If a customer tells you 'I found you on ChatGPT' and you have no way to measure it, at the first budget cut that line item disappears — because you can't defend it with numbers. You're investing in a channel without being able to prove what it produces, and that makes it the first candidate for the chopping block. There are simple systems to track how many customers actually come from AI, and once they're in place they completely change the conversation about budget.
A customer reaches out saying “I read about you on Perplexity”. It’s the anecdote you hear more and more often. But how many leads arrive this way? And how much are they worth compared to the others?
As long as the answer is “dunno, seems to happen pretty often”, at the first budget cut whoever holds the purse strings will ask you to trim the “AI visibility” line because they don’t know how much revenue it brings in. And that’s your problem, not theirs. Let me walk you through the minimal attribution model you can put in place this week, with no new CRM and no consultants, to turn visibility in AI answers from a vanity metric into a number the sales team is happy to look at.
Why AI attribution doesn’t work like Google’s
When someone lands on your site from a Google search, in Search Console you see the query, the landing page, the impression, the click. The referral is clean. But when ChatGPT cites your site in an answer, three things happen:
- The user reads the answer inside ChatGPT and doesn’t click: you’ve generated invisible awareness.
- The user clicks the link in the sources panel: you get a referral, but the HTTP header often doesn’t contain `chat.openai.com`, depending on how the user opened the link.
- The user memorizes your brand, closes ChatGPT, opens Google and searches for “[Brand Name]
+ city”: it reaches you as brand traffic on Google, but the real origin is the AI answer from three days earlier.
It follows that, if you rely only on Google Analytics, a large slice of AI visibility vanishes from your numbers. It’s invisible not because it doesn’t work, but because the measurement tool isn’t calibrated for that source. In the world of classic SEO search we accepted “direct traffic” as a black box for years. With AI that black box becomes the main item, and pretending it doesn’t exist means telling whoever keeps the P&L in your company an incomplete story.
The attribution principle that works here
When an acquisition channel is opaque to technical tracking, the only reliable signal is the one declared by the user. It’s a principle as old as direct marketing: before Analytics existed, companies put “how did you hear about us?” on paper coupons. It follows that for visibility in AI answers you need two overlapping layers of tracking:
- Technical tracking: referral header, UTM parameters if the source is clean, direct traffic to deep non-indexed pages (a typical signal of someone arriving from an AI citation).
- Declared tracking: a mandatory question in contact forms and a question in the sales rep’s first messages.
In practical terms: the first layer tells you what happened technically, the second tells you what happened in the customer’s head. The sum of the two, however imperfect, is infinitely more useful than Analytics alone.
If the form has “Internet” or “Web” as a single option, you’ve destroyed the data.
The minimal setup to do this week
Let me tell you what I did with a new client before writing this article, so you get the operational checklist and not just the theory.
Step 1 — Add a question to the contact form. Required field, label “How did you hear about us?”, options: Google, ChatGPT, Perplexity, Gemini, Word of mouth, LinkedIn, Other. Three minutes of work in Forminator, Gravity Forms or whatever you use. The “AI/ChatGPT/Perplexity” option must be separate from Google: if you lump it into “online search” you lose all the signal.
Step 2 — Add the same question to the sales brief. When the sales rep calls a lead back, the first question after “hello” is “may I ask how you found us?”. The answer logged in the CRM in a dedicated field, not in free-form notes.
Step 3 — Set up a segment in Google Search Console. Open Google Search Console, filter your client’s brand queries and look at the trend over the last 6 months. If brand queries grow without you having run any ADV or PR campaigns, the growth is likely coming from AI mentions: the user saw the brand in an answer and then went to search for it. An indirect indicator, not proof, but useful as triangulation.
Step 4 — Compare with direct traffic to deep pages. Still in Analytics, isolate direct entries on URLs that aren’t the homepage and aren’t pages indexed for generic queries. If the long tail of deep pages with direct access is growing, someone is sharing them with you: either links in chats, or AI citations with clicks on the sources panel.
ChatGPT, Perplexity, Google and LinkedIn are different worlds: they need to be kept separate.
The case I followed: a Prosecco Superiore DOCG winery in Valdobbiadene
For eight months I’ve been working with a winery in Valdobbiadene (TV) that produces Prosecco Superiore DOCG, premium tier, direct sales on site + e-commerce + HoReCa. When we started, the answer to the question “how many leads do you get from AI?” was “we don’t know, some people tell us”. We applied exactly the four steps above. Six months of collected data, 312 total leads in the period (tasting requests + HoReCa contacts + e-commerce orders with the field filled in).
The result: 47 leads out of 312 declared they had discovered the brand through ChatGPT, Perplexity or “an AI answer”. That’s 15% of the total. Three figures struck me:
- The AI leads had a higher average order value than generic Google leads (full tastings instead of single samples, e-commerce orders with more expensive aging bottles).
- The time from first contact to close was shorter: they arrived already informed about the denomination, the rive, the vineyards. The AI answer had done the pre-sale.
- 70% of those leads would never have been attributed to AI without the question in the form: in Analytics they showed up as direct traffic or brand search on Google.
Let me state the limits: single sample, 6 months, a niche sector with seasonality (a spike in November-December for gifts). It’s not a study, it’s an observation on one client. But the pattern is clear enough to justify the setup investment, which costs nothing.
The mistakes I see most often
When I bring this up with clients, these are the four mistakes I run into nine times out of ten.
- Making the “how did you hear about us?” question optional. If it’s optional, 70% of users skip it. It needs to be made mandatory, even at the cost of losing a few conversions (in reality you lose nothing, I tested it).
- Having “Internet” as the only digital choice. If the form has “Internet” or “Web” as a single option, you’ve destroyed the data. ChatGPT, Perplexity, Google and LinkedIn are different worlds: they need to be kept separate.
- A sales rep who doesn’t ask. The form data only covers inbound leads from the site. Direct phone calls, word of mouth, WhatsApp requests don’t go through the form. Without the question asked by the sales rep, you lose half the sample.
- No dedicated field in the CRM. If the answer ends up in free-form notes (“says they saw us on ChatGPT”), six months later you can’t build a report. You need a structured field with predefined values.
What to do concretely this week?
You don’t need new software. You need one hour of work spread out like this:
- 10 minutes to add the mandatory “How did you hear about us?” field to the contact form, with the AI options separate from Google.
- 15 minutes to add the same field to the CRM and train the sales rep to ask it at the opening.
- 20 minutes to create a simple monthly report in Excel: total leads, declared AI leads, average value per source.
- 15 minutes to check in Google Search Console the trend of brand queries over the last 6 months as triangulation.
After 90 days you have a first data point of your own, after 6 months you have a solid baseline to bring to the table when budget is being discussed. It’s the bare minimum, not real analysis: for that you need professional marketing mix modeling tools and server-side tracking, but those are topics to tackle once you’ve already proven the channel produces.
