The AI Attribution Problem: Proxies That Work When Clicks Don't

Why is AI attribution hard?

Because the journey breaks in multiple places at once, and none of the usual fixes apply cleanly. A buyer might ask ChatGPT a question on their phone, see your brand named, open your site in a different browser or device later, convert through a form that never carries a referrer, and never click a tracked link in between. Bessemer Venture Partners' interviews with marketing leaders on how they're measuring AI search are blunt about the state of the art: Natalia Bandach, Senior Director of Growth Marketing at Cloudinary, told Bessemer "whoever says they have it perfectly set up is probably trying to sell something. It doesn't really exist right now." A CMO at a vertical software company interviewed separately put it plainly: "Would I say it's a 100% science? Absolutely not. But that's the nature of the beast right now."

Fewer clicks make the problem worse at the source. Pew Research Center found that people saw a traditional link click in 8% of visits when a Google AI summary appeared, versus 15% when it didn't — a gap measured across 900 US adults and roughly 69,000 searches. That's a Google-specific, click-only measure, not a general "AI traffic drop," but it points at the same mechanic: AI answers increasingly end the interaction before a click happens at all, so click-based attribution is structurally missing more of the real behavior with every passing quarter, not less.

Which proxies work?

None of these alone is proof. Together, moving in the same direction, they're a reasonable case.

Natalia Bandach, Senior Director of Growth Marketing at Cloudinary, told Bessemer she runs three proxies together: AI visibility scores against a share-of-voice target, self-reported attribution surveys, and AI-chatbot referral traffic tracked in GA4 — not one number standing in for the whole picture. Wix VP of Marketing Paula Ximena Mejia added a useful reframe for a world with fewer clicks: "Traffic isn't the goal — conversions are." A smaller, lower-converting pool of non-branded visitors paired with a stronger conversion rate can mean AI is sending more qualified people, even while raw traffic looks flat or down.

How to set up self-reported attribution?

Add an "AI assistant (ChatGPT, Gemini, etc.)" option to whatever source-of-discovery question you already ask — a signup survey, a demo-request form, or a post-purchase "how did you hear about us" field. Keep the list of options short and specific rather than lumping every possible source into "other," since a buried option gets skipped. Review responses monthly rather than per-lead; a handful of noisy data points per week becomes a real trend over a quarter.

Pair this with the branded-search check from GA4/Search Console: a weekly or monthly look at branded-query impressions and clicks versus non-branded, watching for a branded-search lift in the weeks after a known citation or mention event. Neither signal is conclusive alone; a self-report that lines up with a branded-search bump in the same window is a reasonable case.

What should be reported monthly?

A small, consistent set beats a sprawling dashboard that changes every month:

• AI Assistant channel sessions (or your custom channel group's AI rows), trended month over month.
• Self-reported "AI assistant" responses as a share of all attributed leads.
• Branded search impressions/clicks trend from Search Console, split branded vs non-branded.
• Direct traffic trend, flagged against any known citation or mention events that month.
• A running log of sales-call or support mentions of AI engines, counted, not estimated.

Report these as a panel of signals pointing the same direction (or not), with the caveats attached each time — not as a single blended "AI-attributed revenue" figure, which implies a precision none of these inputs actually has.

What would count as overreach?

Claiming a specific dollar figure or percentage of revenue "came from AI" based on proxies alone is the clearest overreach — none of these signals supports that level of precision, and a client or boss who checks the underlying math will find the gap. Crediting a single deal to a single AI mention without corroborating evidence (a self-report, a sales-call note) is the same mistake at deal level instead of portfolio level. And treating a quiet AI Assistant channel as proof AI isn't working is backwards — it may just mean the traffic is arriving with no referrer, as covered in tracking AI traffic in GA4, not that it isn't happening.

The honest framing, and the one every interviewed leader above actually used: contribution, not attribution. Did this touchpoint plausibly influence the decision, alongside everything else that did? That's a defensible claim. A precise credit split across untrackable, fragmented journeys is not.

Checklist

• A self-reported "AI assistant" option exists on at least one lead-capture form.
• GA4's AI Assistant channel or a custom channel group is set up and reviewed monthly.
• Branded vs non-branded search is tracked in Search Console and checked against known AI-mention events — see branded search and AI visibility for why this specific proxy tends to move first.
• Sales and support teams have a simple way to log "a prospect mentioned ChatGPT/Gemini/etc." in a call or chat.
• Monthly reporting presents these as a panel of proxies, not a single attributed number.

FAQ

Is there any tool that gives exact AI attribution today? No credible one. Every proxy described here, including GA4's own AI Assistant channel, has a stated gap (missed referrers, self-report bias, correlation without causation). Treat any tool claiming exact AI-attributed revenue with suspicion.

How many proxies should I track before reporting anything? At least two that don't share the same blind spot — for example, a referrer-based metric (GA4) paired with a non-referrer metric (self-report or branded search), since both fail in different ways and rarely fail together.

Does zero-click behavior make attribution hopeless? It makes click-based attribution weaker, not hopeless. Zero-click research from Pew, Ahrefs and Seer shows fewer clicks, not zero influence — the proxies above exist precisely because influence still shows up in branded search, self-reports and sales conversations even without a tracked click.

Should agencies build this into client reporting? Yes, as a disclosed panel of proxies with their limits stated up front, not as a single polished number. Clients who later ask how a figure was calculated deserve an answer that holds up.

What to do next

1. Add an "AI assistant" option to one existing lead-capture or post-purchase survey this month.
2. Set up GA4's AI Assistant channel (or a custom channel group) and Search Console's branded-vs-unbranded view if you haven't already.
3. Build one monthly reporting panel with at least two independent proxies, each labeled with what it does and doesn't show.

ContextOwl sells one-time AI visibility reports. The sample report shows what a point-in-time evidence snapshot looks like — a complement to the ongoing proxies above, not a replacement for them.