Your company appears in an AI answer. Before calling that a win, ask what a prospective customer actually learned. Was the service described correctly? Did the answer link to a useful page? Could that visitor take the next step?
A useful AI search visibility scorecard connects those questions to decisions. Over 30 days, your goal is to establish a repeatable baseline, identify fixable problems, and determine what deserves more investment. Thirty days is a measurement window, not a promise of rankings, citations, or revenue.
Our support guide, How AI Search Visibility Is Measured with Signal Scanner, introduces the diagnostic. This guide adds a provider-independent, manual workflow your team can run with a spreadsheet, saved answer evidence, and your existing analytics. It does not require assuming that any scanner automatically collects every metric below.
Define five signals before counting anything
Keep these measures separate:
- Mention: An answer names your business. Count answers containing a mention, not every repetition of the name.
- Citation: An answer provides a source link to your website. Record the destination and distinguish owned-site citations from third-party coverage of your brand.
- Accurate description: An answer describes your business correctly against an approved fact sheet: services, location, audience, and important limitations. A name alone does not qualify.
- Referral: Analytics records a visit associated with an identifiable AI source. This measures observed traffic, not everyone who saw an answer.
- Business outcome: A visitor completes a meaningful action, such as an inquiry or purchase. Confirm lead quality or sales in your business records; a form submission alone does not establish either.
A high mention count with incorrect service descriptions calls for corrections. Citations to a weak landing page call for a better customer experience. Neither automatically justifies publishing more articles.
Week 1: Build a baseline you can repeat
Choose ten questions drawn from sales conversations and customer research. Include discovery, comparison, and decision-stage questions. Keep branded fact checks separate from nonbranded discovery: asking about your company by name should not inflate its apparent discoverability.
For example, “Which commercial cleaning companies serve Jacksonville medical offices?” tests a different need from “What should a medical office ask before choosing a cleaning company?” Use questions relevant to your actual customers, not prompts engineered to produce your name.
Select three specific answer surfaces for an initial 30-check round. Treat Google AI Overviews and Google AI Mode as different surfaces. Google explains that their responses and supporting links can differ, and an AI Overview does not appear for every query. See Google’s AI features documentation.
Run two baseline rounds on separate days before making changes. For every check, save:
- Exact prompt, intent category, and prompt ID.
- Platform, surface, displayed model/version, and whether web search was enabled or observed.
- Test location and location permissions, language, account/personalization state, and fresh-chat status.
- Timestamp and time zone, full answer or screenshot, citation URLs, and reviewer notes.
If a platform does not disclose its model, record “not disclosed.” Keep settings consistent where possible, and log anything you cannot control. Keep API tests separate from consumer-interface checks. Retain unfavorable answers and failed checks. A collection timeout is unknown; an observed answer that omits your brand is an actual absence.
Week 2: Find the problem behind the number
Review the evidence by customer question, not just by platform total. Open each cited page. Check whether it supports the answer and whether the customer can verify the claim there.
Create three short worklists:
- Incorrect facts: Wrong service area, outdated offering, or unsupported credentials. Identify the authoritative page that should state the correct information.
- Missing answers: An important buying question has no useful public resource. Ask a subject-matter expert to supply the actual process, constraints, or evidence.
- Broken next steps: The cited page is outdated, inaccessible, or unclear about how to inquire.
Record competitors that appear for the same questions, but avoid treating every competitor citation as a content assignment. First ask whether your business genuinely serves that need.
Prioritize two or three fixes by customer impact. Each needs an owner, a URL, a change date, and a reason. This creates a useful change log rather than a list of speculative optimization tactics.
Week 3: Make focused improvements and verify measurement
Correct the source pages, answer the selected customer questions, and repair the next step. Keep unrelated redesigns and large content launches out of this small measurement cycle when possible.
Check that important pages can be discovered and indexed. Google says supporting links in its AI features require pages to be indexed and eligible for a search snippet; there is no special AI schema requirement. Eligibility still does not guarantee inclusion. Review Google’s technical requirements for AI features.
In GA4, inspect Session source / medium for identifiable AI traffic and review its landing pages. Use the same session-based scope each time; first-user acquisition answers a different question. Google explains traffic-source scopes here.
Verify a meaningful key event, such as a completed inquiry, with a test submission. Google’s key-event setup guide includes verification steps. Separate test activity from business reporting. If tracking starts now, begin its baseline now; missing historical measurement is not zero performance.
For a session-based lead rate, divide AI-source sessions containing the verified lead event by measured AI-source sessions. Report both counts, using the same date range.
Week 4: Compare like with like
Repeat two rounds using the original prompts, surfaces, settings, and similar time windows. Compare mention and citation rates on matched checks that returned answers in both periods; report changes in answer availability separately. If you add a platform or change the questions, establish a separate baseline.
Show numerator and denominator together. Suppose one illustrative round has 30 scheduled checks, 28 completed observations, and 24 AI answers. Four observations showed no AI answer; two checks failed. The following numbers demonstrate the method only, not client results:
| Measure | Illustrative result | What it means |
|---|---|---|
| Collection coverage | 28/30 | Two checks remain unknown |
| AI answers observed | 24/28 | Four completed checks produced no AI answer |
| Brand mention rate | 8/24 | Eight observed answers named the brand |
| Owned-site citation rate | 5/24 | Five answers linked to the business’s website |
| Accurate descriptions | 6/7 | Six of seven substantive descriptions passed review |
| AI-referred leads | Not measured | Tracking is unavailable, not zero |
Use the same definitions in both periods, and report coverage beside every comparison. Split results by surface and intent before interpreting a pooled percentage. This is a directional sample, not market share. A tiny sample can swing after one answer changes; report counts and avoid confident causal claims.
Add Google’s reporting without overstating attribution
As of October 2026, Search Console has a dedicated Generative AI performance report (Search) covering impressions from AI Overviews and AI Mode. It includes page, country, device, and date views. Google says the report may be absent when a site has insufficient impressions. Record that as unavailable, not proof of zero visibility. Its export can also turn unavailable values into zeros, so preserve the original status. Read the current report documentation.
Keep these impressions separate from your sampled prompt tests and GA4 visits. They measure different things. Missing referral information can be classified as direct traffic, so analytics cannot establish every AI-assisted customer journey. See Google’s traffic-source processing guidance.
End day 30 with a decision, not a trophy score
Summarize what changed, which evidence supports it, and what remains unknown. Decide what to fix next, what to continue measuring, and what to stop doing. Describe observed movement without claiming that your edits caused it.
That evidence-led review fits Tridence’s SEO 3.0 approach. If your team needs help turning scattered answers into a practical improvement plan, contact Tridence with your priority questions and current measurement gaps.
