Before comparing this month’s AI numbers with last month’s, write the unit beside each total. A crawler request means something different from a citation or a visit, and a customer inquiry means something different again. Combining them under “AI traffic” hides the steps between a system accessing your page and a person contacting the business.
Keep a small evidence ledger with separate rows for access, visibility, visits, and business outcomes. Each row should name the source of the number and the conclusion it supports, which gives you a way to assess progress even when no single report follows the entire customer journey.
Start with the question the owner needs answered
Start by asking whether the right people are finding the business and whether their inquiries are worthwhile. Technical access matters because a blocked or broken page may prevent discovery, and citations matter because they show a page being used or displayed in a supported answer experience. You need both, along with the inquiries themselves, to answer that customer question.
Choose one reporting period and a short list of commercially relevant pages. A service page and one guide related to a buying decision are enough to begin, plus an important location page if you have one. Keep your general traffic report alongside the AI report, so a fashionable category does not crowd out the channels already producing work.
The ledger below keeps different units in their own rows:
| Evidence | Where it comes from | What it supports | Still needs checking |
|---|---|---|---|
| Crawler request | Server or edge logs | A request reached an observed endpoint | A human saw or liked the page |
| Displayed link or citation | A platform’s publisher report | Visibility within that report’s defined coverage | Every AI mention or a universal rank |
| Referral visit | Website analytics | A visit with an observed source | All visits influenced by AI |
| Submitted inquiry | Form system and customer record | A request was accepted | A qualified sales opportunity |
| Qualified opportunity | Staff-updated customer record | The inquiry met your stated criteria | Revenue already earned |
| Won work | Business records | A recorded commercial outcome | That one channel caused the whole decision |
Write the unit beside every number. Report fifty requests, fifty citations and fifty visits as three separate lines, since adding them into “150 AI interactions” mixes crawlers, mentions and people.
Use each publisher report within its boundaries
As of September 2026, Google’s Generative AI performance report for Search reports impressions for AI Overviews and AI Mode. It supports breakdowns including pages, countries, dates, and devices. Google states that the data comes from the Web search type in the ordinary performance report, so those overlapping impression totals should not be added together.
The report helps you identify pages shown in those Google features, and it covers only those features, leaving out every other assistant that may discuss your business. If the report is missing or has little data, write that limitation down. Leave it blank, since a zero would claim your site never appeared anywhere and the report only covers what it observed.
Bing’s AI Performance report covers citations across Microsoft Copilot, Bing’s AI summaries, and selected partner integrations. Bing distinguishes citation counts from rank or authority, and describes grounding-query phrases as a sample. Use them to understand which topics lead to retrieval, and keep market-demand estimates out of it.
Export or record the original metric names. If a platform changes its report or coverage, note the date, since a sudden increase caused by expanded measurement is different from an increase in customer attention.
Keep crawler activity in a technical section
If your website runs through Cloudflare, AI Crawl Control can show crawler requests and response outcomes. Its documentation distinguishes all requests, including blocked and error responses, from successful responses. Some referral reporting depends on the plan.
For an owner, the useful question is often whether important public pages can be accessed as intended. A spike in failed requests may need an investigation, while a rise in successful requests may simply mean a crawler revisited existing pages more often.
Record the affected page, crawler identity as reported by the tool, response pattern, and any action taken. Keep calling them requests and avoid renaming them “readers,” because the log records a crawler fetching the page, and whether a person saw the answer or contacted you shows up elsewhere, if at all.
Access policy is a separate decision from measuring value. The Cloudflare AI Crawl Control guide explains the policy choices. Keep that decision visible if a reporting change follows a new block or allow rule.
Follow observed visitors to the next meaningful step
In website analytics, inspect source and referral data for the AI products you can identify. Keep the raw source as well as any grouping your team creates, since product domains and traffic patterns can change and a custom “AI” group needs periodic review.
Some visits will arrive without a useful referral label. Someone may copy a link or return later on a different device. That missing information is a measurement limitation, and we would leave unknown traffic labeled as unknown even when the timing makes AI seem plausible.
When someone submits an inquiry, store the source information available at that point in the customer record. You can also ask an optional, short “How did you hear about us?” question and keep the answer in a separate field from analytics attribution. A customer saying “an AI assistant” is useful reported evidence, even if the visit was recorded as direct.
The website lead tracking example explains how to connect a visit with an accepted record. That delivery check matters because a browser event can fire even while the business fails to receive a usable inquiry.
Work through a hypothetical month
For example, suppose a service business records 240 Google AI-feature impressions, 18 Bing citations, six website visits with an identifiable AI referral, and two inquiries whose records retain such a source.
One inquiry is outside the service area. The other is a good lead that is still waiting for an estimate, so the report shows one qualified inquiry and no sale yet. It would be wrong to report two customers from AI, or 258 people who saw the business.
Now suppose another customer reports finding the business through an assistant, but their website session has no usable source. Record one self-reported AI discovery on its own line. Leave it out of the measured referral count, and keep it in the report even though the browser data disagrees.
The useful decision might be to keep the relevant service page accurate and improve the explanation of the service area. With this evidence, we would wait before forecasting a monthly revenue contribution or calling the channel proven.
Add a change log beside the numbers
Each reporting period should include a short list of meaningful changes, such as a service page published or outdated hours corrected, even a new guide shared with an existing audience. Date those changes and link to the affected pages.
A change log helps you form hypotheses, which you then test against later results. If citations rise after a page update, write “citations increased after the update” and identify other changes that could matter. Claiming the update caused the increase requires more evidence than a before-and-after chart.
For persistent factual errors in answers, use a business information correction log. Accuracy work and acquisition measurement benefit from sharing evidence, while each keeps its own definition of success.
Choose the next month’s action
End the review with one practical decision, picked from options like these:
- Continue a page that attracts relevant inquiries.
- Clarify a page that attracts the wrong ones.
- Investigate an access failure.
- Improve a promising guide with a concrete example.
- Leave a low-volume signal under observation when there is not enough evidence to act.
Report unknowns in plain language: source missing, report unavailable, inquiry not yet qualified, or outcome pending. We believe an explained blank helps an owner more than a precise-looking estimate with no source behind it.
Measuring AI search helps a business decide where effort belongs. Keeping the evidence separate lets the team act on what it knows, and crawler activity and citations stay in their own rows until a real customer appears in the records.