GEO referral traffic measurement should not stop at the referral row, because AI answers shape shortlists before anyone clicks. Two published datasets put AI referrals below 1% of Google's volume, yet the visitors who do arrive convert far better. Measure four layers instead: referral tags, branded search lift, assisted conversions and AI share of voice.
Key takeaways
- Low AI referral traffic is a fact, not a verdict: it measures clicks, and AI answers often end the research without one.
- Report four layers side by side: tagged referrals, branded search lift, assisted or self-reported conversions, and share of voice on a fixed prompt set.
- Fix GA4 first: ChatGPT traffic can be split across three channels, so the AI Assistant channel alone undercounts.
- Add an open-text "How did you first hear about us?" field to every demo and signup form this week.
- Credit AI with a fraction of a conversion when evidence supports it. Claiming whole deals is as wrong as claiming none.
What happened: the "GEO is a distraction" argument and the reply
On 23 October 2026, Marie Haynes published GA4 data from 100 websites and over 145 million sessions. In Search Engine Journal, October 2026, she reported that ChatGPT sessions equalled 0.62% of Google organic volume, rising to 0.73% with all other LLMs. Across the 55 sites tracking conversions, ChatGPT visitors converted at 13.8% against 3.02% for Google organic, yet Google still drove 96.5% of conversions.
On 5 November 2026, Dan Taylor replied with his own follow-up. In Search Engine Journal, November 2026, he reported 108 million sessions from 1 January to 2 November 2026 across 57 websites, where trackable AI traffic was 0.66% of the total: about seven AI users for every 1,000 from search engines, and 20 to 25 per 1,000 on some sites. He accepts the numbers and disputes the conclusion, writing that "the strategic conclusion rests on assumptions" that do not match how people use AI tools.
Both writers are right about what they measured. The disagreement is about what a referral row can show, and that is a measurement question you can answer with your own data.
Is GEO worth it if AI referral traffic is low?
Low referral traffic alone does not decide whether GEO is worth it. It is worth it if AI answers influence your buyers' shortlists and you can show that influence through other signals. It is not worth it if your buyers do not use AI tools for research, and your own data should settle that, not an industry average.
Taylor describes the pattern: a buyer compares vendors inside a chat for half an hour, then days later searches your brand name and books a demo. Analytics credits organic search. The AI conversation never appears. SEO Sherpa describes the same last-click problem and warns about the opposite error, giving AI full credit for every conversion that followed a citation (SEO Sherpa, October 2026).
Two further facts matter. Haynes reports that ChatGPT referral growth flattened after August 2025, so waiting for referrals to rise may not work. And the click-through that does happen is a small, high-intent slice, not the whole influence.
How do I measure the influence of AI search on sales?
Measure influence with four layers, because no single one is reliable alone. Each answers a different question, and agreement between layers is the evidence. The table shows what to pull, where it lives and how it misleads.
| Layer | Question it answers | Where to get it | How it misleads |
|---|---|---|---|
| 1. Referral tagging | Who clicked from an AI answer? | GA4 source/medium, server logs | Misses zero-click journeys and in-app browsers that strip referrers |
| 2. Branded search lift | Are more people searching for us by name? | Search Console branded queries filter | PR, ads, launches and seasonality also move it |
| 3. Assisted and self-reported conversions | Did buyers say AI introduced us? | Form field, onboarding question, sales call notes | Self-reports are incomplete and recall is imperfect |
| 4. AI share of voice | How often are we named against rivals? | Fixed prompt set run on each engine | Answers vary, so single runs mislead |
Layer 1: fix referral tagging before you read it
OpenAI's help centre says ChatGPT includes utm_source=chatgpt.com in referral URLs for publishers who allow OAI-SearchBot, so those visits can be tracked in analytics (OpenAI Help Centre). GA4 added an AI Assistant channel to its default channel group on 13 May 2026. Itamar Blauer found chatgpt.com sessions split across AI Assistant, Referral and Unassigned, because in-app browsers can strip the referrer (Search Engine Journal, July 2026).
- In GA4, add Session source/medium as a dimension and filter for chatgpt.com, perplexity, gemini and claude.
- Build one custom channel that groups every AI source across all three channels.
- Start trend lines from your rollout date, not earlier, since tagged and untagged traffic are not comparable.
Referrer behaviour changes between engines and apps. Our guide to Gemini UTM parameters and referral tracking shows how to check what each engine actually sends.
Layer 2: branded search lift
Google's Search Console branded queries filter, announced in November 2025, splits performance data into branded and non-branded queries (Search Engine Land, November 2025). Export three to six months of branded impressions and clicks before you start. Then compare weekly branded impressions against your AI mention rate, and annotate every launch, campaign and press hit.
SEO Sherpa gives an illustrative calculation: branded impressions rising from 12,000 to 17,400 is a 45% increase, (17,400 − 12,000) ÷ 12,000. That figure proves nothing about cause. Repeated, logical timing does: a lift that follows gains on specific prompts, appears for the product the answers named, and lasts beyond a one-week spike.
Layer 3: assisted and self-reported conversions
David Khim lists self-reported attribution among the metrics to track, and recommends testing whether a new form question actually hurts conversion (Search Engine Journal, August 2026). Use an open-text field, because a dropdown without an AI option pushes people to answer "Google". Tag the answers "AI assistant" or "AI named a competitor" and count them in CRM next to the last-click source.
