How much revenue do you lose when AI agents can't buy?

"AI agents can't check out on our store" is a sentence that gets nodded at in a meeting and then deprioritised, because it has no number attached. This post shows how to attach one: an estimate of the revenue a store loses each month when AI agents fail somewhere between finding a product and paying for it, with a calculator you can run on your own figures.

The model is the one tragent uses in its reports. It is deliberately conservative, and every assumption is visible.

How do you estimate revenue lost to AI agent failures?#

Start from the agent business you would have if everything worked, then walk the purchase funnel and remove what each failing stage blocks. Four inputs give the starting point:

• Monthly sessions on your store.
• Agent share: the fraction of those sessions that are AI agents or arrive through AI assistants.
• Agent conversion rate: how often such a session would end in an order if nothing stood in the way.
• Average order value.

Multiplied together they give the monthly revenue agents would bring if every stage worked. A store with 200,000 monthly sessions, 2% from agents, 2.5% conversion and an $80 average order has $8,000 a month of agent-driven sales at stake.

Then the funnel. An agent has to find the product, read it, see the price, check stock, sign in where needed, add to the basket, get through checkout, and pay. A stage that fails blocks all the agent revenue still reaching it. A stage that only partly fails, working for some agents and not others, blocks part of it.

How do you avoid counting the same lost order twice?#

Walk the stages in order and let each one block only what is still reaching it. This is the step most back-of-envelope estimates get wrong.

Suppose checkout fails and payment also fails. Adding "revenue lost at checkout" to "revenue lost at payment" counts every order twice: no agent that was stopped at checkout ever reached payment. In a sequential walk, checkout blocks 100% of what reaches it, and payment is attributed nothing until checkout is fixed. It is masked by the earlier failure.

Two consequences matter for planning:

• The total can never exceed the baseline. However many stages fail, you cannot lose more than the agent business you had.
• The first failing stage is the one to fix. It is the binding constraint. Fixing a later stage first produces no visible result, which is a common way for agent work to look like it did nothing.

Warnings are handled the same way. A stage that worked on some runs and not others is treated as blocking 35% of what reaches it, not all of it, so a later failure can still claim the remaining 65%.

Try it on your store's numbers#

Change the inputs and click a stage to mark it as passing, warning or failing. The figures update as you go, and nothing you type leaves your browser.

Monthly sessionsAgent share of sessions%Agent conversion rate%Average order value$

Estimated revenue at risk

$8,000/ month

100% of $8,000 a month in agent-driven sales.

Fix first: Reach checkout, the earliest stage losing agent revenue. Later stages are only priced on what gets past it.

Estimate, econ-v1. A failing stage blocks all agent revenue still reaching it; a warning blocks 35%; sign-in counts half, since guest checkout is common.

Click a stage to cycle pass → warn → fail

• Find productspass—
• Read the detailspass—
• See the pricepass—
• Check stockpass—
• Sign inpass—
• Add to basketpass—
• Reach checkoutfail$8,000
• Paypass—

The example starts with checkout failing, which on the default inputs puts the whole $8,000 at risk. Mark checkout as a warning and the figure drops to $2,800. Then mark payment as failing too, and the total returns to $8,000, split between the two stages. That is the walk at work.

Where do the inputs come from?#

Sessions and order value from your analytics; agent share from your CDN or bot manager; conversion is the assumption to be most careful with.

• Agent share. Your bot manager or CDN classifies verified bots and signed agents and can report their share of requests. Add sessions that arrive with AI-assistant referrers, such as utm_source=chatgpt.com. If you have neither yet, start with a low figure and say it is a guess.
• Conversion rate. Here is the trap. Your observed conversion for agent traffic is already depressed by the very failures you are trying to price; using it as the "if everything worked" rate counts the damage twice. If you have no clean figure, use your site-wide conversion as the assumption and label it. tragent's measured model handles this by deriving the achievable rate from observed conversion and the share of agents that got through.
• Order value. Your average is a reasonable start. Agents shopping on a brief tend to buy what they were asked for, so check whether AI-referred orders differ.

How big is AI agent traffic today?#

Small for most stores in absolute terms, growing quickly, and converting better than it did a year ago. The figures depend heavily on who is counting what.

Adobe, which measures traffic across a large share of US retail sites, reported AI-referred traffic to US retailers up 393% year on year in the first quarter of 2026, and said it was converting 42% better than non-AI traffic by March, a reversal from a year earlier. A study from a German university, measuring ChatGPT referrals as a share of all ecommerce traffic, found them below 0.2%.

These do not contradict each other. One measures growth and quality across a category; the other measures share in absolute terms. For a single store the honest position is: the number is probably small today, it is the fastest-growing source most stores have, and its visitors arrive closer to a purchase. That is the case for fixing the funnel now, while the cost of the gaps is still small.

What does this model leave out?#

Several real costs, on purpose, so that the figure it does give is defensible.

• Returns and support. A broken returns flow costs support time and repeat custom, but it cannot block an order that already happened, so the model attributes no purchase revenue to it.
• Lost lifetime value. A shopper whose assistant could not buy from you once may not be sent back. That is real and not modelled.
• Visibility. If assistants stop recommending you because your data is unreliable, the loss starts before any session. The model only counts sessions that happen.
• Sign-in is counted at half. Many stores allow guest checkout, so a broken sign-in blocks only some purchases. If yours requires an account, the real figure is higher.

So read the result as a floor for the revenue at stake from the failures you know about, not a forecast of your agentic commerce business.

When should you replace the estimate with a measurement?#

As soon as you have enough agent sessions with outcomes to measure from. An estimate is built from inputs someone chose; a measurement comes from what agents actually did on your store.

tragent's measured model prices your own agent traffic: sessions classified as agents, the stage each one reached, and the orders that resulted. It refuses to produce a figure below 200 agent sessions, or without recorded outcomes or an order value, because a partial feed produces zeros that look exactly like a store with no problems. Until you reach that, the estimate is the better guide, as long as it is labelled as one.

Frequently asked questions

How do I calculate revenue lost to AI agents?

Multiply monthly sessions by the agent share, the conversion rate agents would have if nothing blocked them, and average order value. Then walk the purchase stages in order, letting each failing stage block only the revenue still reaching it.

Why not add up the loss at every failing stage?

Because an agent stopped at one stage never reaches the next. Adding the stages counts the same lost orders more than once. A sequential walk attributes each order to the first stage that blocks it.

What share of my traffic is AI agents?

Your CDN or bot manager can report verified bots and signed agents, and analytics can segment AI-assistant referrals such as utm_source=chatgpt.com. For most stores the share is small today and growing quickly.

Is AI agent traffic worth fixing for yet?

It is small for most stores but growing fast; Adobe reported AI-referred retail traffic up 393% year on year in early 2026 and converting better than other traffic. The gaps are cheapest to fix while the volume is low.

Which failure should I fix first?

The earliest stage in the purchase funnel that fails. It blocks everything behind it, so later fixes show no result until it is resolved.

Sources

1. AI traffic to US retailers jumps 393% in Q1 as agentic shoppers outspend humans — Decrypt. Secondary report of Adobe Analytics data
2. AI-driven retail traffic surges 393% in Q1 2026 — but read the data carefully — Stellagent. Covers the German university study on ChatGPT referral share

Written by the tragent team and checked against the sources listed above on October 19, 2026. See how we write and correct posts in our editorial policy.