A customer found you through ChatGPT, but your analytics missed it
Published: · Reviewed by the geo-rank.ai editorial team · Our methodology
A buyer sees your company in ChatGPT, remembers its name and searches for your website in Google days later. Website analytics may show an organic search visit, while an interview identifies AI as the discovery source. Both describe real but different parts of the journey. To investigate ChatGPT’s contribution, combine visit data with the buyer’s account and preserve the origin of each piece of evidence.
This article concerns discovery before the first observable visit. A report on AI referrals measures recorded traffic; it cannot recover a click that never happened.
How a recommendation disappears from acquisition reports
Consider a fictional example. On Monday, a marketer researching website redevelopment discusses selection criteria with ChatGPT and saves two company names. On Wednesday, they search Google for one name, open the company website and submit a form.
Scroll horizontally to see all columns.
| Stage | Where evidence remains | What can be established |
|---|---|---|
| Receives a ChatGPT recommendation | The buyer’s conversation history | What a particular saved answer showed |
| Records a company name | A note or the buyer’s account | How the buyer explains the choice, subject to available evidence |
| Finds the website through a branded Google search | The search visit and available analytics | The observable route to the website |
| Submits an inquiry | Form records and CRM | The inquiry and its subsequent outcome |
The website did not participate before Wednesday. Its analytics has no ChatGPT referral to recover by changing the attribution model.
Google distinguishes first-user, session and event traffic-source dimensions. Check the scope before comparing reports: First user source and Session source answer different questions. GA4 documentation.
Why Direct cannot all be credited to AI
Optimist’s article on AI and SEO discusses discovery that is not visible in conventional click reports. That is a useful problem to investigate. However, growth in direct traffic or branded search does not identify its cause.
Google defines (direct) / (none) as traffic without a clear referral source. Possible explanations include entering the address directly or losing acquisition information. It is not a built-in label for “visited after an AI recommendation.” GA4’s Direct explanation.
Branded searches can also rise after advertising, a conference appearance, press coverage or a personal recommendation. A trend prompts investigation; it does not attribute demand to ChatGPT.
Ask about the sequence, not just the channel
Use a neutral opening question on a form or in the first conversation: “How did you first hear about our company?” Let the customer answer freely. “Did you find us in ChatGPT?” suggests the explanation you want to hear.
If the buyer mentions AI, ask:
- What task were they working on, and roughly what did they ask?
- Did the service introduce the company, or did they already know it and request a comparison?
- What interested them: the recommendation, approach, price or linked article?
- How did they subsequently find the website, and what prompted the inquiry?
This distinguishes discovery from later evaluation. Someone may hear about you from a colleague and use ChatGPT to prepare supplier questions. AI contributed to evaluation but was not the first source of awareness.
Where practical, invite the customer to share the relevant answer fragment without unrelated information. Do not demand their conversation history for a report. If the answer is unavailable, record their account and that limitation.
Preserve three layers of evidence
Keep automatically captured data, the customer’s account and the researcher’s interpretation in separate fields. Otherwise, a hypothesis can quietly become a supposed CRM fact.
Scroll horizontally to see all columns.
| Layer | Example record | Use |
|---|---|---|
| Observed data | Inquiry L-104; linked session source: google / organic | Reporting registered acquisition |
| Customer account | “I first saw the name in ChatGPT, then searched for the website” | Self-reported discovery through AI |
| Support and interpretation | Original answer unavailable; AI discovery reported by customer | Assessing evidence without inventing a referral |
L-104 is an illustrative identifier. Use your agreed inquiry or deal ID so that several calls with one buyer do not become several buyers.
If analytics cannot be reliably linked to an inquiry, keep it in the aggregate report. Matching the date and company can be a clue but does not necessarily establish ownership of a session.
Store interview notes in the system already used for customer records. Do not put free-text interviews in URLs or analytics event parameters: they can contain personal and commercial details unnecessary for channel measurement.
Classify the strength of the evidence
The customer shares an answer naming the company. This confirms the answer’s content. You still need to establish whether they saw it before knowing the brand and what role it played. A screenshot does not prove an additional sale caused by AI.
A detailed account exists without the answer. Record self-reported discovery or AI involvement. This remains useful evidence with a clear limit; it does not need to become a precise referral URL.
Only Direct or branded-search growth is available. This is an aggregate signal. Compare advertising, publicity, seasonality and analytics changes. Do not assign it to individual customers.
The customer does not remember. Keep “unknown.” Selecting a plausible channel improves the appearance of the report, not its accuracy.
Count each buyer once
Suppose a month brings 12 unique qualified inquiries. These are illustrative data. Eight buyers answer the discovery question. Three name ChatGPT: one supplies the original answer and two recall the journey. Another used AI only after a colleague’s recommendation. Four provide no answer to the question.
A valid report states: “Of eight respondents, three said they first discovered the company through ChatGPT. One supplied the original answer. Another used AI when evaluating the company. Discovery information is unavailable for four of the 12 inquiries.”
Self-reported ChatGPT discovery is 3 / 8 = 37.5% of respondents. Across all qualified inquiries, 3 / 12 = 25% carry that evidence. The four missing responses do not justify labeling the remaining 75% unrelated to AI.
Automatically recorded AI sources may overlap with survey answers. Merge by identifier before counting the combined group. Otherwise, a buyer who clicks a link and tells a salesperson about it appears twice.
Use the findings to choose an action
If several buyers independently describe the same useful scenario, check how clearly your website addresses it. Perhaps they found a supplier by asking about migrating a website without losing inquiries, yet the service page never explains that stage. That is a concrete content and offer question.
If customer accounts conflict with current service terms, correct the information shaping expectations. If you have only branded-demand growth, keep observing and account for other marketing activity.
Start with a few recent inquiries, using the same procedure. Preserve the customer’s words, the observed path and missing information. This complements traffic and inquiry analysis and helps define questions for AI visibility measurement.