Is AI search sending you anyone?

Help a made-up home-energy site check what AI reports say about visits to its pages.

Start the home-energy case

Real example: WireRef’s three reports

WireRef, three reports that count different things

  1. Google AI featuresOct 1–28, 202612impressions
  2. Bing AI PerformanceOct 1–28, 2026149Kcitations
  3. Clarity, visits from chatgpt.comone three-day export, Oct 31, 20262sessions

WireRef: Google and Bing through Oct 28, 2026; Clarity, one three-day export of Oct 31, 2026.Check the visits report before claiming traffic from AI search.

What would you recommend?

Read the reports, choose an action, then explain your reasoning. Use the analysis prompt and worked answer to check your thinking.

Fictional practice data. K means thousand. Not results from my sites or clients.

Beginner: A citation is not a visit

A fictional home-energy guide appears in AI answers. A citation is a reference to its content in an AI response. The owner asks how many visitors came from those answers.

Choose what these two reports let you say about AI search and visits to the site.

Check the source evidence
Fictional reports for the same complete 28-day period; the sources measure different events.
SourceMeasureCount
Bing AI reportCitations10K
Site analyticsSessions attributed to chatgpt.com100
Which sentence can you put in the report?

Analyze the reports

Download the reports and attach them to your AI tool. Paste the analysis prompt, then check its answer against the evidence.

Download the case and reports

Read or select the prompt

Then check its work. Check the reporting periods, one important source row and any calculation. Does its recommendation actually follow from that evidence?

Help checking the analysis

  1. Download the complete case and give it to an AI tool with the analysis prompt.
  2. Ask AI to organize the evidence. Check the dates, one important row and any calculation it uses.
  3. Choose or reject its recommendation. Explain the page action you would take and what you would check next.

If you’re stuck: Read each count together with the source and measure in its row.

Explain your decision

Use the evidence to explain your choice and next check. Your writing is not submitted or graded.

Keep the service name, number and unit together.

Explain what citations and sessions do not establish.

Report the observations without adding unlike counts or connecting unrelated services.

Keep your work

Choose an answer and write your reasoning. Compare with the worked answer.

Notes are not saved automatically. Download your notes before leaving.

Intermediate: Do not add overlapping lookbacks

A fictional energy site checks referral traffic on two consecutive days. Each export looks back three days, so the dates overlap.

Decide whether these two overlapping exports can give you a total for June 1–4. If not, ask for the report you need.

Check the source evidence
Fictional analytics exports; these are totals for lookback windows, not new daily sessions.
ExportCovered datesChatGPT-attributed sessions
First pullJune 1–3100
Next pullJune 2–4100
What would you request before reporting the June 1–4 session total?

Analyze the reports

Download the reports and attach them to your AI tool. Paste the analysis prompt, then check its answer against the evidence.

Download the case and reports

Read or select the prompt

Then check its work. Check the reporting periods, one important source row and any calculation. Does its recommendation actually follow from that evidence?

Help checking the analysis

  1. Download the complete case and give it to an AI tool with the analysis prompt.
  2. Ask AI to organize the evidence. Check the dates, one important row and any calculation it uses.
  3. Choose or reject its recommendation. Explain the page action you would take and what you would check next.

If you’re stuck: Think about which dates appear in both files.

Explain your decision

Use the evidence to explain your choice and next check. Your writing is not submitted or graded.

Explain why the two totals cannot simply be added or averaged.

Specify the dates, analytics source, session measure and referring-site filter.

Keep the timezone and counting rules consistent; do not invent missing settings.

Keep your work

Choose an answer and write your reasoning. Compare with the worked answer.

Notes are not saved automatically. Download your notes before leaving.

Advanced practice: Design a measurement you can stand behind

You handle reporting for a fictional home-energy publisher. Before testing a revised heat-pump cost explanation, the editor needs a reporting plan. Check whether the numbers were counted consistently, identify missing settings and propose a task that tests whether readers understand running costs. The packet includes visibility reports, referral sessions and a small usability observation.

Find what changed in the measurement setup. Decide which comparisons are still usable and what must be checked before trusting the next report.

Check the source evidence
Fictional June 1–28 reports. Metric definitions and collection details are supplied in the attached log.
SourceObservationDefinition or collection detail
Google platform export5K impressionsOccurrences in the provided AI visibility report; not visits.
Bing platform export10K citationsDisplayed citations; no shared event identifier with analytics.
Analytics source report100 ChatGPT-attributed sessionsAn analytics source group. Inspect its dated inclusion rules in the collection log.
Usability notesParticipants struggled to compare annual running costs.Small convenience sample; no baseline task-completion rate collected.

Inspect the fictional metric definitions and collection log (CSV)Inspect the metric definitions and dated collection rules before choosing a comparison. No event-level join between platforms and analytics is provided.

After inspecting the collection log, which specification best meets the editor’s brief?

Analyze the reports

Download the reports and attach them to your AI tool. Paste the analysis prompt, then check its answer against the evidence.

Download the case and reports

Read or select the prompt

Then check its work. Check the source periods, a decisive row and any calculation. Challenge its strongest assumption and ask what would reverse the recommendation.

Help checking the analysis

  1. Download the complete case and give it to an AI tool with the analysis prompt.
  2. Ask AI to organize the evidence. Check the dates, one important row and any calculation it uses.
  3. Choose or reject its recommendation. Explain the page action you would take and what you would check next.

If you’re stuck: Check whether each metric keeps the same definition across its comparison period. Then ask which reader outcome the content test will observe.

Explain your decision

Use the evidence to explain your choice and next check. Your writing is not submitted or graded.

List its source, dates, counting rules and limits; include referral settings where relevant.

