AI visibility tracking template: prompts, citations and failed checks

Use this template to record what an AI search product actually returned for a fixed set of questions. Keep the wording, context, answer and sources so another person can inspect the observation. A saved sample is evidence of that test, not a universal ranking or a count of everyone who saw the answer.

The template is free to copy into your own spreadsheet. It supports the observation part of an AI visibility audit; it does not replace the website checks or an enquiry report. Start with a small set you can review properly.

Set the rules before testing

Write down the business, the buyer questions, the products and the market you intend to observe. Separate branded questions, such as “What does Example Workshop offer?”, from unbranded supplier questions. Keep those groups separate in your reporting: a branded mention answers a different question from being suggested to someone who has not named you.

Choose your repeat schedule in advance. Record whether search was enabled and the relevant location, language, model or product mode and signed-in context where known. Use “unknown” when you cannot inspect a setting. Do not pretend every assistant has a clean, controllable country setting.

Copy these columns

Copy this header into a plain-text file ending in .csv, then import it into your spreadsheet. Keep one row per planned test. If an answer contains several sources, keep them in the cited-URLs cell or a separate sheet linked by test ID.

test_id,question_id,repeat,question,question_type,product,mode,market,context,observed_at,status,brand_mentioned,own_site_cited,cited_urls,evidence_ref,accuracy_issue,next_action

A fictional example with a failed check

The following example is invented to show the arithmetic. “Example Workshop” is a fictional business, “Product A” is a placeholder and none of these rows is a test of an actual search product. Three unbranded buyer questions were each scheduled twice under the same declared conditions.

Completion coverage is 5 of 6 planned checks, or 83.3%. The brand appears in 3 of the 5 completed answers, or 60%. Its website is linked in 2 of the 5 completed answers, or 40%. The failed check is shown separately. It is neither a successful omission nor a citation.

Those percentages describe this tiny, fictional sample. They are not market share, search volume, probability of being recommended or evidence of a trend. The differing repeat for Q3 is worth retaining, not deleting because it weakens the result.

Keep the labels honest

A mention names the business. An own-site citation links to its website as a source. A third-party article discussing the business is a different cited source; log it without calling it an own-site citation. Open the citation to check whether it supports the claim, rather than counting every link as favourable evidence.

A search result that does not present an AI answer should have its own status. Do not mix it silently with tool failures or successful assistant answers. Declare which statuses enter each denominator and report the exclusions. If access fails, record that limitation instead of changing the question until the tool produces a result you like.

Compare like with like

Keep the question set and conditions stable for a comparison. If a product, prompt, market or method changes, mark the break. Keep the old record. Report the numerator and denominator alongside any percentage, and separate platforms when their behaviour or coverage differs.

Bing’s AI Performance reporting is a separate platform source for citations across its supported experiences. Do not combine its counts with manual sample rows as if they shared a population. Likewise, use a distinct report for AI referral visits and enquiries. An answer observed in testing did not necessarily send a visitor.

Turn the log into a useful decision

Prioritise wrong prices, contact details and service descriptions, then missing evidence on pages buyers need. Link each proposed fix to the source you control and keep publication verification separate from any later answer change. For help setting up a bounded review and implementing the findings, see Blue Canvas’s AEO service.