How to attribute ChatGPT-influenced purchases across direct, search, and devices

ChatGPT-influenced purchases cannot be recovered reliably from the converting session alone. Measure them as two separate things: ChatGPT-sourced revenue, where the click is observable, and ChatGPT-influenced revenue, where another signal shows that ChatGPT shaped the journey. Combine source data, first-party identity, post-purchase survey responses, and controlled tests; report the evidence and confidence level instead of assigning every direct or branded-search order to ChatGPT.

Start with the distinction your reports are hiding

A ChatGPT click and a ChatGPT recommendation are different measurement events. A shopper may click a cited product page from ChatGPT, leave, search your brand later, and buy. Another shopper may read a recommendation in ChatGPT, open your site manually on a phone, and convert on a laptop. The final order contains no reliable record of the original recommendation unless you captured another signal.

Use three reporting buckets:

Bucket	What it proves	Typical evidence
ChatGPT-sourced	ChatGPT was the observed source of a visit in the tracked journey	Referrer, campaign parameter, landing session, and order path
ChatGPT-assisted	ChatGPT was reported or identified as an earlier touchpoint, but did not receive the final click	Post-purchase response, known-user journey, or an experiment
ChatGPT-influenced	ChatGPT likely changed consideration or demand, supported by an aggregate lift or repeated evidence	Survey data, matched cohorts, or incrementality testing

Do not add these buckets together without deduplication. One order can be ChatGPT-sourced and also involve an on-site recommendation, while an influenced order may never show ChatGPT as a referrer.

Google Analytics defines attribution as assigning credit to touchpoints along a path to a key event. Its paid-and-organic last-click model ignores direct traffic when another non-direct touchpoint exists, so a path ending in direct traffic can give credit to an earlier channel rather than prove that the earlier channel caused the sale. Read Google’s attribution overview before comparing channel reports.

Capture the ChatGPT click before it disappears

The first step is to make observable ChatGPT visits distinct from ordinary referral and direct sessions. Keep the raw source, medium, referrer, landing page, timestamp, campaign parameters, device type, and session ID with the order or customer journey.

When you control a link that will be shared, use a consistent campaign convention such as utm_source=chatgpt. Add a medium and campaign value that your analytics team has defined, and preserve those parameters through redirects and checkout. This identifies tagged clicks; it does not identify people who saw an untagged recommendation and later navigated manually.

Also inspect server and storefront data. A missing referrer can place an AI-assisted visit in direct traffic, while a present referrer may place it in referral traffic. Shopify describes a referrer as the method a customer used to reach the store and provides marketing reports with attribution-model options; its documentation also says an order with only direct traffic is attributed as direct. See Shopify’s marketing reports documentation.

Create a dedicated ChatGPT view rather than renaming all direct traffic. Useful dimensions include:

observed source and medium;
first landing page and product page;
new versus returning customer;
order ID, revenue, refund status, and customer ID;
first-touch source, converting source, and all known touchpoints;
whether the source was referrer-based, tagged, self-reported, or experimental.

Label evidence by collection method. “Referrer” and “survey” are not interchangeable, and a report that combines them without labels is difficult to defend.

Stitch sessions and devices with first-party identity

Cross-device attribution requires a stable first-party identifier that the shopper has deliberately supplied or that your platform can lawfully use. A login, account ID, email captured before purchase, loyalty ID, or checkout customer ID can connect an earlier visit to a later order; anonymous browser activity generally cannot be joined with certainty across devices.

Send the same identifier consistently to your analytics and commerce systems, subject to consent and your privacy policy. Store event times and the identifier’s source so an analyst can distinguish a deterministic match from an inferred one.

A practical event sequence looks like this:

The shopper arrives from a measurable ChatGPT link, or completes a “How did you hear about us?” interaction.
The site records the first-touch source and landing page.
The shopper identifies themselves through login, email capture, loyalty, or checkout.
The commerce system connects the known identifier to the order.
The report retains both the original source and the source of the converting session.

