TL;DR
- Keep known AI referrals as a useful traffic group. They are not a full record of AI influence.
- A click with no source label and a later visit through search leave different gaps in the record.
- Panel research links AI advice with later visits. It does not show B2B sales gains or predict results for your site.
- Keep referrals, buyer source notes and saved AI answers distinct. Do not scale up known visits with a ratio borrowed from a study.
A Visit Label Is One Part of the Evidence
For a generative engine optimization (GEO) budget, “we have few AI referrals” and “AI has little effect on buyers” are different claims. The first may describe a channel report. The second needs evidence from before people reach the site. A large direct segment cannot fill that gap. A favorable AI answer cannot do so either.
A Missing Referrer and a Later Search Are Different Gaps
Patrick Stox’s May 2025 link tests at Ahrefs found that the source data passed by a link varied with the assistant, account and link type. Some links passed no referrer; others from the same service did. These are dated tests. They do not show how often each AI app loses source data today.
That gap can affect a click straight from an answer. The panel study concerns another path: a person may not click the answer at all. They may search later. The visit can have a valid search label, yet the site has no record of the earlier AI answer. Fixing a link label would not join those two events.
Google’s current GA4 guide says “(direct) / (none)” is traffic with no clear referral source. It lists several possible causes. Direct traffic is not a hidden-AI total. Some people type a URL; other visits lose source data for other reasons.
| Direct evidence | Actual result or test | What it cannot establish |
|---|---|---|
| Iannelli and Ai, 2026 | Seven-day navigation contrasts: search +4.3 pp; own site +2.4 pp; retailer page +1.0 pp | A causal sales effect, a B2B forecast or the missing share in your analytics |
| Ahrefs link tests, 2025 | Observed mixed referrer passing across products and link types | A current market-wide proportion of untracked AI visits |
| Google GA4 guidance | Direct/none has no clear referral source; several causes are possible | Which unknown visits were preceded by an AI answer |
Figure 2. Evidence comparison. The panel study tests later browsing patterns. The link checks test a technical handoff. Google defines a report label. These sources cover different parts of the problem. They do not repeat the same test or give a shared ratio for missing visits.
The panel design is observational, not a random trial. Its controls cannot fully rule out a change in the user’s intent. That change could drive both the chat and the visit. A retailer-page visit is not a sale either. The study gives reason to question a referral-only account. It does not prove revenue that was never measured.
Keep Three Records Before Changing the Business Claim
This example is fictional. An industrial software supplier has few AI referrals. A buyer’s sales note says an AI answer helped them form a shortlist. The site visit is logged as Google organic. The two records describe different parts of the buyer’s reported path.
| Evidence record | Completed fictional entry | Supported reporting choice |
|---|---|---|
| Site visit source | Google organic in the saved visit record | Retain the observed acquisition label |
| Buyer-reported context | Buyer says an AI answer helped form the shortlist | Record reported AI influence separately; it is self-report |
| AI answer evidence | Buyer’s saved answer and date are available; it names the supplier | Review what was actually said and linked |
| Proposed dashboard fix | Move the organic visit into AI referrals | Reject: this would replace one kind of evidence with another |
| Revised report note | Organic arrival; buyer-reported AI influence; saved answer reviewed | Preserve both signals without counting two visits or claiming causal credit |
The analyst has made the record more useful without inventing a new channel total. If the old answer is missing, record that gap. Do not try to recreate it with a new prompt. A fresh answer cannot show what the buyer saw before.
Use the same care in later reports. Keep the source rules and dates beside the data. Check whether a rise in referrals followed a change in tags or channel rules. Review qualified inquiries in the sales record apart from site visits.
A voluntary buyer-source question can add context if the right consent and process are in place. It remains a self-report. It is not a full or unbiased account of influence. You can start by reviewing records you already hold; this does not require a new form field or a change to analytics.
For the separate question of visit quality, see our AI referral conversion guide. A higher conversion rate and a fuller record of the buyer’s path are different findings.
AI Referral Attribution: Common Questions
Can we estimate missing AI visits by scaling up the known referrals?
Not from these studies. The gap can vary with the service, link type, device, buyer path and tracking setup. A ratio from another panel cannot tell you which unknown visits on your site came through AI.
Does direct traffic rising prove AI is working?
No. Direct/none has several possible causes. Check changes in tracking and campaigns, and review buyer evidence before assigning a cause. Two numbers rising at once is not a causal test.
Should we stop reporting AI referrals?
No. Keep them with a clear source rule and outcome definition. Add context where it exists, while keeping visits, reported influence, qualified inquiries and revenue distinct.
Does a saved AI mention prove it caused an inquiry?
It shows what that answer said under the saved conditions. A buyer’s account may link it to their research. Neither record alone shows what would have happened without the answer.
Bring the Reporting Gap You Need to Explain
If a small AI referral total is driving a large budget choice, bring the channel rules, available buyer-source notes and saved answers to SemShift’s AI visibility audit. We can discuss what each record supports and where a claim needs to be narrower or checked further.