In one online store's analytics, a payment provider's checkout page appeared as a "referral" credited with 13% of revenue over 28 days. It sent no customers. Shoppers left the store to pay, came back to the confirmation page, and the analytics tool treated their return as a new visit from the payment site, giving it the sale.
That is one small example of a larger problem. As search sends fewer clicks and AI sends visitors with no referrer, the default reports in most analytics tools describe a website less accurately every month.
When clicks fall, measure value per visit and per channel rather than total traffic, find the visitors your tools mislabel, and accept that some of a website's value now happens in answers you cannot see.
Why sessions stopped meaning much
Two forces push the session count away from reality. Search sends fewer clicks for informational questions, so traffic falls even when the business is doing well; we explain why in who is eating your clicks. Meanwhile automated traffic rises, so sessions can climb while nothing real happens.
On our own site, more than half of recent sessions were classified as Direct, and fewer than one in five of them showed any engagement. Read at face value, Direct was our biggest channel. In reality most of it was automated, and the visitors who mattered were in two smaller rows lower down.
Where AI traffic hides
Google Analytics 4 now groups visits from assistants such as ChatGPT into an "AI Assistant" channel when they carry a referrer. Many do not. Visits from apps, from copied links and from AI agents browsing on a person's behalf often arrive with no referrer and land in Direct. Every AI figure you see is therefore a floor, not a total.
Search Console does not help separate it either. It reports clicks and impressions from Google's AI features inside the normal web search totals. On our property, the dedicated generative-AI view showed exactly the same totals as the standard report for the same three months, so there is no way to isolate AI clicks there today.
Six numbers that still mean something
1. Conversion rate by channel. Not sessions, but the share of each channel's sessions that complete the action you care about. On the store, ChatGPT visitors completed a key event in about 24% of sessions over four months, against about 15% from Google organic, which says far more than either channel's traffic.
2. Revenue or leads per session, by channel. This is the number that justifies effort. On the same store, revenue per ChatGPT session was about 2.8 times Google's over the same period, although ChatGPT sent a fraction of the visits.
3. Engagement rate, not bounce rate. Engaged sessions separate people from noise. A channel with high sessions and very low engagement is usually automated, and should be filtered or ignored.
4. Branded search impressions. When people learn about you from an AI answer, many search your name afterwards. Rising impressions on brand queries in Search Console are one of the few visible traces of citations you cannot see.
5. Clicks on the queries that still click. Split Search Console queries into brand, buying, comparison and informational groups. Judge buying and brand queries on clicks; judge informational queries on impressions and position, and expect their clicks to fall.
6. Landing pages for AI visitors. Pair the AI channel with landing page in your analytics. On the store, only 16% of ChatGPT visitors landed on the homepage. Knowing which pages AI sends people to tells you which pages to improve first.
Three attribution traps to fix this week
Payment and login providers as referrals. Add your payment gateways, card authentication pages and single sign-on providers to the referral exclusion list in your analytics settings, or they will keep taking credit for sales, as in the example above.
Your own domains as referrals. Subdomains, booking tools and old domains that redirect can appear as referrers to your own site. Exclude them.
Untagged campaigns. Email, social posts and partner links without tracking parameters fall into Direct alongside AI agents and bots. Tag every link you control so that Direct contains only what you cannot explain.
Where this stops applying
No setup recovers the value of being quoted in an AI answer that nobody clicks. That influence is real, as anyone who has been told "ChatGPT recommended you" knows, but it shows up only indirectly: in brand searches, in direct visits and in what new customers say when you ask how they found you. Asking that question, in a form field or on a first call, is still the most reliable attribution tool there is. The concession is that it is a survey, not a dashboard, and it takes discipline to keep asking.
Questions to ask of your own reports
- Which channel has the highest revenue or leads per session, and is it the one you spend the most on?
- What share of your Direct traffic is engaged, and what is the rest?
- Does any payment, login or booking provider appear in your referral report?
- Do you ask new customers where they heard of you, and do you record the answer?
Fixing measurement is usually the first month of our digital marketing retainers, because every later decision depends on it. If you want to see how AI search is changing what visitors are worth, start with the role of your website in AI search.