Build your AI visibility baseline

Choose one question a buyer might ask, record a completed AI answer, and check the page that should help answer it. That gives you a specific issue to investigate before buying software or commissioning a batch of content.

This kit is for B2B SaaS founders and small marketing teams. You can complete the task with a permitted search assistant, a browser, and a spreadsheet. The method is public below; the optional downloads save you the work of setting up the files.

What should your first question test?

Choose either description accuracy or discovery. A branded question asks what a named product does. An unbranded question describes a buyer’s task without supplying your product name. Keep those purposes separate when reading the answer.

Write where the question came from. A sales question or a recorded search query is observed evidence. A question you generate for this exercise is a hypothesis. It can still be useful, but it doesn’t establish demand or search volume.

Give the question an ID, such as Q01, and keep its wording unchanged for the first repeat. In the buyer-question sheet, record the intended decision, origin, evidence date, branded status, and panel version. You should be able to explain why that question belongs without pointing to a score.

What belongs in the answer record?

Save the complete answer before interpreting it. Record the question ID, attempt time, platform, consumer app or API, visible model, search setting, language, and session context. Mark anything the interface doesn’t show as unknown. Store evidence somewhere the people doing the review can access without exposing customer information.

For a consumer check, start a fresh conversation. Inspect its memory and personalization state rather than assuming a new chat is neutral. OpenAI’s crawler documentation also distinguishes search crawling from model training and user-triggered visits. Allowing one doesn’t establish the others.

Use one row per attempt. A failed request stays in the sheet with its failure reason and blank outcome fields. A completed answer needs a separate review of each outcome:

Outcome	Check	Record
Mention	Does the answer identify your company?	Yes, no, or unknown
Owned citation	Is there a displayed link to a verified owned URL?	Yes, no, or unknown, plus the link
Recommendation	Does the answer suggest the product for the stated need?	Yes, no, or unknown, with the passage
Accuracy	Do material claims match current product facts?	Match, mismatch, or unassessed, with evidence

A branded answer describing services isn’t automatically a recommendation. A citation to your homepage doesn’t prove that the answer used its current contents. The AI visibility measurement article explains how these distinctions affect a comparison.

What does a completed record look like?

JAC’s initial records show two different kinds of evidence. On October 9, 2026, the native Google generative AI report showed 24 property impressions for September 9 through October 6. Separately, one unpersonalized ChatGPT consumer search answer used retired JAC service labels while linking to the homepage and Services page.

Those observations don’t belong in the same percentage. The first is a platform report over a date window; the second is a reviewed answer to one branded question. Neither establishes new leads or a growth effect.

The completed JAC file preserves that separation. It also records the main search-service URL as available to Google in a live test, while its index record still said it was unknown. You can inspect the JAC experiment log for the dated work and its limits. Use the question and observation sheets for your own answer checks. Keep native reports and page checks in their separate records.

Which page should you inspect?

Inspect the page that answers the buyer’s decision, not every URL an assistant happens to mention. Confirm its main answer, product limits, supporting facts, working links, intended canonical URL, and a relevant next step. Compare the published text with the source documentation before calling a claim wrong.

When account access is available, check indexing separately from a browser fetch. Google’s generative search guidance ties eligibility to indexing, snippet eligibility, and inclusion in its Search generative AI features. A successful public fetch alone doesn’t establish those conditions.

Write one repair in the page-check sheet. Attach the question and observation IDs, evidence, change owner, required access, and completion test. For JAC’s stale service description, the controlled work is to keep current service facts clear on the cited pages and check the relevant discovery records. A later assistant adopting those facts is an outcome to observe, not a completion promise.

How should you use platform reports?

Keep native reports in their own records with dates, filters, and aggregation. The Google report instructions distinguish property totals from page totals and warn that unavailable values can become zeros in exports. Save the displayed state before interpreting a spreadsheet.

Bing’s AI Performance documentation describes aggregated citations across supported surfaces. Citation counts don’t establish rank or the role of a page in one answer. Check which reports your account provides before adding a metric to the sheet.

Visits and inquiries need separate analytics and qualification records. Keep known test visits out of demand claims. Leave the discovery source unknown when it can’t be established. The audit workbook helps connect an observed description issue to an agreed repair.

When is your baseline ready to repeat?

Repeat only after you’ve recorded enough context to recognize the same check. Keep exact questions, surfaces, settings, and review definitions fixed. Record a page change and its publication date. A changed question or model needs a separate comparison, and an unavailable later answer stays missing.

Count assessed outcomes over assessed answers. List failed attempts and unreviewed answers beside the counts. A small panel can reveal a useful issue without representing every buyer or supporting a market-wide visibility score.

Finish one question, one answer record, and one page check first. The full search and SEO guide covers the ongoing page work. If you’d like that research, implementation, and reporting handled, explore AI Search & SEO.