How to Measure Chatbot Attributed Revenue: A Technical Guide (2026)

The five parts of a chatbot attribution pipeline

1. The trigger: what counts as a touch

The trigger marks a visitor as touched by the chatbot. Loose triggers count widget impressions or opens. Tighter ones require the visitor to send a message or the bot to give a real answer: Alhena publishes that a conversation qualifies only if an AI response found an answer in product data, and Gorgias voids attribution when a human takes over. Choose a trigger that requires an actual exchange. A visitor who saw a chat bubble was not sold anything by it.

2. Identity stitching: linking the chat to the order

A conversation happens in a browser session and an order happens at checkout, so something has to connect them. Session or cart tagging writes an identifier into the session or cart when the visitor chats; on Shopify, cart attributes are copied onto the order, so the evidence lives in your own store data. Cookie matching works within one device but breaks when a shopper chats on a phone and buys on a laptop. Email matching catches cross device purchases, but only for visitors who shared contact details. Chat only discount codes are a clean secondary signal that undercounts, because most chat assisted buyers never use them. SparkGPT uses session tagging: when a visitor chats, the session is tagged, and orders are matched against that tag.

3. The window: how long a touch stays valid

The window is how long after the conversation an order still counts. It is the biggest lever on the reported number, because a longer window always sweeps in more purchases that would have happened anyway. For stores selling products under a few hundred dollars, a short window is the honest default: a shopper who asks about sizing and buys that evening was plausibly helped, while one who chatted two weeks ago and returned through an ad was mostly moved by other things.

4. Order value: gross or net

Decide whether revenue is counted before or after refunds, cancellations, taxes, and shipping. Gorgias documents that its Shopping Assistant revenue is net of refunds and excludes voided and pending orders. Whatever you choose, apply it to every channel, or a gross chatbot figure will beat a net email figure for no real reason.

5. Deduplication: one order, several claimants

Almost no chatbot guide covers this part. Klaviyo's default windows are 5 days for email opens, email clicks, and SMS clicks, and Google Analytics 4 defaults to a 90 day lookback for most key events. A shopper who clicks an email on Monday, asks the chatbot about shipping on Tuesday, and buys that night is counted by both Klaviyo and the chatbot, and possibly by an ad platform too. Add up every dashboard and the total can exceed what your store actually sold. The fix: export attributed order IDs from each tool, join them against your order list, and report orders claimed by more than one tool as shared credit.

Attribution windows vendors publish, compared

The table below lists published rules as of September 2026. It measures the rules, not the quality of the products.

Two honest observations. First, SparkGPT is not the only tool using a 24 hour window. Alhena publishes the same length and adds a stricter trigger and a separate experiments layer for causal measurement, which is careful work and worth crediting. Second, the last column is where the incentives differ. When a vendor's fee is independent of the attribution number, a generous rule costs the vendor nothing and makes the dashboard look better. When the fee is a share of the attributed number, as with SparkGPT's 5%, a generous window would raise the merchant's bill, which gives both sides a reason to keep the rule tight and public.

From attributed to incremental: run a holdout test

Attribution answers "which orders followed a conversation." It does not answer "which orders would not have happened without it." People who start a chat are already more engaged than average, so some of them would have bought anyway. To measure the difference, you need a randomized holdout.

Randomly assign each new visitor to one of two groups, for example by a hash of their visitor ID, show the chatbot to one group only, run for at least two full weeks, and compare revenue per visitor. The difference is the chatbot's incremental effect, already net of shoppers who would have bought anyway.

This is the same design researchers use. In a working paper by Fang, Yuan, Zhang, Donati, and Sarvary (arXiv 2510.12049), a randomized experiment covering 44,614 shoppers found that a presale generative AI chatbot raised sales by 16.3% against a control group that did not get it. That is the kind of number an attribution dashboard cannot produce on its own. We review that research in more depth in Does an AI Chatbot Increase Ecommerce Sales?.

Be realistic about traffic. Detecting a small lift needs a lot of visitors. As a rough guide from standard sample size math, at a 2% baseline conversion rate, reliably detecting a 10% relative lift (2.0% to 2.2%) takes on the order of 80,000 visitors in each group. A store with 10,000 monthly visitors will not get a statistically clean answer in a month. In that case, run the holdout longer, focus on revenue per visitor rather than tiny conversion differences, and treat attributed revenue with a short window as your working estimate in the meantime.

