AI search attribution asks a harder question than "How many visitors arrived from ChatGPT?" It asks whether an assistant helped somebody discover, evaluate or choose your business, including journeys where the final visit came through another channel.
The practical answer is to measure what you can identify, record what customers tell you, and keep inference separate. Do not turn an unobserved journey into a precise conversion claim.
Reviewed 8 September 2026.
What changed with NIQ and Similarweb?
On 2 September 2026, NIQ announced a measurement collaboration with Similarweb. The planned solution aims to connect AI product discovery with traffic and purchasing outcomes. An initial version is planned for Q4 2026, starting with selected categories and markets. It is not available today according to the announcement's FAQ.
Our interpretation is that measurement providers see a commercial need beyond counting assistant referrals. The announcement itself does not prove that every AI-influenced purchase can already be attributed accurately, or that your business is covered by the planned launch.
Direct referral and assisted discovery are different
Consider an illustrative customer journey. Someone asks an assistant to compare booking software, remembers a product name, and later searches for it before signing up. A website visit alone may not reveal the earlier conversation.
For reporting, we suggest three evidence classes. The first is an identifiable visit associated with an AI source. The second is a customer's self-reported AI discovery. The third is an analytical hypothesis supported by broader patterns, such as a change in observed recommendations alongside branded demand.
Keep the original evidence visible in each class. Do not give all three the same "verified" badge. They have different coverage, failure modes and uses.
Start with a clean GA4 traffic view
Google's Traffic acquisition documentation explains session-level source and medium reporting alongside engagement and key events. Begin there to inspect identifiable traffic, keeping the scope consistent rather than mixing first-user acquisition with session acquisition.
OpenAI's publisher guidance says ChatGPT referral URLs include utm_source=chatgpt.com. Use observed source values and test your own measurement setup; do not assume that every assistant or every subsequent visit will carry an equivalent identifier.
Our suggested implementation check is to follow a known link through your own site and verify what reaches analytics, including any redirects. Use a non-production test where appropriate and respect your normal consent setup. Document what the test can and cannot establish.
Then compare useful outcomes, not only sessions. For a service business, a relevant event could be a qualified enquiry. For software, it could be a completed activation rather than an account that never uses the product. Define the event before evaluating a channel.
Add a customer-reported discovery field
We recommend a short, optional question at a suitable point in the journey: "How did you first hear about us?" Give respondents a way to specify an assistant and to add context without forcing that answer.
Keep the exact response and collection date where appropriate. A customer's recollection is valuable, but it is not the same as an independently observed click path. Do not overwrite the analytics source with the survey answer; retain both.
For longer sales cycles, distinguish initial discovery from later research. A buyer might first encounter your brand at an event and use an assistant to compare it weeks later. Both can matter without either deserving all the credit.
Avoid double-counting the same customer
Here is an illustrative example. Your records show 12 customers associated with identifiable AI-referred visits. Eight customers also say an assistant helped them discover you. Five people appear in both groups.
The deduplicated union is 15 people, not 20: twelve plus eight minus five. Even then, "15 people with one or both AI signals" is a more defensible label than "15 customers caused by AI".
If you cannot reliably identify the overlap, display the two totals separately rather than inventing a combined figure. Keep your existing privacy and data-governance boundaries intact. Measurement should not depend on collecting information you have no appropriate reason to retain.
Use visibility as context, not a conversion substitute
An improvement in prompt-test visibility can suggest a useful hypothesis. It does not tell you how many real buyers saw those answers, clicked, or purchased.
Likewise, more branded searches could have several explanations. A product launch, advertising campaign, event or recommendation programme may overlap the same period. Put those events in your analysis notes before assigning credit.
We recommend reporting confidence explicitly: observed, self-reported or inferred. For high-value decisions, consider whether a carefully designed controlled test is possible. A dashboard correlation alone should not be presented as causal lift.
Build a monthly decision report
Our proposed report has three sections: identifiable AI-associated visits and outcomes; customer-reported discovery; and visibility observations with known limitations. Finish with the decision those signals support, such as improving a comparison page or testing a clearer product explanation.
Use UpSearch's SEO Dashboard as part of your wider performance review and AI Mention Tracker for the observation layer. This framework does not assume that UpSearch can reconstruct private assistant conversations or supply complete cross-channel attribution.
For exposure inside Google, read the Generative AI reporting guide. Keep that evidence separate from the visit and customer records used here.
The objective is not to give AI the most credit. It is to make a better investment decision with the evidence you actually have.
FAQ
Can I calculate AI search ROI from referral traffic alone?
You can evaluate the outcomes associated with identifiable visits, but that is narrower than all possible AI influence. Label the scope rather than presenting it as a complete acquisition picture.
Does a rise in branded searches prove AI-assisted growth?
No. Treat it as a hypothesis to investigate alongside campaigns, customer feedback and other evidence. It does not reveal the cause by itself.
Is the NIQ and Similarweb solution already available?
Not according to the announcement reviewed for this article. NIQ says it is being built, with an initial version planned for Q4 2026 and a limited initial market and category scope.