Start with the event, not the dashboard
“Are we visible in AI search?” is not one measurement question. It bundles at least three different events: an assistant included the brand in an answer; AI infrastructure reached a page; or a person clicked from an assistant and arrived on the site. One can occur without either of the others.
Brandvane therefore does not treat an answer, a crawler hit, and a GA4 session as interchangeable units of visibility. Each observation keeps its source, time window, and denominator. Reconciliation happens only after the lanes have been reported on their own terms.
The method should let a skeptical client trace a sentence in the report back to the observations that support it.
Independent measurements
The three signals Brandvane measures
Sampled answers: what the assistants said
Current SEO and Pro+ fresh checks each sample one OpenAI API answer to one question. Recorded-mention searches separately retrieve DataForSEO evidence; they are not fresh queries to every consumer assistant. Preserve the question, source, disclosed model, run time, raw answer and available citations with each observation. A stable panel is a research design you can organize from separately approved checks, not an included automatic multi-engine monitoring service.
Google AI Overviews and AI Mode are measured as a separate Search surface. When available, first-party Search Console impressions and controlled snapshots retain their own dates, markets, devices, and denominators rather than being blended into assistant-answer incidence.
The primary unit is sampled appearance incidence. If the brand appeared in 7 of 12 completed answers, the report says “appeared in 7 of 12 sampled answers.” It may also show 58.3%, but the fraction remains beside it. Engine and question-cluster breakdowns stay available so an aggregate cannot hide where the pattern occurred.
The practical guides to tracking AI citations and testing citation improvements carry that rule into source registers and before-and-after work.
It supports: a claim about this disclosed panel and window. It does not support: a universal rank, every answer users received, or the next answer an assistant will produce.
Verified crawler access: what infrastructure did
Server or edge logs can show that identified AI infrastructure requested a URL, when the request occurred, and whether the page was reached or blocked. Where verification is possible, Brandvane uses more than a self-declared user-agent string and labels the crawler’s purpose rather than pooling unlike bots.
A separately prepared audit can include a crawler lane when suitable logs are supplied. Current SEO and Pro+ plans do not automatically ingest crawler logs. Without supplied evidence, label the lane not connected; do not infer crawling from a citation or fill the space with synthetic data.
The practical AI crawler analytics guide explains how to verify a claimed bot, interpret edge and server outcomes, and test what remains visible when essential content depends on JavaScript.
It supports: an infrastructure-access claim for the observed requests. It does not support: a claim that a page was used in an answer, cited, recommended, or seen by a person.
AI-referred human visits: who arrived
GA4 can report sessions whose source contains a recognizable AI-assistant referrer. When analyzing a supplied export, preserve its date window, total sessions, AI-referred sessions, and available source or landing-page dimensions. Our free traffic tool processes a supplied GA4 CSV locally; it is not an automatic GA4 connection included in SEO or Pro+. The denominator is all recorded sessions in the same window, not an unrelated estimate.
Referral attribution is useful and incomplete. A person may learn about a brand in an assistant, then arrive through direct or branded search. Privacy controls can remove attribution. An assistant may mention a brand without producing a click. The number is therefore described as AI-referred human visits, not all AI influence.
It supports: a claim that GA4 recorded referred sessions from known sources. It does not support: crawler activity, unclicked mentions, or the assistant answer that caused a visit.
Noise control
What replication does—and does not do
Generated answers vary. A single run can be true as an observation and weak as a trend. For an illustrative research design, select priority questions and repeat each three times within a stated window. A multi-engine study requires separately arranged access to each engine. Current SEO and Pro+ plans do not run this panel automatically: each fresh check uses one separate unit and samples OpenAI only. State the actual observation window rather than implying a default weekly cadence or four-week reporting period.
Three repetitions make obvious variability visible and reduce the influence of one unusual answer. They do not eliminate sampling error, reconstruct private user conversations, or establish the probability of appearance for every possible prompt. The number is a practical operating choice, not a claim of statistical completeness.
The guide to why generated answers vary separates sampling, model changes, retrieval, prompt wording, conversation state, and product surface without pretending an outside observer can assign every change to one cause.
Failed samples stay visible in method notes. If 31 of 36 planned samples returned usable answers, an appearance fraction uses the 31 completed answers while the report also discloses 31 completed of 36 planned. A model change, prompt edit, or other break in the instrument is labeled rather than silently joined to the old series.
Composite metrics
Why Brandvane does not publish a 0–100 AI visibility score
A single AI visibility score is attractive because it makes a complicated channel feel comparable. The compression is also the problem. To create one number, a vendor must decide how to weight prompts, engines, appearance, placement, citation, sentiment, market, and time. Different reasonable weights can produce different scores from the same observations.
A score is not automatically dishonest. It can be a useful index when its inputs, weighting, coverage, and changes are disclosed. But the number should not outrun the evidence beneath it. A polished 73/100 can conceal that only a handful of answers ran, one engine dominated the weighting, or several samples failed.
This reporting framework keeps the decision closer to the evidence. Report fractions by engine and question cluster, citation gaps with their supporting answer counts, completed versus planned samples, verified access when available, and referred visits with a date-window denominator. The summary is a plain-language bearing—protect, investigate, review the questions, or build authority—not a synthetic claim of rank.
The practical template for reporting AI search visibility to clients and leadership shows how to carry those definitions, denominators, method changes, and limitations into a deliverable built with any tool.
Interpretation
Reconcile only the evidence you have
| Observed pattern | Defensible reading | Next move |
|---|---|---|
| Sampled appearance + referred visits | The measured path is producing an observed outcome. | Protect cited pages and test adjacent buyer questions. |
| Sampled appearance + no recorded visits | Appearance exists in the sample; traffic impact is not established. | Review citations, calls to action, attribution, and the window. |
| Referrals or verified fetches + no sampled appearance | The prompt panel may not cover the real discovery path. | Use landing pages and logs to improve the questions. |
| No evidence in connected lanes | Nothing was observed in this window. | Improve access and authority; preserve the baseline. |
If a lane is missing, the combined call narrows. Brandvane does not turn “not connected” into zero, and it does not synthesize a crawler or referral conclusion from answer samples alone.
Claim boundary
What Brandvane cannot know
No AI search visibility measurement can directly know:
- every private prompt buyers ask or how frequently each variation occurs;
- every answer delivered to every user, market, account state, or conversation;
- whether the next generated answer will repeat a sampled observation;
- whether a crawler fetch caused a later citation or recommendation;
- the assistant answer behind a GA4 referral unless the journey supplies that context;
- AI influence that ends in direct traffic, branded search, an offline action, or no visit; or
- causation from a before-and-after trend without a design capable of supporting it.
Brandvane can know what its disclosed samples returned, what verified logs recorded, and which recognizable referrals GA4 attributed during a defined window. That is narrower than omniscience. It is also enough to replace guesswork with a repeatable operating baseline.