A worked example of the full audit

The figures below are hypothetical and chosen only to show the arithmetic.

A chatbot dashboard lists 60 attributed orders worth $4,800 for the month. First, confirm each order links to a conversation where the visitor actually sent a message. Second, match the order IDs in Shopify: if 4 were refunded for $320, net attributed revenue is $4,480. Third, compare against Klaviyo's attributed orders: if 9 of the 60 appear there too, report them as shared credit rather than exclusive chatbot credit. Fourth, until a holdout says otherwise, assume half of the chat buyers would have purchased anyway, which leaves roughly $2,240 of incremental revenue.

On SparkGPT's Growth plan, the month on the dashboard figure would cost $49 plus 5% of $4,800, which is $289. Even against the discounted incremental estimate, that is still a return of more than 7 times cost. For the broader cost and savings math, see AI Chatbot ROI for Small Business.

Six questions to ask about any attributed revenue number

Use these on any vendor's dashboard, SparkGPT included.

1. What exact event marks a visitor as touched, and does a widget impression count?
2. What is the window, and is it measured from the first message or the last one?
3. Can I see every attributed order individually, with its order ID, so I can find it in my store admin?
4. Are refunds, cancellations, and test orders removed?
5. How are orders handled that my email or ad tools also claim?
6. Has the tool ever been tested against a holdout group, and can I run one myself?

A vendor that answers all six clearly is giving you a measurement. A vendor that answers with a conversion rate and a case study is giving you marketing.

How SparkGPT handles attribution

SparkGPT's rule fits in one sentence: when a visitor chats with the agent, the session is tagged, and any purchase within 24 hours is attributed to the agent. Every attributed order is listed in the ROI dashboard for you to match against your store records; the Shopify setup guide shows how. The rule is also the billing rule: the Growth plan costs $49 per month plus 5% of attributed sales, with unlimited conversations and prepaid credits, so the agent pauses instead of charging unexpectedly when the balance runs out. Because the fee depends on the attributed number, the 24 hour window is not a reporting preference we can quietly widen. It is the published price. For more on why pricing on results changes vendor incentives, read Performance Based AI Chatbot Pricing.

Frequently Asked Questions

What is chatbot attributed revenue?

It is the total value of orders linked to a chatbot conversation under a stated rule. The rule defines what counts as a conversation, how the conversation is matched to the order, and how long after the conversation an order still counts. Without all three written down, the number cannot be compared or audited.

What attribution window should I use for an ecommerce chatbot?

For most stores selling products under a few hundred dollars, a short window of about 24 hours is the honest default, because purchases soon after a conversation are the ones the answer plausibly influenced. Longer considered purchases can justify a longer window, but every extra day adds orders that would have happened anyway.

How do I stop my chatbot and email tool from counting the same order twice?

Export the attributed order IDs from each tool for the same period and join them against your store's order list. Orders claimed by more than one tool are shared credit. Report them as overlap rather than adding every tool's total together, which can add up to more than your store actually sold.

What is the difference between attributed revenue and incremental revenue?

Attributed revenue counts orders that followed a conversation. Incremental revenue counts only orders the chatbot caused, meaning orders that would not have happened without it. The only reliable way to measure incremental revenue is a randomized holdout that hides the chatbot from a share of visitors.

How much traffic do I need to run a chatbot holdout test?

More than most small stores expect. At a 2% baseline conversion rate, reliably detecting a 10% relative lift takes roughly 80,000 visitors in each group. Smaller stores should run the test longer, compare revenue per visitor, and rely on short window attributed revenue as a working estimate.

Can I verify SparkGPT's attributed orders myself?

Yes. Every attributed order appears in the ROI dashboard. Pick a few, find the same orders in your store admin, and confirm each was placed within 24 hours of the conversation it is linked to. Since the 5% fee is charged on exactly these orders, the dashboard is also your invoice detail.

Measure it on your own store

The fastest way to understand attribution is to watch it work on your own traffic. Build an agent from your website URL on SparkGPT's Free plan in about 10 minutes, with no credit card, and test it without limits. When you go live on Growth at $49 per month plus 5% of attributed sales, every order the agent earns within 24 hours shows up in the dashboard for you to audit. Questions about how attribution works for your store can go to hello@sparkgpt.ai.

Build your free AI sales agent at www.sparkgpt.ai