GA4 shows visits from chatgpt.com, but the enquiry report is empty. It is tempting to conclude that AI visitors only read and never buy. That conclusion skips several checks: whether the source filter describes the right visits, whether a relevant service is reachable and whether a real enquiry would appear in the report at all.
Start with one landing article and follow a customer’s task to contact. We recommend checking delivery and measurement before rewriting the entire content programme. An empty event column and an empty sales pipeline are different observations.
Establish which visits you are actually examining
Use session-level acquisition to investigate the source of incoming visits. Google distinguishes First user, Session and event scopes. A user first acquired elsewhere may return from an AI platform; a First user report answers a different question from the source of that return session.
Inspect the observed source/medium values and landing pages instead of assuming every AI visit falls into one default channel. Retain the actual filter in the report. A referring domain or a campaign parameter can help identify a visit, but not every route preserves the same information. Missing referral data does not prove direct demand or absence of AI influence.
Keep ChatGPT referrals separate from Perplexity or other identifiable sources before combining them for a specific purpose. Google AI Overviews and AI Mode are not a chatgpt.com referral cohort. Google reports their search performance within Web search; do not infer a universal AI channel from a list of external domains.
Exclude your own diagnostic visits from the business interpretation, and label suspected automation rather than declaring every short visit a bot. Save the period, device, language and country filters. On a small site, one internal test or a few visits can materially change the apparent pattern.
Open the page as a potential customer would
A person may arrive halfway through research with a narrow question, not an intention to request an agency immediately. Read the article’s promise and answer. Then ask what problem a relevant buyer could reasonably want help with next.
The transition should name that task. A reader of a tracking explanation needs a suitable tracking service, not an unrelated homepage pitch. The service page must clarify what work is offered, what information is useful and what happens after contact. Check the actual locale and market, especially where an English answer links to a German business page.
On mobile, use the real article link and continue through the service and contact route. Look for an obscured button, horizontal scrolling, a broken anchor or a form that rejects a reasonable input. A working desktop screenshot does not establish that this route is usable on a phone.
A generic “contact us” link may work technically while leaving the commercial question unanswered. Improve the explanation where the next step is unclear. Do not force a sales request into every informational paragraph or remove a useful answer merely to make the article shorter.
Separate each stage and its evidence
Keep the stages below distinct even if your implementation uses different event names. They form a diagnostic map, not a requirement to install every possible tracking event.
| Stage | Evidence to inspect | What it does not establish |
|---|---|---|
| Observed incoming visit | Session source/medium and landing page under stated filters | Every AI-influenced visit or a unique potential buyer |
| Relevant service reached | Page view or a verified article-to-service interaction | Purchase intent or an enquiry |
| Contact attempted | Form start, valid button interaction or chosen contact method | Successful delivery |
| Enquiry received | Application acceptance and destination record, reconciled by an appropriate identifier | A qualified opportunity |
| Sales outcome | CRM status, response and agreed qualification | Causation by the last article viewed |
For the form, distinguish a browser click, application acceptance and downstream receipt. A success-looking page can coexist with a failed notification. Conversely, a delivered email or phone enquiry can exist without your expected browser event. Reconcile a sample with the receiving system before deciding which side is wrong.
Technical verification should use an authorised test destination or controlled test mode. Do not fill a production pipeline with fictitious customers to manufacture evidence. Preserve consent requirements and send only the data needed for the agreed measurement; personal message content does not belong in analytics parameters.
When zero events and a real contact coexist
Consider a hypothetical diagnostic, not a Canem Errant client result. A filtered group contains 80 observed sessions, 8 observed service transitions and 3 form starts. The submission event shows 0. Separately, the CRM contains 1 new enquiry whose source is unknown.
The CRM record does not prove an AI conversion. The zero in GA4 does not prove there were no enquiries. Check whether the submission event fires after genuine acceptance, whether another contact method was used and whether consent or filtering affected observation. Keep the unknown-source record unknown until evidence supports a link.
Do not divide that CRM record by the 80 sessions and announce an AI conversion rate. The numerator has not been assigned to the cohort. Even the observed stages need compatible counting rules: interaction counts, sessions containing an interaction and unique enquiries are different units.
This is the useful dead end to recognise early: adding more events will not recreate a missing historical identity or prove a source retrospectively. Repair measurement for future observation and record the historical limit. A voluntary “how did you find us?” answer can supplement the record, with its own recall limitations.
Decide whether the problem is measurement, the offer or volume
If delivery fails, fix and verify delivery first. If delivery works but the event definition is wrong, correct that definition. If the article attracts research visits but provides no relevant service transition, test a clearer task-specific explanation and destination.
If suitable enquiries arrive and then stall, inspect response time, ownership and sales qualification. The website cannot close a conversation nobody answers. If every stage works but the sample is small, extend observation without declaring the channel good or bad from a handful of visits.
Select one change and an evaluation period suited to the buying cycle. Compare compatible cohorts and record other changes to content, offers or traffic. A before-and-after improvement is a useful signal, but it does not isolate the effect when several conditions changed together.
Make the next step small enough to finish
Prepare a short worksheet with the article, relevant service, observed source filter, receiving destination and current evidence at each stage. Mark missing evidence explicitly. That is enough to assign the first correction and its acceptance check.
Canem Errant’s CRO service can inspect this route from a concrete customer task. Begin with one article and one contact method. Confirm that the customer can complete the journey and that the business receives the enquiry; evaluate source attribution and lead quality as separate questions.
Frequently Asked Questions
Does no key event mean AI traffic produced no leads?
No. The event may be absent, misdefined or unobserved, while a different contact route may have been used. Check actual receiving records before drawing the business conclusion.
Can I identify every visit influenced by ChatGPT?
No. Observable referral or campaign information describes a subset of identifiable routes. Copied links, later returns and missing signals limit attribution; do not label the unknown remainder as confirmed AI traffic.
Should I remove an article that gets visits but no enquiries?
First verify intent, the next step, delivery and the observation period. An informational article may support later decisions. Removal needs a reason grounded in its role, not just a zero in one report.