Use the dated log; explain whether earlier data can be counted under the same rules or needs a marked break.

Name missing settings needed to interpret the referral report rather than guessing them.

Define a running-cost task, a fair comparison, observable success and what the result cannot establish.

Keep your work

Choose an answer and write your reasoning. Compare with the worked answer.

Notes are not saved automatically. Download your notes before leaving.

Something wrong, or a different approach? Include your evidence and lesson link. Leave out private information.

Optional: explore a worked real report

Use this as a reference when checking your analysis. You do not need to complete a second set of notes.

Would you change the plan for a bigger AI number?

Imagine you report on two sites I own: WireRef, an electrical reference, and The Quack Report, an Oregon Ducks football site. Your current assignment is to improve Google visibility and help the people who reach these sites. Someone suggests shifting the content budget because WireRef has a large Bing AI citation count. Start with what Google reports.

Google’s Generative AI performance report counts appearances of links in supported AI features, including AI Overviews and AI Mode. This export combines those features for text-based searches. An impression is not a visit or a unique person.

Google Generative AI performance report: Web (text-based), October 1–28, 2026. Property-level impressions across supported AI features; not visits or a split by feature.
SiteReported impressions
WireRef12
The Quack Report3.2K
Download these rows (CSV)

Now read the Microsoft report separately. A Bing citation is a source displayed in a supported AI answer; the report covers Microsoft Copilot, Bing AI summaries and selected partner experiences. It does not report clicks.

Bing AI Performance, October 1–28, 2026. These are displayed source citations across supported Microsoft AI experiences, not unique people or visits.
SiteReported citations
WireRef149K
The Quack Report20.4K
Download these rows (CSV)
Which recommendation fits the assignment and these reports?

Do the visit rows complete the story?

Clarity records some sessions with an identifiable referring source. In the latest selected pull, it lists the following sessions under the source name chatgpt.com. Each row covers one three-day export collected on October 31, 2026, not the October 1–28 period above.

Clarity source exactly chatgpt.com: individual three-day exports collected October 31, 2026. These are not monthly totals or attributed Google/Bing clicks.
SiteReported sessions
WireRef2
The Quack Report1
Download these rows (CSV)
What could you honestly put in the report?
Why matching the dates would still not produce a citation click rate

A click-through rate needs compatible click and impression events from the same measured population. These files provide Google AI impressions, Microsoft citations and separately recorded ChatGPT referrals. They cannot be joined into a visitor funnel.

If the question is whether known AI assistants send useful visits, obtain a suitable analytics report for the intended full period, keep source definitions consistent, and inspect the landing pages and relevant outcomes. That still does not isolate Google AI-feature clicks from ordinary Google search traffic.

Can you add yesterday’s and today’s pulls?

The collection runs repeatedly, but each pull requests a three-day lookback. Here are two WireRef pulls for the same source name.

WireRef source exactly chatgpt.com: two separate three-day exports. Most of their time coverage overlaps.
Pull collected (UTC)Reported sessions
Oct 30, 10:342
Oct 31, 10:362
Download these rows (CSV)
How would you use these rows?
10-second walkthrough. A drawn operation using the dated export, not an account recording.
Read the video explanation

Citations and visits measure different things.

These three-day reports overlap.

Keep the reports separate.

Guided review: what would you do next?

You can already decline a budget change justified only by a large citation counter. That is a decision, not a reason to ignore AI search. Keep the engine-specific visibility reports, state what the referral data covers, and connect any proposed content work to a page that could better serve a real reader.

Which plan would you send to the person approving the work?
Apply this to your own reports

Write the decision first: are you measuring appearance, a recorded visit, or useful activity after someone arrives? Export the relevant reports for that decision and note their engines, units, dates and time zones. Preserve the original files.

Ask AI to identify which numbers can be compared and which cannot, then check its answer against the report definitions. Do not let it turn citations into people, fill gaps with zero, or add totals from overlapping exports. If the available report cannot answer the question, name the report or observation you need next.

For a proposed new page, test whether the content gap is real. For a page that appears without earning clicks, compare the evidence in the impressions lesson.

Your real-report answer review

Use the explanations beside each question to review your choices, then write your recommendation.

Only your latest choice for each real-report question counts. Your written recommendation is not graded.

Write what you would do next

Write your recommendation

Write a short recommendation that names what you can report now, what you would not claim, and the next page or measurement you would inspect. Your notes stay on this page until you download them.

Compare with a worked recommendation

I would report WireRef’s 12 Google AI impressions from October 1–28, 2026, and keep its 149K Microsoft AI citations on a separate line. Neither figure establishes visits or justifies moving the content budget. I would authorize a page review: use Google’s page report to choose a relevant page, then check which reader question it should answer before proposing more content.

Clarity’s individual three-day exports collected October 31 recorded 2 sessions under chatgpt.com for WireRef and 1 for The Quack Report. To judge useful visits over a full period, I would request a report covering that period, its referring sources, landing pages and relevant outcomes. Adding overlapping snapshots would not supply that answer, and these rows do not attribute clicks to Google or Bing.

My priority is work that helps the reader, with any new investment tied to an identified need. The effects of these proposed decisions have not been measured in this exercise.

Compare the evidence and next check with your own. If you would choose a different action, explain what supports it.

Continue with the Moat Map course

What would you like to know?

Arrow keys move. Enter opens. Esc closes.

Ask Bella

Ask Bella Hayden's corgi, who answers from this site and cites every fact.

BELLA

WOOF. I'm Bella, Hayden's corgi. He's my human. Ask me about him; every answer cites its source.

Reading this page? The top button is my short version of it.

I have written answers to common questions. For other questions, I show what I found on this site. You can open the sources under each answer. What you type stays in your browser. Read every answer on one page

Top