On Indie Hackers recently, an app founder mentioned that ChatGPT had become their third-largest organic traffic source, behind Google and Bing. A few years ago that sentence would have sounded odd. Now it's common enough that people mention it in passing.
But my first question when I read it wasn't "how much traffic?" It was "do those people buy?"
A visit is cheap. A sign-up costs the visitor some effort. A purchase costs them money. If you're a small team deciding where to spend your next month, only the last two really matter. So this post is about measuring AI referrals by what they turn into, not by how many there are.
Your AI numbers are probably too low
Start with an uncomfortable fact: whatever your analytics says about AI traffic, the real number is likely higher.
Referrers get lost. When someone taps a link inside a mobile app, the in-app browser often doesn't pass a referrer along, so the visit lands in "Direct" with no source attached. Desktop web versions of assistants tend to behave better than their apps, but you can't count on either.
There has been some progress. On 1 October 2026, Roger Montti at Search Engine Journal reported that Google Gemini had started adding UTM parameters to outgoing links. Google hadn't documented it, so nobody outside Google knows exactly when the tags appear. The same article notes that ChatGPT already adds UTM parameters, but only to links in answers grounded in web sources, not to answers drawn from its training data.
So some AI visits arrive tagged, some arrive with a referrer, and some arrive looking like someone typed your URL from memory. If you only count the first two groups, you undercount.
Crawls are not customers
The other trap runs in the opposite direction. Server logs full of AI bots can make you think you're being "discovered" a lot.
On a Search Engine Journal podcast published 7 October 2026, Newfold Digital CEO Sachin Puri (Newfold owns Bluehost) shared numbers from a 90-day Bluehost analysis. Across Bluehost-hosted sites there were 83.4 million verified AI crawler requests per day. Roughly 90% were for model training, about 9% for search indexing and around 1% for live assistant fetching. For OpenAI specifically, Bluehost estimated one observable referral visit for every 5,350 verified HTML crawl requests.
Puri was clear that this is not a conversion rate. It's one host's sample, and it can't see the customer who reads an AI answer and then phones you or walks in. Still, the gap tells you something useful: bot activity and human visits are separate measurements. Track them separately.
A 30-day routine you can run in a spreadsheet
You don't need a new analytics stack for this. You need a habit and one extra form field.
1. Add a question. Put "How did you first hear about us?" on your sign-up form or checkout. Make it a free-text field, or a dropdown with an "Other" box. Include ChatGPT, Gemini, Perplexity and Claude as options. People will tell you things your analytics can't see.
2. Keep three kinds of rows. In one sheet, log events as they happen:
- AI visit: a session with an AI referrer or an AI UTM tag.
- Sign-up: a new account, trial or email subscriber.
- Purchase: money received.
Each row gets a date, a source (ChatGPT, Gemini and so on) and an ID or email when you have one, so you can link a purchase back to a sign-up.
3. Give every row one label.
- Confirmed: your tools show the AI source directly. A referrer, a UTM tag, or a sign-up session you can trace to one.
- Declared: the customer told you. They picked "ChatGPT" in the form, or mentioned it in an email.
- Assisted: AI played a part but wasn't the first touch. Someone found you on Google, then asked ChatGPT to compare you with a competitor, then bought.
Don't merge these. A confirmed sale and a declared sale are both real, but they're evidence of different strength, and you'll want to know the mix.
4. Review weekly, decide at day 30. Once a week, spend fifteen minutes filling gaps: match form answers to accounts, check "Direct" sign-ups for a declared source. At the end of the month, count sign-ups and purchases per source and per label. Compare AI sources with Google and Bing on the same terms.
You might find that a small number of AI visitors convert unusually well. They often arrive having already been told what you do. Or you might find they bounce. Either way, you'll have a number to argue with instead of a feeling.
What this won't tell you
This method has real limits, and it's better to know them up front.
Self-reported sources are fuzzy. People misremember, and they pick the first plausible option. A few dozen answers is not a statistically meaningful sample, so treat your month as a direction, not a verdict.
Some influence stays invisible. If an assistant recommends you and the person searches your name on Google a day later, your analytics will credit Google. The form field catches some of this. It won't catch all of it.
The tagging can change. Gemini's UTMs are undocumented, so they could change or disappear. Note the date tags first show up in your data, because any comparison across that date is comparing two different ways of measuring.
And 30 days is short. Seasonality, a launch or one viral post can swamp everything. Run it again next month before you move budget around.
I'll be running this on my own projects for the next few weeks. If you already track this: are your AI-referred visitors converting better or worse than your Google visitors, and how did you find out?
Disclosure: I made a kit for this. The AI Referral Ledger ($19) is the spreadsheet version of this routine, and The 5-Prompt Citation Snapshot is a free way to check whether AI assistants mention you at all.
Top comments (0)