Add a "How did you hear about us?" field to your high-intent forms, with AI assistants named as an explicit option, because a large share of AI influence never shows up in your analytics. A buyer who asks ChatGPT for a shortlist and then searches your brand name looks like branded search; one who types your URL looks like direct traffic. Asking is the most direct evidence you can get. This post covers why analytics misses AI influence, where to ask, how to design the field, how to code answers and how to report them honestly.
Why analytics misses AI influence
Referral tracking only sees the last click, and AI research often ends without one. Common paths that leave no AI fingerprint:
- The assistant names you; the buyer later searches your brand on Google
- The buyer types your URL from memory or from the answer
- The link is opened from a mobile app or copied into another browser, losing the referrer
- A colleague shares the AI answer in Slack, and the click comes from there
- The answer appears in Google's AI Overviews, and the click is counted as ordinary Google organic traffic
Tagging helps but doesn't close the gap. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral links, but testing by Seer Interactive found the tag isn't present on every link. Referral analytics remains useful as a floor; see tracking AI referral traffic in GA4. For the bigger picture of why so much AI influence is invisible, see zero-click answers.
Where to ask
Ask at the moments when buyers have just decided to engage and still remember how they found you. In priority order:
- Demo and contact forms, where answers are tied to pipeline
- Free trial or account sign-up, either on the form or as the first onboarding question
- Post-purchase surveys for ecommerce, on the order confirmation page or in a follow-up email
- Sales discovery calls, as a standard question logged in the CRM
- Event and webinar registrations, for B2B brands with heavy event programs
Sales conversations catch the richest detail ("I asked Perplexity to compare three tools and you came up twice"), so add a CRM field for it even if the form field exists.
Designing the field
A hybrid design gets you clean categories and the detail behind them. Use a short dropdown with AI assistants named explicitly, followed by an optional text field.
<label for="heard">How did you hear about us?</label>
<select id="heard" name="heard_about_us">
<option value="">Select one</option>
<option value="ai_assistant">AI assistant (ChatGPT, Gemini, Perplexity, Claude, Copilot...)</option>
<option value="search_engine">Search engine (Google, Bing...)</option>
<option value="colleague">Colleague or friend</option>
<option value="social">Social media</option>
<option value="review_site">Review site</option>
<option value="community">Community or forum (Reddit, Slack groups...)</option>
<option value="podcast_event">Podcast, event or webinar</option>
<option value="other">Other</option>
</select>
<label for="heard_detail">Which one, or any detail? (optional)</label>
<input id="heard_detail" name="heard_about_us_detail" type="text" maxlength="200">
Design choices that improve the data:
- Name the assistants in the option label. "AI assistant" alone is abstract; people recognize "ChatGPT."
- Keep the list under ten options, so people read it instead of picking the first one.
- Randomize or rotate order of the middle options if your form tool supports it, to reduce first-option bias (keep "Other" last).
- Separate search engines from AI assistants, and accept that the line is blurry: someone who read an AI Overview may pick "Google." The text field helps catch this.
- Allow multiple selections on long B2B journeys if your form tool supports it; many buyers will have used both an assistant and a review site.
- Store the raw text alongside the coded value, and pass both to your CRM.
Coding open-text answers
Free text needs a consistent taxonomy before it can be reported. Start with simple keyword rules, then review what's left by hand each month.
function codeHeardAbout(text) {
const t = (text || "").toLowerCase();
if (/chat\s?gpt|openai|\bgpt\b/.test(t)) return "ai_chatgpt";
if (/perplexity/.test(t)) return "ai_perplexity";
if (/gemini|bard/.test(t)) return "ai_gemini";
if (/claude|anthropic/.test(t)) return "ai_claude";
if (/copilot/.test(t)) return "ai_copilot";
if (/ai overview|ai mode|google ai/.test(t)) return "ai_google";
if (/\bai\b|assistant|chatbot|llm/.test(t)) return "ai_unspecified";
if (/google|bing|search/.test(t)) return "search_engine";
if (/reddit|forum|slack|discord|community/.test(t)) return "community";
if (/g2|capterra|trustpilot|review/.test(t)) return "review_site";
return "uncoded";
}
The order of the rules matters: AI-specific terms are checked before generic "google" or "search," so "Google AI Mode" isn't coded as plain search. Review every "uncoded" answer monthly and add rules for new patterns.
Asking in sales conversations
Sales discovery is where you learn how AI was used, not just whether it was. A short script:
- "How did you first come across us?"
- If they mention AI: "Which assistant was it, and what did you ask?"
- "Who else did it suggest?"
- "Did you check anything else before reaching out, such as reviews or our site?"
The second and third answers are gold for your prompt set and competitor tracking. Log them in a dedicated CRM field rather than in call notes, so they can be reported. Feed the actual questions buyers mention into your prompt research.
Reporting it honestly
Self-reported data is directional evidence, so report it alongside other signals and as a trend. A useful monthly view:
| Metric | What it shows |
|---|---|
| Share of responses coded to any AI assistant | Overall AI influence among people who answered |
| Breakdown by assistant | Where to focus platform efforts |
| Pipeline and revenue from AI-coded leads | Business value of the channel |
| Response rate for the field | How much of the funnel the data represents |
| AI referral sessions and conversions (analytics) | The measurable floor |
Then combine signals into a range for AI-influenced pipeline: tracked AI referrals as the low end, and self-reported AI plus AI-referred leads as the higher estimate. See attribution for AI search for how the pieces fit together, and put the result on your AEO reporting dashboard.
Pitfalls to watch
- Recall bias: people remember the last or most memorable touch, not the first
- Option order bias: the first options get chosen more often
- Ambiguity: "Google" can mean classic results, AI Overviews or Gemini
- Junk answers on required fields: "asdf" and "n/a" mean the field is annoying people
- Small samples: a few dozen responses a month can swing widely; look at quarterly trends
Example: what the data might show
This example is illustrative; the numbers are made up.
A B2B analytics company adds the hybrid field to its demo form in January. By the end of the quarter:
- 620 demo requests, 540 with an answer to the field
- 9% of answers coded to an AI assistant, most of them naming ChatGPT
- In GA4, only a small fraction of demo requests had an AI referrer
- Sales notes show several prospects asked assistants for "alternatives to" a competitor
The team reports AI-influenced demo requests as a range between the referral-tracked count and the self-reported count, and adds "alternatives to [competitor]" prompts to its tracking set.
The bottom line
Analytics shows the AI traffic that clicks through with a referrer; buyers can tell you about the rest. Add a hybrid "How did you hear about us?" field with AI assistants named explicitly to demo, sign-up and checkout flows, ask the same question in sales calls, code the answers consistently, and report the results as a trend and a range next to referral data and prompt tracking.
