrics describe observations of your brand or website in selected AI-generated answers. They can include the proportion of usable answers mentioning the brand, the proportion linking to the official domain and the order of a link in a source list. These are observations of a chosen question set, not access to a model's private confidence scores. Their value depends on the sample definition and evidence behind each record. A composite visibility score without a published formula is not interchangeable with a raw mention or citation count.
Begin by separating four layers: website readiness, answer presence, attributable sessions and qualified outcomes. GAS, or GEO Authority Score, belongs to the readiness layer and is a proprietary 23SEOGEO assessment. It does not forecast a citation probability or describe how accurately an engine answers every question. Mention and citation records belong to observed presence. Referrals belong to website analytics, while qualified leads require business-side confirmation. Keeping these layers separate prevents an improved technical score from being reported as revenue or a crawler hit from being counted as a customer.
Google's AI features guidance, reviewed on October 14, 2026, states that existing SEO practices remain relevant for AI Overviews and AI Mode. It also describes reporting within Search Console's web search performance data. Do not assume that ordinary web impressions reveal all conversations in ChatGPT or provide individual source-link positions inside every generated answer. A platform-specific measurement needs its own evidence and clearly described coverage. Treat unavailable breakdowns as unavailable rather than manufacturing a dashboard field from unrelated data.
Deep analysis: a metric dictionary your team can audit
Define the denominator before drawing the chart. A usable answer is a successfully returned response that meets the test's basic requirements; an unavailable engine or timed-out call belongs in a separate failure count. Removing failures from a rate does not justify hiding the failure total. State whether repeated tests count as individual responses or are collapsed by unique question. Preserve the same rule in exports, charts and narrative conclusions so the reporting does not change depending on which screen a client opens.
| Metric | Numerator or record | Denominator or scope | What it cannot prove | | --- | --- | --- | --- | | Brand mention rate | Usable answers containing a verified brand alias | Usable sampled answers | Official-domain citation | | Official-domain citation rate | Usable answers with a verified link to the target site | Usable sampled answers | A visit or purchase | | Citation position | Order of the official link in the returned source list | Responses with that link | A universal search rank | | Mention frequency | Occurrences of the brand in an answer | State per-answer or total aggregation | Unique customers reached | | Sentiment label | Context classified by a disclosed method | Reviewed brand-related statements | Purchase intent or endorsement | | AI referral sessions | Sessions with documented AI attribution | Selected site and reporting period | Total AI answer exposure | | Qualified leads | Confirmed inquiries meeting business criteria | Attributable leads or sessions, explicitly stated | Causal uplift from one edit | | GAS readiness | Evidence-backed site checks | Scorer version and sampled pages | Citation probability |
Mention rate is calculated by dividing usable answers that mention the brand by all usable sampled answers. Official-domain citation rate uses the same denominator but counts answers with a validated link to the chosen site. Count an answer once for each rate even if it repeats the brand name several times. If you also want mention frequency, show that as a separate count. A citation to a review about your brand is useful evidence, but it is not an official-domain citation. Do not merge the two just because both contain the company name.
Sentiment needs especially careful interpretation. A keyword-based label can misread negation, quotations or a comparison involving another company. An AI classifier can also make mistakes and should expose its definition and uncertainty. Keep positive, neutral, negative and unclassified records inspectable rather than claiming that a sentiment number measures recommendation strength. Review unexpected changes against the original answer. Unless a vendor publishes how a score is produced, do not treat it as a model-provided confidence value or a universal industry standard.
This conceptual illustration emphasizes matched samples and source review. It is not a real report and contains no measured uplift. In a real report, each folder should represent a documented test window with the same question-set version. Every cited link needs a check for its destination and relation to the target brand. Keep changes in model, language or browsing capability visible because they can affect comparability. Good charts summarize evidence; they do not substitute for it.
How do you build a fair question library?
Separate brand questions from category, comparison and problem questions. A prompt that names the company tests a different situation from a neutral request for suppliers. Use customer interviews, support questions, sales objections and relevant search queries to form the library. Search demand can help prioritize a question, but a keyword's estimated volume is not the volume of AI conversations. Record the intended market and language without implying a geographic test that the integration did not actually perform. Avoid leading prompts that instruct the model to praise or select your brand.
