Ecommerce Analytics

Tracking AI traffic in Shopify: what you can prove and what you can't

ChatGPT, Perplexity and Gemini send traffic that most Shopify stores record as direct or unattributed. Here is how to capture what is capturable, why the rest is structurally invisible, and how to report a number that survives scrutiny.

August 4, 2026 · 12 min read

AI-sourced Shopify traffic can be identified in three tiers: confirmed, when the landing URL carries a UTM parameter naming the assistant; high-confidence, when the referrer is an assistant's domain; and not attributable, which covers every shopper who asked an assistant and later arrived directly. Because the third tier is large and permanently invisible, any AI revenue figure is a floor rather than a total.

A merchant asks a reasonable question: how much revenue is coming from AI assistants? The honest answer starts with an uncomfortable admission — you can measure part of it precisely, part of it probably, and part of it not at all. Tools that answer with one confident number have quietly merged all three.

This article covers what signals actually exist, how to capture them in Shopify, why the invisible portion is structurally invisible rather than just hard, and how to report a figure you can defend when someone asks how it was calculated.

The three tiers of evidence

Every order either carries evidence of where it came from or it does not. The discipline is refusing to promote a weaker signal into a stronger claim.

Tier	Signal	What it proves	Failure mode
Confirmed	utm_source parameter naming an assistant on the landing URL	The shopper arrived via a link the assistant generated	Lost if the parameter is stripped by a redirect or a link-shortening app
High	HTTP referrer from an assistant's domain	The click originated in the assistant's interface	Stripped by privacy settings, some browsers, and referrer-policy headers
Unknown	Direct traffic, no referrer, no parameters	Nothing about the source	Contains an unknown quantity of AI-influenced sessions
Invisible	Answer given without a link, or a purchase made later on another device	Cannot be observed from your side at all	No technical fix exists

What each signal actually proves

Capturing what is capturable in Shopify

Shopify records a landing site and a referring site on orders, which is the raw material. The work is preserving those signals and classifying them consistently.

1
Stop stripping parameters
Audit anything that redirects on entry — geo-redirect apps, forced locale routing, HTTP-to-HTTPS chains, marketing link wrappers. Every redirect is an opportunity to drop the query string that carries your only confirmed signal. Test by loading your storefront with a test UTM and checking the parameter survives to the final URL.
2
Preserve the referrer across the session
Capture the referrer on first landing and persist it for the session, rather than reading it at checkout when it has long since been overwritten by your own pages. Shopify's own landing/referring fields do this, but custom analytics frequently do not.
3
Classify with an explicit, versioned rule set
Write the mapping from parameters and referrer domains to sources as code with a version number, not as an ad-hoc set of report filters. When you add a new assistant domain next quarter, you need to know which historical rows were classified under which rules.
4
Store the tier alongside the source
Every order gets both a source and a confidence tier. Without the tier, a confirmed ChatGPT order and a guessed one are indistinguishable a month later, and your totals become uninterrogable.
5
Never overwrite unknown with a guess
Orders with no signal stay unknown. Resist the backfill that assigns them proportionally to known sources — it manufactures precision and makes every subsequent comparison meaningless.

Why Google's AI surfaces are a special case

One distinction is worth stating explicitly because it is the most common source of overclaiming in this category. When Google shows an AI-generated summary above the results and a shopper clicks a link within it, the referrer is Google. It is not distinguishable from an ordinary organic search click by anything visible on your side.

So any tool reporting a specific figure for revenue from Google's AI answers is producing an estimate, not a measurement. The correct handling is to classify those sessions as Google search and say plainly that the AI-summary portion is not separable. That is less satisfying than a breakdown and it is the only version that is true.

Interpreting what you get

Once orders carry sources and tiers, three questions become answerable — and one does not.

Answerable: is confirmed AI traffic growing over time; do AI-sourced sessions convert differently from other sources; which products do AI-sourced shoppers actually buy. That third one is frequently the most useful output, because it tells you which parts of your catalogue assistants are successfully recommending, which is a much sharper signal than an overall citation rate.

Not answerable: what share of total revenue is AI-influenced. You will never know. What you can say is that at least the confirmed and high-confidence portion is, and that the true figure is higher by an unknown amount.

A note on sample size. AI-sourced sessions are usually a small fraction of total sessions, which means conversion-rate comparisons on them are noisy. A week with eleven AI sessions and one order does not demonstrate a 9% conversion rate. Set a minimum sample before you report a rate at all, and show the count next to every rate so nobody reads too much into it.

What good reporting looks like

The presentation matters as much as the collection, because a correctly computed number presented without its caveats will be quoted without them.

State the floor explicitly: "At least X in revenue came from AI assistants, confirmed by link parameters and referrers."
Show the tier breakdown, not just the total — confirmed, high confidence, and the size of the unknown bucket.
Show the underlying counts beside every rate, so a 9% conversion rate on eleven sessions is visibly weak.
Show the date window and the classifier version, so a number from March can be reconciled with a number from August.
Never present a trend line smoothed enough to hide a drop. If AI-sourced revenue fell, that is information.

Frequently asked questions

Does Shopify Analytics show AI traffic?

Only indirectly. Assistant referrers appear under referral sources by domain, and UTM-tagged sessions appear by campaign source, but there is no built-in grouping that says "AI assistants" or that distinguishes an assistant's link from any other referral. You have to define the mapping yourself.

Should I add my own UTM parameters for AI?

You cannot — you do not control the links assistants generate. What you can control is not destroying the parameters they add, which is why the redirect audit is the highest-value step in this article.

Why does GA4 show different numbers from Shopify?

Different attribution models, different session definitions, different handling of consent and blocked scripts. Pick one as your system of record for revenue — Shopify's order data is the stronger candidate because it is server-side — and use the other for behavioural context rather than reconciling them.

Is a growing 'direct' bucket evidence of AI traffic?

It is consistent with AI traffic and with several other explanations, including app changes, referrer stripping, and email clients. Treat it as a prompt to investigate, never as attribution. If you want evidence, probe the assistants directly and see whether your citation rate moved in the same period.

How small is too small to report?

Set a floor before you look at the data, so the threshold is not chosen to flatter the result. A common approach is refusing to report a conversion rate below a fixed minimum session count and showing the raw counts instead.

Will attribution get better over time?

Partially. More assistants are tagging outbound links, which grows the confirmed tier. But the structurally invisible portion — answers with no click, and purchases made days later on a different device — will not become visible, because there is nothing to observe.

The verdict

AI attribution is a small measurable core surrounded by a large unmeasurable halo. The temptation is to model the halo and report the total. The discipline is to measure the core precisely, label it a floor, and say that the halo exists and cannot be sized.

Do the unglamorous parts: stop stripping parameters, persist the referrer from first landing, classify with versioned rules, and keep the confidence tier on every row. Then report the floor with its counts visible. It is a smaller number than the one a less careful tool would show you, and it is the one you can still stand behind in six months.