What exactly is the boss trying to prove with attribution?

“Coming from AI” involves at least source identification and effect judgment. Source identification answers where a visit entered from; effect judgment answers whether the work brought in buyers worth talking to. The two are related, but one screenshot cannot stand in for all the evidence.

I break the evidence into a sequence of actions: the brand appears in an answer, the cited page is found, someone enters the website, a valid contact is produced, sales confirms the need, and then later results are reviewed. If any link is missing, that gap stays unknown. Do not take a single citation at the beginning and write it directly as an order at the end.

This breakdown is my proposed measurement approach. If you are looking at a visibility report, first compare the difference between brand mentions, page citations, and traffic to see whether the report crosses the evidence boundary.

Can visits from AI assistants be identified today?

Partly. And current documentation can no longer be summed up simply as “all of it is just referral.” The official GA4 default channel group documentation I checked this time lists “AI Assistant,” covering assistant sources that meet its identification rules, and states that the related medium is ai-assistant. Whether it appears in your data still depends on the actual sources and applicable conditions.

In practice, I keep the channel, source, medium, and landing page together. The channel is useful for aggregation; the specific source helps with follow-up checks. Do not look at one group labeled “AI” and assume it contains every platform and every entry point. Custom rules with overly broad domain matching can also pull in unrelated visits.

The site also needs to record visits normally. If the analytics script did not run, the user did not allow the relevant recording, or the entry point has no recognizable source information, a report cannot restore the facts out of thin air. My approach is to note the recording capability and the scope of what is missing, rather than guessing a flattering source for a missing visit.

Basic data checks can follow the method for analyzing sources and landing pages in GA4. This article does not read back your account; what the official documentation supports and what a specific account actually has complete data for must be stated separately.

Can Google’s AI search be measured separately?

Google AI Overviews and AI Mode must be separated from standalone AI assistants. The current GA4 channel documentation classifies them under organic search, not the AI Assistant category above; Google Search Central states that their performance is included in the “Web” search type in the Search Console Performance report. These two official positions do not support simply renaming total organic search as AI search volume.

I keep the original organic search definition first, then compile observable AI answers and citation records separately. The two data sets can be compared, but “organic search increased during this period” plus “there were AI citations in the same period” is not enough to prove that all the added clicks came from those citations.

The basis for this check is Google’s official documentation on AI search features and measurement. If a tool claims it can precisely break out the path from every AI impression to an inquiry, ask further what raw fields it relies on, which entry points it covers, and how it handles missing records; do not accept a column name inside the tool at face value.

For the reporting boundary between Search Console and AI Mode, you can continue reading the explanation of AI Mode data definitions. When an older article conflicts with current documentation, verify the feature against the current official description first, then judge whether your own data applies.

When the source is invisible, why can’t you fill in attribution yourself?

An unknown source means an evidence gap; it does not mean the visit naturally came from AI. A buyer may remember the domain and type it in again, or return through another channel; even if they saw an AI answer first, a given visit may not preserve that experience. Without the corresponding record, these are only possible explanations.

I suggest adding a short, voluntary question at the inquiry stage to capture the source, such as asking where they learned about the company. This is a way to collect the buyer’s own account, not to reconstruct the full browsing path. The sales record should say “buyer’s own account,” and it should not be mixed with website source fields as one kind of evidence.

If they clearly say they came from a specific AI platform, you can ask further whether they saw the company name in an answer or clicked through to a page. If they cannot say, keep the original account and do not lead them toward a more precise answer. What foreign-trade work needs is a trustworthy procurement judgment, not every lead pushed into a channel you hoped would succeed in advance.

Likewise, crawler visits cannot substitute for customer visits. Crawl activity in server logs only helps check visiting behavior; it cannot prove that a buyer read the page, sent a message, or closed a deal. For how to keep evidence of citation sources, see the tracking method for AI citation text and URLs.

How do you connect identifiable visits to real inquiries?