Do not use device, IP address, timing, or similar browsing patterns as if they were a confirmed identity match. If you use probabilistic modeling, report it separately as modeled influence and publish the matching rules, lookback window, and error assumptions.

Ask buyers what analytics cannot observe

A short post-purchase survey is the most direct way to capture ChatGPT-assisted discovery that ended in direct traffic, branded search, or another device. Ask immediately after purchase or in the order follow-up while the decision is still recent.

Keep the question neutral and include an explicit option, for example:

Where did you first learn about or research this product?

Search engine

Social media

Friend or recommendation

ChatGPT or another AI assistant

Email

I already knew the brand

Other

Not sure

A follow-up can ask which assistant and what they searched for, but do not make the survey so long that response quality collapses. Save the response with the order ID, customer ID where available, date, and whether the answer was single- or multi-select.

Survey responses are evidence of reported influence, not proof that ChatGPT caused the purchase. Report response rate, the share selecting ChatGPT, net revenue from those orders, and the uncertainty created by nonresponse. A respondent who says “ChatGPT” may have also encountered paid search, reviews, or a recommendation from another person.

Report revenue in separate views

A defensible dashboard should show observed acquisition and inferred influence side by side, not force both into one precise percentage.

View	Recommended measures	Do not claim
ChatGPT-sourced	Sessions, orders, gross and net revenue, conversion rate, average order value, landing pages	That it includes people who saw ChatGPT but did not click through
ChatGPT-assisted	Orders with a survey response or known earlier ChatGPT touchpoint, response rate, revenue	That survey-selected revenue is fully incremental
Direct and branded-search overlap	Returning visitors, time from first known ChatGPT touchpoint to order, customer-level overlap where consented	That every direct or branded order was caused by ChatGPT
Incrementality	Conversion or revenue lift versus a comparable control, with confidence intervals	That correlation in a path report is causal proof

Keep gross revenue, discounts, refunds, and cancellations defined consistently. If your business reports contribution margin, include it alongside revenue rather than presenting revenue as profit.

Anagram’s AI Visibility product is relevant to the discovery side of this measurement problem: its page says it shows how a brand appears in ChatGPT, how it compares with competitors, and what customer questions reveal about where to focus. That can help connect changes in visibility, mentions, competitors, and citation sources with later demand analysis. It does not, by itself, turn an unobserved ChatGPT recommendation into a confirmed customer-level conversion path.

Use tests to estimate influence, not just paths

Incrementality testing is the strongest way to estimate whether ChatGPT visibility or recommendation exposure changes purchases beyond the conversions you can directly track. A path report describes what happened; a test compares what happened with exposure against a credible counterfactual.

Possible designs include:

compare matched geographic or audience cohorts exposed to different ChatGPT visibility conditions;
run a before-and-after analysis around a clearly documented change, while controlling for promotions, seasonality, stock, and media;
use a holdout where practical and measure branded demand, direct orders, new customers, and net revenue;
compare survey-reported ChatGPT influence with changes in observed ChatGPT-sourced traffic rather than treating either as the whole market.

Document the test window, eligible population, control method, primary outcome, exclusions, and other campaigns running at the same time. If you cannot create a credible control, call the result an association or directional signal.

Set an attribution policy your finance team can audit

Choose the lookback window, identity rules, source hierarchy, survey treatment, and deduplication method before the next reporting period. Then keep the raw events so a later model change does not rewrite history without explanation.

A useful policy might say:

direct referrer evidence qualifies an order as ChatGPT-sourced;
a self-reported ChatGPT response qualifies it as ChatGPT-assisted, with survey response rate shown;
a cross-device journey qualifies only when the identifier match is deterministic and consented;
modeled matches remain in a separate modeled-influence view;
tested lift is reported as incremental revenue or conversion change, not assigned order by order;
source-reported and influenced revenue are never summed as if they were mutually exclusive.

The result is less dramatic than claiming that ChatGPT caused every direct or branded-search purchase. It is more useful: your team can see what ChatGPT demonstrably sent, what shoppers say it influenced, what devices and sessions can be joined, and what a controlled test supports.