Choose the unit of measurement
A prompt is a question, not a person. A sampled answer is an observation, not a ranking that every buyer sees. Decide whether you are measuring a brand mention, a linked source, an explicit recommendation or a correct product fact. Keep these fields separate in the underlying record even if your reporting tool also offers a combined score.
Define what qualifies. A competitor comparison may mention you only as an unsuitable option. That is different from being recommended. A citation to your blog may help someone understand a category without endorsing your service. Classification should preserve these differences so the response you choose matches the observation.
Build the sample around buying decisions
Start with the market, language, company size and use cases that matter to revenue. Create a fixed set of discovery, comparison and factual questions. Include non-branded prompts so the sample does not reward you simply for putting your own name in the question. Keep a separate accuracy set for product claims.
Repeat the same prompts on the same platforms and record the collection method. If you add a platform or double the number of prompts, mark the change before comparing aggregate percentages. A larger denominator can change a score without any page or answer getting worse. Preserve older samples rather than rewriting their context to match a new reporting format.
Store enough evidence to investigate a change
For each observation, store the prompt, answer, platform, collection time, location, cited page URLs and classification. Link a correction task to the affected source. This turns a decline into something a writer, developer or product specialist can inspect.
Imagine an answer suddenly says your product lacks an integration. If the source is a retired help article, the next step is to update or redirect that document appropriately. If the source is a current pricing page, the plan entitlement may genuinely differ. If the answer offers no supporting source, you have an uncertain observation rather than proof that a particular competitor displaced you.
Keep business attribution honest
Use analytics to examine referral visits from answer engines where available, then connect conversions to qualified enquiries using your normal definitions. A platform can influence a purchase without sending a trackable click. Conversely, a referral visit can be irrelevant. Show these limits rather than assigning every direct or unattributed lead to AI discovery.
Separate leading observations from commercial evidence. Citations and accurate mentions can help explain discovery, but a booked call still needs qualification. Record what the prospect says about how they found you without forcing the answer into a single-channel story. Keep personal customer information out of general visibility reports.
Use a weekly decision report
A practical report has four parts: work delivered, answer observations, website/referral performance and the next decisions. Show changes against the same baseline, note the sample size and attach representative evidence. Include incorrect claims still awaiting correction and pages that remain undiscovered.
End with a short backlog. One action might be to improve a missing implementation explanation. Another might be to verify a competitor claim before changing your own positioning. If evidence is too sparse to support a conclusion, continue the sample and say so. Reporting is useful when it changes what the team does next, not when it produces a larger chart each week.