For assisted conversions, build a GA4 or CRM segment of converters who also had an AI referral session earlier in the path, then report them as assisted, never as last-click wins.
Layer 4: AI share of voice
Share of voice is the one layer that sees the zero-click conversation itself. Run a fixed set of unbranded buyer prompts on each engine, repeat each run, and count the share of answers that name you against named rivals. Our guide to measuring AI share of voice across engines covers sample sizes, and AI citation stability explains why one run is noise.
What the published numbers look like as a reporting example
The arithmetic below uses Haynes's published rates only. It shows why last-click makes AI look negligible while telling you nothing about influence.
| Per 10,000 Google organic sessions | Sessions | Conversion rate | Conversions |
|---|---|---|---|
| Google organic | 10,000 | 3.02% | 302 |
| ChatGPT (0.62% of that volume) | 62 | 13.8% | about 8.6 |
On last-click, ChatGPT earns under 3% of the Google organic conversions. That row cannot include the buyer who read your name in an answer and later typed it into Google. Those conversions sit in the Google row, and only layers 2 to 4 can find them.
Vendor figures claim the gap is large: SegmentStream says across its customers, AI search's contribution is five to eight times higher than last click suggests, and up to 15 times for B2B and high-value products. SEO Sherpa flags that this comes from an attribution vendor's own customer data. Treat it as direction, not a benchmark to paste into a board deck.
What to check this week
- Confirm OAI-SearchBot and other AI crawlers can reach your pages with the free AI crawler checker.
- Run the free Noma AI SEO checker on your top commercial pages. It shows whether ChatGPT, Claude, Perplexity and Google AI Mode can read each page, with the top fixes.
- Rebuild your GA4 AI grouping across AI Assistant, Referral and Unassigned.
- Export 90 days of branded impressions from Search Console.
- Add the open-text "first heard about us" question to demo and signup forms.
- Write 20 unbranded buyer prompts and record who each engine names. See our guide on which prompts to track.
How to measure the effect over a quarter
Set a 90-day window with three pass criteria agreed beforehand: share of voice on the fixed prompt set moves by more than run-to-run noise; branded impressions rise in the weeks after; and self-reported AI mentions in the CRM grow. If all three move, you have a defensible case. If only share of voice moves, GEO is working on visibility but not yet on demand. If none moves, change the pages or the prompt set before spending more. The AI Visibility Business Case gives a pilot structure with go/no-go criteria.
Search Console will not close the gap for you. See why in our note on Search Console's AI report and clicks, and the wider method in the AI search attribution guide.
Where Noma fits
Noma by Nurdd covers layers 1 and 4 of this plan. It asks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, your buyers' questions every day, and reports visibility, position and sentiment per engine, competitor share of voice and the sources cited. It also reports which assistants send visitors back and which AI crawlers fetch each page. Noma also connects to Google Search Console, where the branded queries filter lives.
See it on your own brand: book a demo or start the 3-day free trial. A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged.
How we researched this
We read the two Search Engine Journal articles at the centre of the debate, Search Engine Journal pieces on GA4 and measurement, OpenAI's publisher documentation, a Search Engine Land report on Google's branded queries filter and a published ROI guide. Facts and figures were checked on 2026-11-06. We did not run our own experiment, and every dataset cited belongs to its named author.
What to do next
Pick one money page, and run all four layers on it for 90 days. If you want the share-of-voice and referral layers running without building spreadsheets, book a demo or start the free trial.
See what AI engines say about your brand
Free · no account · or start a 3-day trial
Questions people ask
Is GEO worth it if AI referral traffic is low?
It can be. Referral traffic counts clicks, and AI answers often finish a buyer's research without one. Haynes found ChatGPT sessions at 0.62% of Google organic volume, while Taylor argues influence shows up later as branded search and demo requests. Test it on your own data with branded search lift, self-reported attribution and share of voice.
How do I measure the influence of AI search on sales?
Use four layers: tagged AI referrals in GA4, branded search lift in Search Console, assisted and self-reported conversions, and share of voice on a fixed prompt set. Look for agreement between layers over 90 days. Credit AI with a fraction of conversions where evidence supports it, not whole deals.
How much traffic does ChatGPT send compared with Google?
Marie Haynes reported ChatGPT sessions at 0.62% of Google organic volume across 100 sites and 145 million sessions. Dan Taylor reported trackable AI traffic at 0.66% across 57 sites and 108 million sessions, with some sites at 20 to 25 AI users per 1,000. Your own site may differ, so check GA4.
Do AI visitors convert better than Google visitors?
In Haynes's 55 sites with conversion tracking, ChatGPT visitors converted at 13.8% against 3.02% for Google organic. Google still drove 96.5% of total conversions because of its volume. The rates describe visitors who clicked, not everyone AI influenced.
Why does GA4 undercount AI traffic?
GA4 picks a channel using source and medium together. Itamar Blauer found chatgpt.com sessions split across AI Assistant, Referral and Unassigned, because in-app browsers can strip the referrer. Group all AI sources into one custom channel and start trends from your rollout date.
How do I show branded search lift from AI visibility?
Export three to six months of branded impressions from Search Console's branded queries filter as a baseline. Compare weekly branded impressions with your AI mention rate and annotate launches, ads and press. Treat repeated, logical timing as evidence, not proof.