We measured AI-assistant referral traffic and orders across 1,000+ Shopify stores running Aimerce, on two independent data rails. In the year to September 2026, ChatGPT’s share of store pageviews grew 1.97x. It is still under one-tenth of one percent of traffic. It is also the fastest-growing referral source we can see, and most analytics tools under-count it by roughly 5x.
Two views of the same data. The pooled share is all ChatGPT pageviews divided by all pageviews, so the largest stores dominate it. The median store share gives every store equal weight, which is what you want when the set of stores changes month to month. Both roughly doubled, which is the point: the growth is not one large store joining the sample.
The dip through late 2025 and the first quarter of 2026 is worth naming rather than smoothing away. Both measures fall to a low in April 2026 and then climb steeply. We are not claiming a cause. Note only that the denominator swings with the retail calendar: peak-season campaign traffic in November and December grows the bottom of the fraction faster than the top. What survives that caveat is the same-month comparison, which is free of seasonality: 0.0564% in September 2025 against 0.1111% in September 2026 on the pooled measure, and 0.0379% against 0.0883% on the median.
The order rail is built from Shopify order records rather than pageviews, so it is an independent check on the traffic finding rather than a restatement of it. It agrees, including on timing: the step change lands in the second week of August 2026 on both rails.
Merchant penetration is the more useful number for anyone deciding whether this matters to them. It counts, for a single day, the share of stores that took at least one ChatGPT-attributed order. Because it is a daily figure it does not inflate simply because a month is longer.
ChatGPT-attributed orders also carry a higher median value than the rest. In US dollars, the median ChatGPT-attributed order was $75.46 in August 2026 against $60.25 for orders with no AI referral, about 25% higher. The gap has held in every month we measured, though it has narrowed since May. We report a median per currency rather than a total, because order totals across the fleet are recorded in many currencies and summing them would be meaningless.
Every other assistant combined is a rounding error today. The ordering differs slightly between the two rails: Gemini sends more traffic than Perplexity, while Perplexity converts more of what it sends. Both tails are small enough that we would not plan against the difference.
This is the part most likely to change what you do on Monday. ChatGPT strips the HTTP referrer on most outbound clicks, so those visits land in your analytics as direct traffic. What it does instead is append utm_source=chatgpt.com to the outbound link. Detection therefore has to be referrer OR utm_source. On our most recent month, referrer alone sees only 18.2% of ChatGPT pageviews.
The undercount is not a fixed discount you can correct for after the fact, which is what makes it dangerous. It has grown: the share of ChatGPT visits arriving with no referrer went from 68.1% in September 2025 to 81.8% in September 2026. So a referrer-only measurement does not simply read low, it reads flat while the real number nearly doubles.
The order rail shows the same shape independently: 73% of ChatGPT-attributed orders across the window had no referring site and were identifiable only from the landing URL. One practical note for anyone reproducing this: the parameter is spelled both chatgpt.com and chatgpt, and a substring match on the referring domain alone will also sweep in merchants’ own internal link tagging, so the two signals need to be read together rather than either one trusted by itself.
Traffic comes from first-party pageview data collected on each store’s own domain, one dataset per store, queried store by store and then aggregated. Every store dataset in the fleet was queried, 1,000+ of them; every query returned successfully, so no store was dropped or estimated. A little under two-thirds hold no pageviews in the window at all and contribute nothing. Orders come from Shopify order records for the same fleet, read directly rather than inferred from browser events. The two rails share no code path and no storage, which is why we treat their agreement on timing as meaningful.
A pageview is attributed to an AI assistant when either its referring domain or its utm_source names one. An order is attributed when either its referring site or its landing URL does. Referrer alone would miss 81.8% of ChatGPT pageviews, as Figure 5 shows. For assistants other than ChatGPT and Perplexity we match on the referring domain and on an explicit utm_source only, never on a bare substring of the landing path, because a product or collection named after an assistant would otherwise be counted as a referral.
A separate 90-day census of every referrer and utm_source spelling we matched puts the referrer-only miss rate at 81.9%, which is the figure quoted above arrived at independently. It also shows two things the rule counts that a stricter reading would not. About 4.2% of matched ChatGPT pageviews carry an ad-style source such as chatgpt-ads, which is paid placement rather than an organic assistant referral. A further 3.1% arrive with the merchant’s own domain as the referrer, which is a ChatGPT-originated visit whose tag has survived a click within the site. Neither is large enough to change the trend, and we have left both in rather than apply a judgement call that would be hard for anyone else to reproduce.
