1. Choose the question before the metric
Start with the decision, then choose evidence that can answer it. A public-page check describes implementation. A provider report describes activity within its defined coverage. A manual observation describes one recorded answer. An operational record describes a business outcome. These layers can inform the same investigation, but their units and denominators are different. The following lab is entirely hypothetical: its page, counts and outcomes are teaching examples, not Suprara results or connected account data.
2. Read the readiness score correctly
Suprara fetches one public URL’s initial HTTP/HTML, robots.txt, and one declared or default sitemap. Its 18 checks span six equally weighted dimensions: Identity & metadata, Indexability, Content structure, Link discovery, Accessibility signals, and Structured data. Each dimension scores passed checks divided by passed plus failed checks, expressed as a percentage. The overall score averages evaluated dimensions. Unknown and not-applicable checks are excluded; always read coverage beside the score. Missing optional canonical, sitemap, or JSON-LD features do not automatically fail. An unusable page or fewer than three evaluated dimensions means no overall score. No JavaScript rendering, full-site crawl, or external model is involved. Save the scan PDF with its evidence, timestamp, scope, and rubric version.
3. Lab: classify four evidence cards
A · HTML: a fictional service page returns a readable heading and contact link; one content check is unknown because the initial HTML omits a JavaScript-rendered section. B · Provider report: a fictional Search Console export records 120 generative AI impressions for a defined period. Preserve the property, dates, filters and export alongside it. C · Manual observation: twelve planned requests produce ten completed answers, three brand mentions, one linked citation and two failed requests. D · Business outcome: the receiving team verifies four inquiries, classifies two as qualified and records one sale; the acquisition source is unknown. Before reading the answer below, write one supported conclusion and one unsupported claim for each card.
4. Worked answer: keep the units separate
A supports what was present in the fetched HTML; inspect the rendered section separately. It does not prove visibility in an assistant. B supports reported exposure within that report, not 120 people, visits or sales. C gives a 3/10 mention rate (30%) and a 1/10 linked-citation rate (10%) among completed answers; disclose both failures separately. Neither rate estimates market share. D supports two qualified inquiries and one recorded sale under the team’s definitions, without attributing either to AI. Do not divide the sale by provider impressions: the records do not establish a common visitor cohort. A later increase could reflect changed coverage, prompts, demand or tracking rather than the page edit alone.
5. Missing, zero and unavailable are different
Google documents a generative AI impressions report for AI Overviews and AI Mode. Insufficient impressions can leave it unavailable. Exported zeros can also represent unavailable values shown as ~ or - in the interface. Keep that distinction. In this lab, an unavailable report is not evidence of zero exposure; a failed manual request is not a completed answer without a mention; an unknown HTML check is not a failure. Record status and scope before calculating. If exclusion is a possible explanation, ask an authorized owner to inspect the Search generative AI setting, including inheritance; do not change it merely to complete a report. Its inclusion control is distinct from Google-Extended and does not guarantee appearance. Suprara’s page scanner does not connect to or import provider reports.
Apply it in your own notes
Create four rows, one per lab card. Record source, date, unit, denominator, coverage, status, supported conclusion and next check. Then replace card B with “report unavailable” and card C with “zero mentions in ten completed answers.” Explain why these are different findings. Answer key: B has no usable exposure measurement; C has an observed zero for that sample. Keep the failed requests and unknown acquisition source visible. Never fill either gap with an invented zero or attribution.
Get a worksheet →Practice questions
1. A public-page scan score rises from 50 to 70. What is established?
2. Three mentions in ten completed answers means what?
3. Why keep the rubric version?
4. The provider report is unavailable. What can you conclude?
Primary sources
Google Search Console Help: Generative AI performance report (Search)
Google Search Central: Optimizing for generative AI features