Measure what exists, say what it means
After this lesson you can build a small, honest measurement routine for AI visibility, and write a report that does not claim more than it knows.
The theme is humility. These systems are probabilistic, personalised, change often and are only partly documented, so every number has caveats. Dated 2026-11.
What the vendors give you (documented)
Google Search Console. Google's AI features page says traffic from AI Overviews and AI Mode is included in the overall Search performance data under the Web search type. In June 2026 Google also introduced Search Generative AI performance reports. According to its post, they show impressions in AI Overviews, AI Mode and Discover AI features, by page, country, date and (for Search) device. Clicks and queries are not listed in that report, per the post. The post says the reports reached all sites worldwide by the end of August 2026. Verify menu names in your own account; they have been changing.
Bing Webmaster Tools. The AI Performance report, in preview, shows citations of your content in Copilot, Bing and select partner experiences, plus grounding queries and counts. The June 2026 views (Intents, Topics, Citation Share, Compare) add grouping. Microsoft says Citation Share is observational: it is not a ranking, a traffic share or a quality score and does not show competitors or causes.
Other assistants. I found no comparable official publisher dashboard for other major assistants. Anything else you see is third-party estimation.
What analytics can show (established practice, with limits)
Your web analytics can report visits by referrer. Some assistants send referrer information and some do not, and tools may file visits under different channels. Practical approach:
- Look at referral sources for assistant domains over months.
- Treat the totals as a floor, not a total.
- Do not compare periods when tracking or your tool's channel rules changed.
- Never invent conversions. Count only what your analytics or orders actually record.
Server logs can also show crawlers by user agent. Remember that user agents can be forged, and a crawl is not a citation.
Manual prompt testing
Third-party "AI visibility" tools mostly work by sending prompts to assistants and recording answers. You can do a small, honest version yourself.
- Write 10 to 20 realistic questions your audience asks, in their words.
- For each, test in the products you care about, signed out or in a clean profile where possible.
- Record: date, product, exact prompt, whether you were cited or mentioned, which other sources appeared, and whether what was said about you was accurate.
- Repeat monthly at most, on the same list.
- Report mentions, citations and accuracy separately.
Date: 2026-11-14
Product: (name and version/mode shown)
Prompt: "How do I percent-encode a value for a URL?"
Result: cited / mentioned / absent
Accuracy of what was said: correct / partly wrong / wrong
Other sources shown: (list)Why results vary
Expect different answers for the same prompt because of: wording changes, the time of asking, the user's location and language, personalisation or history, the product's model version and mode, whether the system searches at that moment, and randomness in generation. Google itself says AI Overviews and AI Mode may use different models and that fan-out issues varied searches. So repeat tests across days before concluding anything, and treat a small difference as noise.
What no tool can do
- Guarantee citations. Only the vendor controls selection.
- Reveal "the AI's opinion of your brand" as a fixed fact.
- Show Google's internal metrics. Google's guide warns about trusting tools that claim to.
- Replace reading the actual answers.
Third-party tools can still be useful as a way to run many prompts for you. If you use one, ask what prompts it sends, how often, from where, how it detects mentions, and what it cannot see. Do not treat its score as a measure of anything the vendor publishes.
Reporting honestly
Good phrasing: "In our test of 15 prompts on 2026-11-14, our page was cited in 3 answers in one product. Results vary by day, so this is a snapshot." Bad phrasing: "We are visible in AI search." Never write "up X%" unless the baseline and method are shown, and never claim traffic or revenue that your own analytics do not record.
This site
The repository has an audit script for technical page quality but no measurement of AI visibility, and nothing in this course states that the site is cited anywhere. If Himanshu verifies the site in Search Console and Bing Webmaster Tools, those reports would be the first honest data source.
Common mistakes
- Reading impressions as visits.
- Celebrating a single screenshot.
- Comparing your Citation Share to competitors; it does not show them.
- Changing five things at once and claiming one caused a change.
- Ignoring accuracy: being cited with wrong information is not a win.
Exercise
- Build the 10-question test list for a fictional bookshop or for your site.
- Run it in two products and record results using the template.
- Write a four-sentence report that states the sample size, date, result and limits.
Show a sample answer
"On 2026-11-14 I tested 10 prompts in two assistants. Our site was cited in 2 of 20 answers. Two answers misstated our delivery area. The sample is small and answers vary, so I will repeat in a month before drawing conclusions."
Previous: Rendering, speed and clean HTML. Next: Risks, ethics, hallucinations and opting out.
Key points
- Google's generative AI performance reports in Search Console show impressions by page, country and date, but not clicks or queries (per Google's June 2026 post).
- Bing's AI Performance report shows citations and grounding queries for Microsoft's AI experiences and is observational, not a ranking.
- Analytics can show referrals from some assistants, but attribution is incomplete and varies.
- Manual prompt tests are small, noisy samples; record the date, product, wording and result.
- No tool can guarantee citations, and tools claiming internal vendor metrics should be treated with caution.