A store counts as active in a month if it recorded at least 3,000 pageviews that month. Between 210 and 300 stores clear that bar in any given month. The median store share uses a higher bar of 20,000 pageviews in the month, so that a single visit cannot move a small store’s percentage by a large amount. Stores below the threshold are kept in the underlying data and excluded only from the aggregates.
The traffic rail covers Sep 25 to Sep 26, 13 months. The order rail covers 2026-05-19 to 2026-09-16, every day in that range with no sampling. It starts in May 2026 because that is where the order record begins, not because nothing happened before it. September 2026 is partial on both rails. All headline figures are ratios, which is what makes a partial month safe to plot next to a complete one; absolute counts are reported only for complete periods.
This is last-click attribution on a link the assistant generated. A shopper who asks ChatGPT for a recommendation and then types the brand name into a search engine appears here as search traffic, so the figures are a floor on AI’s influence and not an estimate of it. We also do not measure in-assistant checkout, which does not touch the store’s own domain at all. Finally, these are stores running Aimerce: direct-to-consumer Shopify merchants weighted toward higher order volume, not a random sample of all ecommerce.
Every figure on this page is a fleet-level aggregate. No individual store is named, described or separable from the totals, and store counts are rounded.
| Month | Pooled share | Median store share | Referrer-only share | Active stores | Stores with ChatGPT traffic |
|---|---|---|---|---|---|
| Sep 25 | 0.0564% | 0.0379% | 0.018% | ~210 | 95.3% |
| Oct 25 | 0.051% | 0.0335% | 0.0167% | ~230 | 92.1% |
| Nov 25 | 0.0401% | 0.0249% | 0.0126% | ~230 | 94.8% |
| Dec 25 | 0.035% | 0.0204% | 0.0108% | ~230 | 93.9% |
| Jan 26 | 0.0486% | 0.031% | 0.0138% | ~240 | 93.4% |
| Feb 26 | 0.0348% | 0.0281% | 0.009% | ~240 | 90.3% |
| Mar 26 | 0.0343% | 0.0247% | 0.0086% | ~250 | 92.3% |
| Apr 26 | 0.0334% | 0.0278% | 0.0077% | ~260 | 93.4% |
| May 26 | 0.0524% | 0.0372% | 0.0113% | ~270 | 95.1% |
| Jun 26 | 0.0544% | 0.0307% | 0.0108% | ~280 | 95.4% |
| Jul 26 | 0.062% | 0.0313% | 0.011% | ~290 | 92.8% |
| Aug 26 | 0.0852% | 0.0635% | 0.0155% | ~300 | 96.3% |
| Sep 26(1-17 only) | 0.1111% | 0.0883% | 0.0202% | ~280 | 96% |
| Month | Days | All orders | ChatGPT orders | Share | Stores/day | Median order value (USD) |
|---|---|---|---|---|---|---|
| May 26(partial) | 13 | 1,678,430 | 1,123 | 0.0669% | 8.25% | $88.21 vs $64.93 |
| Jun 26 | 30 | 4,238,959 | 2,242 | 0.0529% | 7.1% | $84.27 vs $61.36 |
| Jul 26 | 31 | 4,846,174 | 2,742 | 0.0566% | 7.73% | $73.11 vs $61.1 |
| Aug 26 | 31 | 5,101,705 | 4,417 | 0.0866% | 11.54% | $75.46 vs $60.25 |
| Sep 26(partial) | 16 | 3,413,337 | 2,722 | 0.0797% | 12.7% | $71 vs $60.25 |
Median order value compares ChatGPT-attributed orders against orders with no AI referral, in US dollars only. Store counts are rounded to the nearest ten.
Two things would change the picture. The first is whether the August 2026 step holds or decays; one month of data cannot tell you, and we will republish this page as the window extends. The second is in-assistant checkout, which would move the transaction off the store’s domain entirely and make referral-based measurement blind to it. Both rails here depend on the shopper landing on the merchant’s own site. That assumption is the one most likely to break.
If you want to run this measurement on your own store, the detection rule is the whole trick: match the referring domain OR the utm_source, and treat the two as one signal. Anything less and you will conclude that ChatGPT sends you nothing.