Your AI visibility report shows more mentions. Google Analytics shows no traffic from those assistants. Before you announce that the campaign worked—or failed—check whether the two reports are measuring the same thing.
They usually aren't. A saved AI answer measures an observation in a chosen sample. GA4 measures activity collected on your website. Someone can read a recommendation without clicking, and a website visit doesn't tell you every source that influenced the buyer.
Start with a narrower question: which visits can we identify as AI referrals, and what did those visitors do? Here's a practical way to answer it without turning missing data into imaginary revenue.
Which GA4 report should you start with?
Open Traffic acquisition and use a session-scoped traffic-source dimension, such as Session source / medium. Navigation can vary with your property's report collections; look for the report name rather than assuming every account has the same menu.
Google distinguishes this from User acquisition, which uses first-user dimensions. The two reports answer different questions: a source associated with someone's first visit isn't necessarily the source associated with the session you're reviewing. Keep the report and dimension name in your notes. Google explains the difference between the acquisition reports.
Choose a complete date window and record it. For a small site, a 28-day view can be easier to discuss than a single afternoon. Compare equivalent windows, and note tracking or consent changes that make them less comparable.
Find the source values your property actually receives
Search the source rows for assistants you want to investigate. Start with values you can inspect, rather than assuming a fixed list catches every AI visit. Look at both source and medium before grouping them.
ChatGPT provides one documented example: OpenAI says it adds utm_source=chatgpt.com to referral URLs from ChatGPT search results. That gives you a useful value to look for; it isn't a promise that every AI-assisted journey will arrive with that parameter. Source: OpenAI's publisher FAQ.
Keep a small source register: the exact value observed, why you classify it as an assistant referral, the date checked, and any uncertainty. Don't automatically relabel all referral, direct, or unassigned traffic as AI traffic. If you build a custom grouping later, preserve the raw source values so you can audit your rule.
A useful sanity check is to follow a public link from the assistant experience you're evaluating, with your normal permitted analytics setup, and inspect the destination URL and collected attribution. Label this as a test and keep it out of your business-performance conclusions. A test visit shows that one path works; it doesn't validate every visitor's path.
Check landing pages and meaningful actions
For the identified sessions, inspect the pages people entered on and what they did next. Use a report comparison or an exploration if your standard report doesn't expose the dimensions you need. Keep the same source criteria and date window while moving between views.
| Question | What to inspect | What it doesn't establish |
|---|---|---|
| Did identifiable visitors arrive? | Sessions and their recorded source / medium | Every person influenced by an AI answer |
| Did they find a relevant page? | Landing page and engagement | That the visitor was a qualified buyer |
| Did they take the intended action? | A verified signup, lead or other selected key event | That every event became a sale |
| Did the business earn money? | Validated revenue or matched sales records, with a stated attribution method | Incremental revenue caused by your content edit |
Before using a key event in a report, check that it records the action you care about. A button click and a successfully submitted demo request are different events. Confirm your implementation instead of choosing whichever number looks strongest. Google's key-event guide covers setup and reporting.
Also distinguish event counts from rates. Imagine a fictional site with 12 identified assistant-referred sessions. One session produces two lead events because the form fires twice. There are two events, but only one session with a lead event: 1 divided by 12, about 8.3%. The duplicate needs investigation before anyone reports two leads. Google's Traffic acquisition documentation defines session key-event rate as the percentage of sessions with a key event.
What if GA4 shows zero AI referrals?
Write down the observation precisely: “We found no sessions matching these assistant-source values in this date window.” That is more defensible than “AI search has no effect on our business.”
Check the property, date range, source filter and whether website measurement is functioning as intended. Confirm the page actually received visits before diagnosing its attribution. Ask whoever maintains analytics to review consent handling and collection problems without bypassing visitor choices.
Missing source information is another limitation. Google describes (direct) / (none) as traffic without a clear referral source. It is not a hidden AI bucket. A buyer who discovers you in an assistant and later types your address may be relevant to your research, but you can't identify that journey from a direct-traffic row alone. Source: Google's direct-traffic guidance.
Use voluntary sales feedback as additional evidence if you have it. Keep self-reported discovery separate from measured referral sessions, and don't add the two totals without checking for overlap.
Don't manufacture attribution with internal tracking links
You don't need to label every link inside your blog as an AI referral. A click from one of your own pages to another isn't evidence of a new assistant visit, even if the first page sometimes gets cited.
Use campaign parameters to describe real campaigns you control. Preserve incoming parameters where appropriate and verify redirects don't accidentally discard them. Don't assume an assistant will copy a tagged internal link into a citation.
Google's campaign and traffic-source documentation explains that campaign values collected during an existing session can affect event-level attribution without changing that session's source. Avoid introducing misleading values just to fill a reporting gap.
Bring three separate lines to the review
A useful weekly summary can be short:
- Visibility: what happened in your defined sample of answers, with mentions, recommendations and citations separated.
- Measured visits: the sessions, landing pages and verified actions recorded for your identified assistant sources.
- Business evidence: qualified leads or sales you can substantiate, with the attribution method and gaps explained.
Our ChatGPT tracking guide explains how to keep the first line consistent. The share-of-voice guide explains why a mention total isn't a citation or revenue metric. WildSEO helps investigate tracked answers and sources alongside search and traffic evidence; keep those different measurements visible in your report.
Only call a figure ROI when you've defined the return, relevant costs, time period and attribution assumptions. Revenue credited by a reporting model isn't automatically incremental profit. You can report useful progress before claiming either.
Start by checking one source value, one relevant landing page and one verified action. If you can't explain those three, adding a larger visibility chart won't fix the report.
