Why every AI attribution tool undercounts, and why a minimum is the number you want

If you read the terms of almost any attribution product carefully, you will find a sentence like “figures are estimates” or “some conversions will not be attributed”. Ours says it plainly: AI figures are minimums, not totals, and some AI-influenced customers will never be identified. That can look like a weakness. It isn’t. It describes how the web actually works in 2026, and any tool that doesn’t say it is hiding the same limits.

Here are the five places AI customers disappear, what each one means for your numbers, and how to get the most complete picture you honestly can.

Blind spot 1: the customer never clicks

This is the big one. Someone asks ChatGPT “who’s the best mover in Tampa?”, gets three names, and then does one of two things: types your business name into Google, or calls the number straight away. There is no link between the AI answer and your website, so no analytics tool, pixel or plugin can connect them. In your reports, that customer shows up as a branded Google search or a phone call from “unknown”.

What you can do: ask. A “How did you first hear about us?” field with ChatGPT, Gemini and “another AI assistant” as named options, plus the same question on phone calls, recovers a large share of these customers. We explain how to word it in the most valuable question you’re not asking.

Blind spot 2: Google’s own AI looks like normal search

Clicks from Google’s AI Overviews and AI Mode arrive as ordinary Google search traffic. Google counts them inside the normal Search Console performance data rather than as a separate source (see Google’s AI features and your website documentation). No third-party tool can split them out reliably, so responsible tools leave them under Google instead of guessing.

Blind spot 3: app links arrive with no source

Links opened from the ChatGPT, Gemini or Perplexity mobile and desktop apps don’t always pass a referrer. If the link also has no UTM tag, the visit lands as “Direct”. ChatGPT helps by adding utm_source=chatgpt.com to many links, but not every assistant does, and not every link carries it. Our AI traffic tracking guide lists the sources worth checking.

Blind spot 4: people switch devices

A customer researches on their phone at lunch and books on a laptop that evening. Without a login that ties the two together, they are two different visitors. The AI visit and the booking can’t be joined, so the booking is credited to whatever brought them back, usually a Google search or a typed-in URL.

Blind spot 5: memory is imperfect

Self-reported answers fill gaps, but people misremember. Someone who asked ChatGPT and then Googled you may say “Google”. Someone who saw you in a Facebook group may say “ChatGPT” because it’s top of mind. This cuts both ways, which is why self-reported answers should be labelled separately from tracked ones, never blended in silently.

Why a minimum beats a big estimate

Some tools respond to these gaps by modelling them: applying a multiplier or an algorithm to fill in what they can’t see. The result is a larger, rounder number. The problem is that nobody can check it. When your boss, client or accountant asks “how do you know ChatGPT brought in $40,000?”, a modelled answer is “the model says so”.

A minimum works the other way round. Every customer in it has evidence: a tracked visit from an AI assistant, or the customer telling you an AI sent them. So:

• It can be defended. You can open any lead and show why it counts.
• It only moves for real reasons. It rises when more AI customers arrive, not when a model is re-tuned.
• It is safe to act on. If the minimum already justifies spending on AI visibility, the real return is at least that good.

How to read your AI numbers

Label	Evidence	How to treat it
Confirmed	The visitor arrived from an AI assistant before becoming a lead	Hard evidence
Declared	The customer said an AI assistant sent them	Strong, especially in aggregate
Assisted	An AI visit happened earlier in the journey	Supporting evidence
Everything else	No AI evidence	Stays in its normal channel, including AI Overviews clicks

The full walk-through is in Confirmed vs Declared: reading an AI attribution report. The practical rule: report Confirmed plus Declared as your AI minimum, show Assisted alongside it, and say out loud that the true figure is higher.

Getting as close to the full picture as you honestly can

1. Track AI visits properly. Include every assistant, not just the ones in GA4’s default AI channel. Our AI referral regex generator builds the list for GA4, GTM and Looker Studio.
2. Follow visits to the lead or sale. Visits don’t pay the bills. Connect the source to the form, booking or order (see WooCommerce orders from ChatGPT for a store example).
3. Ask every customer. On forms and on calls, with AI tools named as options.
4. Keep the evidence types separate. Never merge tracked and self-reported answers into one unlabeled number.
5. Watch the trend, not the decimal. A minimum that doubles over a quarter is a clear signal, even though neither number is the full total.

This is exactly how Which AI Sent is built: it tracks AI visits to the lead or sale, collects the “how did you hear about us?” answer, and labels each AI customer by its evidence, so the figure you share is one you can stand behind.