How to Put a Dollar Figure on Your AI Revenue Leak

Nobody has ever approved a budget for a score of 66 out of 100.

If you've run the tests, asking the assistants what they say about your products and sending an agent through your buying flow, you now have a list of failures. That list is interesting to you and completely unpersuasive to whoever controls spend, because it's written in the wrong unit. It needs to be in dollars per month.

Here's how to do that arithmetic honestly, including the parts most vendors quietly skip.

The core formula

For any single failure:

Monthly loss  =  AI sessions with this intent
               × probability you'd have won it
               × conversion rate
               × average order value

Four terms. Two you already know, two you have to estimate. The credibility of the whole exercise depends entirely on how you handle the two you estimate, so let's be precise about each.

Term 1: AI sessions with this intent

Not your total traffic. Not even your total AI traffic. The traffic that would have been in the market for this specific thing.

Three ways to get it, in descending order of confidence:

From your logs. Filter for known assistant and agent signatures, then bucket by the product or category pages they touched. This is the most defensible source you have, and most merchants have never looked at it. Do this first.

From the delta. Compare your server-side session count against what your analytics reports for the same window. Direct traffic that grew while nothing else changed is a reasonable proxy for arrivals your analytics couldn't classify. Cruder, but directional.

From category share. If you know your total AI-attributable traffic and this product is 3% of your catalog revenue, start with 3% of that traffic. Weakest of the three. Label it as such.

A note on the trend: AI referral traffic to US ecommerce has been roughly doubling every two months. Whatever number you land on is a floor for next quarter, not a ceiling. Say so, but don't build the projection into the headline figure, or you'll get argued out of the whole thing on the one assumption you inflated.

Term 2: Probability you'd have won it

The judgement call, and where honest analysis separates from vendor theatre.

If an assistant recommends four products and you sell one that genuinely fits, you weren't guaranteed the sale, you were guaranteed consideration. Your odds of being picked are roughly your competitive share among the products that would have qualified.

Practical approach: for each failed query, count how many products the assistant did recommend. If it named four and yours legitimately belongs in that set, use 1/5 as your baseline: the four named, plus you. Adjust up if you have a real advantage (better price, in stock when others aren't), down if you were marginal.

This number is going to feel uncomfortably low, and it should. A model that assumes you win every query you failed produces a headline figure your CFO will discard in about four seconds, taking the credible parts with it. Conservative here buys you everything later.

Terms 3 and 4: Conversion rate and AOV

These you have. Two cautions:

Use the right conversion rate. Shoppers arriving from AI convert around 30% better than shoppers from traditional search. If you use your blended site-wide rate, you'll understate the loss. If you have a segment for it, use that; if not, use blended and note that it's conservative.

Use AOV, not item price. A shopper who came for the boots buys socks. Your AOV already reflects that; your product price doesn't.

A worked example

Say the assistants get one of your best-selling boots wrong: the page never states it's waterproof, so it's excluded from every waterproof query.

• AI sessions with this intent: 900/month, from log analysis of your boot category
• Probability you'd have won: 1/5 = 20%, since the assistant named four alternatives, yours legitimately competes
• Conversion rate: 2.4%, your AI-segment rate
• AOV: $165

900 × 0.20 × 0.024 × $165  ≈  $713/month

Seven hundred dollars. For one attribute, on one product.

That number is unimpressive alone, and that's exactly the point of doing it this way. It becomes a business case when you run the same arithmetic across the catalog: forty products with the same class of problem, at similar volumes, is roughly $28,000 a month, and the fix for all forty is the same mechanical change.

A CFO will fund $28,000/month against a day of work. Nobody funds a 66.

Four rules that keep this honest

Show every assumption inline. Every estimate visible next to the number it produces. The first question you'll be asked is "where did 20% come from," and having the answer already on the page is the difference between a credible analysis and a sales deck.

Give the range, lead with the low end. Model pessimistic, expected, and optimistic. Put the pessimistic number in the headline. If the conservative case justifies the work, you never have to defend the optimistic one.

Don't count the same session twice. One shopper asking about hiking boots who hits three of your failures is one lost sale, not three. Deduplicate at the session level before summing, or your total will be inflated by exactly the factor someone will catch.

Measure the fix. This is the one that matters most and gets skipped most. Re-run the identical query after the change and record the new answer. "The assistant now recommends us for this query, and it didn't in June" is worth more than any projection, because it's not a projection.

What this gets you

Two things, and the second is the real one.

You get a number you can defend in a meeting. And you get a ranked list, because once every failure has a dollar figure, prioritisation stops being an argument about which problem feels worse and becomes arithmetic. Fix the $4,000/month one before the $200/month one. Obvious, and impossible without the numbers.

This is the entire reason we built revenue leak attribution into Product Gale rather than shipping another scanner. Every finding carries its own figure, built from your traffic, your AOV, and your conversion rate, with the assumptions on screen, and after a fix ships, the same test runs again so you can see whether the answer actually changed.

But you don't need us to start. You need your logs, twenty questions, and a spreadsheet with four columns.

Traffic growth and conversion comparisons: Adobe Analytics, reported via Forbes and Marketing Week, 2025-2026.