1. Define the local scope before testing
Record the business name, website, service category and city or service area. Include known name variants so a similarly named business is not mistaken for the client. Ask the client which services and locations matter commercially.
Build a small, fixed set of questions across discovery, specific services, urgent problems, comparisons and price. For example, “best roofing in Raleigh” is a question from the fictional Raleigh roofing sample. It is an illustration, not verified search demand.
Keep branded questions separate from unbranded discovery questions: asking about a named business tests a different situation from asking which local businesses to consider. Do not turn prompt counts into estimated search volume.
2. Keep the exact question and observation conditions
Save each question verbatim, the date, provider, model or interface where disclosed, location settings, and whether live search actually ran. Record the complete answer and visible source URLs. An API using web search and a signed-in consumer ChatGPT session are different test conditions.
Repeat the same questions under consistent conditions to observe variation. PresenceCurrent’s Full Audit V2 design plans 25 prompts with three repetitions each: 75 observations. Three repetitions do not establish statistical reliability or a stable market ranking.
If search use cannot be confirmed, label it unknown. Do not call a model’s recalled answer a live search measurement. Keep personalized or follow-up conversations separate from fresh-session tests.
3. A failed request is not a business absence
For every attempt, keep its status and a usable-response decision. A successful answer that does not name the client is a measured absence. A timeout, failed request, skipped attempt or malformed answer is not.
Observed visibility is target-mentioned usable responses divided by all usable responses. In the approved synthetic sample, 20 / 71 = 28.2%. Four of 75 planned observations are excluded. These are fictional demonstration counts, not a customer result or a completed Full Audit.
Report the denominator beside the percentage. Also inspect how many distinct questions include the business: appearances across repeated answers and breadth across questions answer different client questions. Never replace unknown observations with zero.
4. Review mentions and citations separately
A business name in an answer is a mention. A source URL is a citation. A source appearing elsewhere in the response does not automatically support the client’s mention. Inspect the saved relationship in that same observation before making a business-specific claim.
In Synthetic Demo V3, 59 of 71 usable responses have saved sources: 83.1% evidence coverage. That measures source availability, not accuracy, client visibility or business-linked support. A high coverage percentage cannot validate the content of every source.
Open the saved-evidence inspector, select a question and repetition, then use View context on a source. Read the complete synthetic answer as well as the extracted business relationship. See citations versus business mentions for the distinction.
5. Give the client an inspectable summary
Use this reporting pattern, filling it only with actual observations:
Within [questions, location, provider/interface and dates], the business appeared in [X] of [Y] usable answers. [Z] attempts were excluded. [N] usable answers included saved sources. Here are the specific responses behind the finding, the limitations and one hypothesis to investigate next.
Keep observed businesses separate from verified direct competitors. Separate a possible website improvement from a demonstrated cause. A later comparison needs compatible prompts, provider conditions and denominators; a percentage change alone does not prove an SEO edit worked.
AI visibility is one layer of reporting. Keep it alongside search impressions, qualified visits and enquiries. An AI mention does not establish that anyone saw the answer or contacted the business. Explore the agency review workflow before deciding what belongs in a client deliverable.
What this workflow cannot establish
A bounded audit cannot measure all users’ answers, guarantee recommendations, or explain a provider’s internal reasoning. PresenceCurrent’s public example is synthetic; it demonstrates inspection, not current live monitoring or cross-provider coverage.
For the underlying search context, read Google’s guidance on AI features and website eligibility and Bing’s explanation of observed citation activity. Neither makes sampled answer mentions equivalent to customer traffic.