Define what counts as an inquiry first, then arrange the events. A submit button being clicked only shows that someone started an action; a saved record after successful form processing is the basis for further checks; and a salesperson actually receiving and confirming the content is the evidence after that. The measurement design should revolve around these actions, not treat a click event directly as real customer demand.

I suggest giving each inquiry a record identifier and linking it to approved source information that is actually recorded. Later checks use the same identifier to verify storage, notification, and the sales assessment. Do not pass customer email addresses, phone numbers, or inquiry text through public URLs or analytics events; business data belongs in the relevant management system.

The WhatsApp entry point must also be separated. An outbound click that the site can record does not automatically prove the message was sent, the recipient received it, or the person fits your supply terms. If the sales side does have a message record, add it; if not, keep the result at the click level. Test submissions should also be clearly marked and excluded from real lead counts.

The source field also needs a clear time meaning: does it record first awareness, the current session, or an identifiable visit before the inquiry? These records can coexist; you cannot switch definitions temporarily to make a channel look good. Google’s explanation of traffic source fields is in the official campaign and traffic source documentation.

Completing these actions still does not equal proving incrementality. Buyers may already know the company, or may be influenced by other content at the same time. I report “inquiries that can be linked under these rules” first, and discuss additional effect only once a suitable control design exists. For how to avoid misreporting visit performance as customer acquisition, see the diagnostic method for checking traffic and inquiries separately.

What can a fixed panel prove, and what is still missing?

The fixed GEO panel in the work order measures answer performance using dsh (DeepSeek) headless, and it cannot be extrapolated to other models. Strict mentions went from 3/72 to 6/72 and provider name mentions from 0 to 2, which shows that observations differed between those two rounds under the specified conditions; it did not record any real buyer’s click-to-inquiry path. The project research explicitly keeps the gap “citation-to-inquiry still has no metric.”

Panel data also contains sampling noise, so the change cannot be attributed entirely to content edits. Environment checks, false-positive corrections, and labeling conditions should all be kept alongside the results. My approach is to let answer records prove answers and inquiry records prove inquiries, not to substitute one metric for another.

A control design also cannot be completed by drawing a before-and-after arrow in a report. Advertising, seasonality, brand exposure, and on-site changes can all shift at the same time. Without addressing those explanations, you can at most say the two coincided in time; you cannot write it up as established causation.

How should the report for the boss be written?

The report should list visibility, identifiable visits, inquiry linkage, and sales feedback separately, and state the source and the unknown parts at each layer. If a layer has no data, write that it has not been recorded yet rather than covering it with a growth number from beside it. Readers need to know which results support further investment and which are still only observations.

I keep the source rules currently in use, the recording dates, and a change log. After a new channel classification is added, applicable data can be rechecked, but you cannot assume the historical records have all been fully rebuilt. Adjusting a report and improving the business are different tasks; once the former is done, the latter still needs to be observed.

For a reviewable format, see the evidence fields for an AI visibility report; the actual experiments and limitations can be found through the research records index. My acceptance standard is that every conclusion can be traced to a corresponding record, not that every channel cell must be filled with a result.

What other common questions come up about AI attribution?

Can GA4 identify AI assistant sources today?

The official default channel groups already list an AI Assistant category, but actual coverage depends on the identification rules and available source information. You should check your own channel, source, medium, and data conditions.

Do visits from Google AI Overviews belong to the AI Assistant channel?

The current official channel documentation classifies Google AI Overviews and AI Mode under organic search; Search Console includes the related performance under the Web search type. Total organic search cannot be treated as AI-only traffic.

If the brand is cited, does that count as an AI inquiry?

No. An answer citation, a site visit, and a real inquiry are different actions and must each be recorded. The fixed panel has not yet proven that citations bring real inquiries.

If a buyer says they learned about the company from AI, can that count as evidence?

It can serve as a buyer’s own account, labeled as such. It cannot replace the full visit path, nor can it alone prove incremental customer acquisition.