AI search attribution is the practice of connecting revenue to the points in a buyer's journey where an AI assistant, such as ChatGPT, Copilot, or Perplexity, appeared. For one B2B client, orders whose journeys included ChatGPT totaled over $230,000 between February and August 2026, and we traced every one of those journeys from May through August. In that period, about $56,000 of the $116,000 in ChatGPT-assisted revenue, or 49%, came from orders where ChatGPT was not the final touch, and a last-click report would not have given ChatGPT any credit.
Attribution can only credit the touches it can observe, and AI assistants can influence a buying decision well before any trackable click happens. A buyer might ask ChatGPT which suppliers carry a part, read the answer, and then, days later, arrive via a Google search or by typing the URL, at which point the report credits whichever channel happened to come last.
Kevin Indig makes this case in The Usefulness of Attribution for AI Search, arguing that click-based attribution was always a weak fit for non-paid channels and that AI makes the weakness impossible to ignore. We agree with that diagnosis, and running an attribution platform has led us to a practical conclusion: attribution stays useful when you are precise about which part of the journey it covers and build the rest of your measurement around that boundary. Our client's data shows where that boundary sits for one business.
Our client is a regional B2B distributor, and its ChatGPT-assisted revenue stayed above $26,000 in each of the seven months in our data. Using Galileo, Arcalea's multi-touch attribution platform, we tracked every order from February through August 2026 whose journey included at least one visit from ChatGPT, and we refer to the revenue from those orders as ChatGPT-assisted revenue.
| Month | ChatGPT-Assisted Customers | ChatGPT-Assisted Revenue |
|---|---|---|
| February | 32 | $41,000 |
| March | 41 | $41,000 |
| April | 41 | $32,000 |
| May | 29 | $34,000 |
| June | 32 | $27,000 |
| July | 32 | $27,000 |
| August | 46 | $29,000 |
| Total | 253 | ~$230,000 |
From April through August, ChatGPT appeared in 198 orders totaling roughly $116,000 (some customers placed multiple orders). In 121 of those orders, ChatGPT was not the final touch, and Google Ads, organic search, or a direct visit completed the sale.
A last-click model would have assigned those 121 orders, or 49% of the period's ChatGPT-assisted revenue, entirely to other channels. We found the gap by comparing a last-touch view of these orders with a view that credits every journey containing ChatGPT. When an AI assistant appears early in the purchase process, and another channel closes the sale, a model that rewards only the final click gives the assistant no credit. The journeys were also long, and in August the median ChatGPT-assisted order included six touches across about three distinct channels before checkout.
From April through August, between 72% and 84% of ChatGPT-assisted customers each month checked out as guests, with no account and no lead record. That matters because self-reported attribution, the "how did you hear about us" field, depends on a buyer filling out a lead form, which none of these guest buyers did. Galileo still connects these orders to revenue because it links on-site activity to transactions, whether or not the buyer ever becomes a named lead.
Of the 198 orders we reviewed, 186, or roughly 94%, came from outside the client's home market. If ChatGPT-assisted orders skew further outside the home market than overall orders do, AI assistants may be connecting the brand with buyers in markets where its local presence does not reach.
Our research on how AI citations differ by industry offers one possible explanation. It found that in B2B industrial categories, where buyers rarely discuss purchases in public, AI assistants tend to rely on what companies publish on their own websites. An answer built on a company's own site has no reason to favor buyers near its locations, which would fit the pattern we saw here.
Small amounts of revenue from Copilot and Perplexity also appeared in July and August. Those numbers are too small to support conclusions, although they are worth watching in case buyers begin spreading their research across multiple assistants.
Attribution describes only the paths it can observe, which imposes two limits on how we use this data: it cannot see influence that occurred before the first observable touch, and it cannot explain why a trend moved.
The first limit is that $230,000 should be read as a minimum for AI-touched revenue. Galileo sees the buyers who reached the site through ChatGPT at some point in their journey. It cannot see the buyer who read a ChatGPT answer, closed the tab, and searched for the brand on Google a week later, because that visit carries no AI referral, and its revenue is attributed to branded search or direct traffic. That delay appears in outside research as well, since a Profound consumer panel study found that most first visits to a brand after an AI mention arrive more than a day later. The figure should also be interpreted as revenue associated with ChatGPT, since some of these buyers may have purchased anyway. Establishing cause would require a holdout or incrementality test, which journey data cannot replace.
The second limit is that attribution can show movement without showing its cause. Customer counts reached a seven-month high of 46 in August, while revenue per customer fell. That could reflect seasonality, a shift toward smaller replenishment orders, or changes in which questions surface the brand, and attribution alone cannot tell us which, so we treat the divergence as the next thing to examine.
We measure AI search through triangulation, which means combining several signals with different blind spots so that each one checks the others and no single metric has to carry the whole argument.
For our clients, that system has three layers:
When all three layers move together, our confidence in the story rises. When they diverge, as they would if AI citations rose while assisted revenue stayed flat, the divergence tells us where to investigate. A split within one layer, such as rising customer counts alongside falling revenue per customer, deserves the same scrutiny. The exposure layer also contains measurement traps, which we cover in why AI visibility reports are often wrong and in our guide to reading an AEO analysis.
For brands that AI assistants do not yet cite, the exposure layer is usually where progress first appears. Putting attribution in place at the same time means that AI-referred orders are identified as they arrive, while the visibility data helps estimate the influence, ensuring no referral is left behind.
Triangulation also helps in the conversation with finance, where alignment is still rare: a July 2026 WFA and Ebiquity study of 71 senior leaders found that only 14% said marketing and finance in their organizations were aligned on how to define effectiveness. A triangulated system gives both teams a shared body of evidence that each side can inspect.
For this client, a meaningful share of ChatGPT-assisted revenue sat inside journeys that single-touch models were never built to read, and any business whose buyers research with AI before purchasing has reason to check whether the same pattern shows up in its own data. We build measurement systems that capture everything observable, state clearly what they cannot see, and give leadership evidence it can weigh when deciding what to invest in.
If you want to know how visible your brand is in AI search and whether that visibility is already translating into revenue, talk to our SEO, AEO, and GEO team about an AEO Index and a Galileo attribution assessment.