Freeze a version of the question set for a comparison period. If you add questions, show the new set separately or compare only the overlapping questions. Distinguish a new brand alias from a previously valid alias to avoid artificially increasing historical mentions. Repeated testing can reveal variability, but repeated answers to one question should not masquerade as independent coverage of many buyer needs. Report sample size and usable count alongside rates. Small samples remain exploratory; larger samples still reflect only the chosen testing design, not the entire market.
Why are crawler logs not citation evidence?
OpenAI's crawler documentation, reviewed on October 14, 2026, distinguishes OAI-SearchBot, GPTBot and ChatGPT-User. OAI-SearchBot relates to search discovery, GPTBot is used for training-related crawling, and ChatGPT-User is associated with user-triggered actions. These roles should not be collapsed into one AI-visitor metric. A crawler request shows an access event, not that an answer mentioned your business. A user-agent string alone can also be spoofed, so verify request identity using the provider's documented method where available.
Server logs help investigate availability, blocked requests and which URLs were fetched. They do not reveal the text of all later generated answers or prove which user clicked a citation. Conversely, website analytics can observe a real browser session but will miss referrals when source information is stripped. Use logs for access evidence, answer records for citation evidence and session attribution for visits. Missing attribution should remain an unknown, not be reassigned to AI because a bot visited the site around the same time.
How should global brands compare periods and engines?
Match the question set, language, test configuration and reporting window before comparing percentages. Keep engine or model identifiers in the record and note configuration changes that affect browsing or grounding. An API result is not necessarily the same as the consumer application's answer. Do not label a provider-backed model as a verified test of a different product unless the integration supports that claim. If an engine is unconnected, show that status and exclude it from success-rate calculations without hiding the coverage gap.
Compare counts as well as rates. A higher rate with fewer usable responses can be weaker evidence than a stable rate on a larger comparable set. Show the sampled period's failures and question-type mix. If the site has no monitoring history, say that the baseline is not yet established rather than importing another project's mentions. Scope referral analytics to one selected site and one explicit period. Keep relative changes descriptive unless a stronger evaluation supports causality, particularly when site updates and engine changes occurred together.
A practical measurement workflow
Start with the GEO results measurement topic to see the four-layer reporting model. Use AI citation tracking to preserve sampled answers and AI traffic tracking to inspect actual attributed visits. The dual-engine scoreboard can help compare available Google-position and AI-citation records for the same URL. The GEO score tool investigates readiness, not those market outcomes. None of these tools should create historical results for a site that has not been measured.
When a gap appears, open the original answer before choosing a fix. Check whether the official page answers the question, whether the cited alternative has stronger relevant evidence and whether the URL is accessible. Update verified facts and page structure only when the diagnosis supports it. Publish the change and repeat the same questions on a defined schedule. Record what changed without promising that a model will adopt the updated page. Successful implementation and verified outcome are different stages of the work.
Frequently asked questions
Is a high GAS score the same as a high AI visibility rate?
No. GAS evaluates website readiness under a stated rule set. Visibility rates come from sampled answers. A ready page may not appear in the tested questions, and a mentioned brand may still have weak site readiness.
Does a citation mean a visitor arrived?
No. A source link in an answer establishes a citation for that answer. A visit requires a browser session and usable attribution evidence. A qualified lead additionally requires business-side confirmation.
Can one engine result represent the whole market?
No. Keep results separated by engine, question type, language and test configuration. Aggregated charts should disclose their composition and limitations rather than imply universal coverage.
Editorial scope and sources
Prepared by the 23SEOGEO team; source guidance reviewed on October 14, 2026. Metric formulas in this article are operational reporting definitions, not Google or OpenAI scores. Google supports the Search-feature eligibility and reporting distinctions; OpenAI supports crawler-role distinctions. No customer uplift, unsupported universal threshold or private model-confidence figure is asserted.
Research context and structured evidence
The GEO research paper supplies a research framework for generative visibility evaluation, not a forecast for these metrics or proof of customer uplift. Use it to understand why sample design matters, while keeping live observations separate from laboratory findings. W3C JSON-LD defines how linked data can be expressed consistently. It supports the representation of article and source entities, not the truth of a claim or automatic inclusion in an answer. Both references were reviewed on October 14, 2026.
