We put GA4 and WISLR.ai side by side on the ecommerce sites we analyze, for August 2026, to find how much of the AI channel each tool can see, what those visitors bought, and which report to trust for the number. A customer asks an assistant which product to buy, taps the link in the answer, and orders from your store. Whether your analytics can see where that order came from depends on which app the customer used. Some apps tell GA4 where the visit came from. GA4 files the rest under Direct, next to people who typed your web address.
That leaves a store owner with three questions this piece answers. How many people is each assistant sending? How many of them buy? Which report should you trust for the number? The answers change by assistant, and a single average hides the change. The analysis shows GA4 and WISLR.ai within 3% of each other on ChatGPT. On Gemini, Claude, Perplexity, and Copilot, GA4 records between 1 in 3 and 1 in 22 of the visitors WISLR.ai counts. On orders, GA4 books about 60% of the ones that follow an AI visit within the hour. We analyze multiple sites, and the one shown here in full is the one where both tools can be compared like for like, host by host, for the same month.
The featured comparison covers a US ecommerce brand on our platform, selling across several country storefronts, for the 31 days of August 2026. GA4 is the standard implementation, tagged through Tag Manager, with the default channel grouping. WISLR.ai reads the same edge request logs, every request, and identifies assistant visits from what is in the request. Both tools were asked the same two questions: how many people arrived from each assistant, and what did they buy?
What the gap looks like on a $1 million month
GA4 attributes traffic from a script that runs in the visitor’s browser. WISLR.ai reads every request at the edge, before the browser runs anything. The two report the same $1 million month differently.
Your store does $1,000,000 a month in orders, and your average order is $100. You open GA4 and the AI Assistant channel shows $30,000 in revenue for the month. That is the number you would take into a budget meeting.
WISLR.ai on the same store shows about $50,000 from AI assistants, 1.6x what GA4 shows. The extra $20,000 is AI revenue GA4 put in its Direct row, because the visits that produced it arrived without a tag or a referrer. Over a year, that is $240,000 of AI revenue GA4 never showed you.
GA4 shows AI visitors buying at 1.1 per 100. WISLR.ai shows 1.7 per 100 on the same visitors, because it sees the orders GA4 did not record. On either report, AI visitors buy at two to four times the rate of your average visitor.
| A $1,000,000 month, illustrative | GA4 reports | WISLR.ai reports |
|---|---|---|
| AI assistant visitors | about 28,000 | about 30,000 |
| Visitors from Gemini, Claude, Perplexity, and Copilot | about 780 | about 3,700 |
| AI orders | 300 | about 480 |
| AI revenue | $30,000 | about $50,000 |
| AI conversion rate, per 100 visitors | 1.1 | 1.7 |
AI revenue GA4 files under Direct: about $20,000 a month.
The sections below show where each ratio comes from, starting with the visits.
Visits, assistant by assistant, for August 2026
Visits come first, because a tool can only capture a purchase to a visit it saw. Every count here is per 1,000 AI assistant visitors WISLR.ai counted, or a share, so no client can be identified; the ratios come from the unrounded counts.
1x means GA4 and WISLR.ai agree. Ratios are as measured from the unrounded counts. ChatGPT: about 895 per 1,000 in GA4 against about 875 in WISLR.ai.
ChatGPT is most of the traffic on both tools, so a blended total hides the gap. Summed across all five assistants, WISLR.ai counts about 8% more visitors than GA4. Assistant by assistant, GA4 and WISLR.ai are within 3% on ChatGPT, and GA4 records between 1 in 3 and 1 in 22 of the visitors WISLR.ai counts on the other four.
ChatGPT is left off: both tools count it the same, and its bar would dwarf the other four.
What those visitors bought, and which tool captures it
The gap in visits carries into the orders, and it carries in a second way as well. GA4 captures a purchase as AI only when the purchase happens in the session that arrived from the assistant. WISLR.ai captures it when the buyer’s IP address, the internet address the request came from, had an assistant visit shortly before. GA4’s rule and WISLR.ai’s rule are not the same count, and the figures below sit side by side for that reason.
One hour is the closest match to GA4's rule, which captures a purchase only when it happens in the session that arrived from the assistant. The 24-hour window is the widest one the data supports; matches past a day are mostly coincidence and are not shown.
How GA4 captures an AI order
- The purchase has to happen in the same session that arrived from the assistant.
- That session has to have carried a campaign tag or a referrer GA4 could read. An app visit with neither was filed under Direct on arrival, so any purchase in it is Direct too.
- A buyer who leaves and comes back later by typing the web address starts a new session, and the purchase is booked to Direct or Organic.
- Result, August 2026: about 1 purchase for every 100 AI-attributed storefront sessions, all but one of them from ChatGPT.
How WISLR.ai captures an AI order
- Any request from the buyer’s IP address that named an assistant counts as a touch, whether it named it through a tag, a referrer, or the app itself, and whether or not the page ran JavaScript.
- The order is matched to that touch if it came within a window: one hour, the closest match to GA4’s same-session rule, or 24 hours.
- Leaving and coming back does not break the match, because it rests on the IP address and the clock rather than on a session.
- Result, August 2026: about 1.6 orders for every one GA4 booked within the hour, and about 1.9 within 24 hours, which is about 2 orders per 100 AI-touched visitor-days. ChatGPT accounts for nearly all of them; Gemini, Claude, Perplexity, and Copilot are a handful each, too few to print as a rate. Past 24 hours the matches are mostly coincidence and are not printed.
Where the gap comes from
- GA4 recorded about 56% of the month’s paid orders across the whole site. That is general tracking loss (blockers, Safari’s storage limits, consent banners, purchase events that never fire), and AI orders are lost at about the same rate as everyone else’s: GA4 captured about 62% of the orders WISLR.ai matched to an AI visit within the hour.
- Return visits. A buyer who read the answer, tapped through, left, and came back within the hour by typing the web address starts a new GA4 session, booked to Direct or Organic. This adds a few points at most, and it has not been matched order by order.
- App visits. Orders whose only AI touch was a Gemini or Claude app visit were about 3% of AI-touched orders, and GA4 filed each of them under Direct.
- The average order is about the same on both tools, so the gap is orders GA4 never recorded rather than a different way of valuing them.
The order and revenue figures below are indexed to GA4’s count rather than printed as dollars, so no client can be identified.
| August 2026, storefront hosts, GA4 = 100 | GA4, AI Assistant channel | WISLR.ai, AI visit within 1 hour | WISLR.ai, within 24 hours |
|---|---|---|---|
| Orders | 100 | 160 | 190 |
| Revenue | 100 | 165 | 223 |
| Average order | 100 | about the same | about the same |
| Share of orders from ChatGPT | 98% | over 90% | over 90% |
| Orders whose only AI touch was an app visit | not visible | about 3% |
Put plainly, in the August 2026 data:
- For every 10 orders that follow an AI visit within the hour, GA4 captures about 6 in its AI channel. WISLR.ai captures all 10.
- GA4 undercounts visits from Gemini, Claude, Perplexity, and Copilot by 3x to 22x, and AI-touched orders by about 40%.
- AI-touched orders on the 24-hour window are under 2% of all orders and under 1% of revenue for the month, so the dollars behind the visits GA4 cannot see are small today.
- What those visits show is a channel growing at a rate GA4 cannot see, on storefronts where most AI-driven purchases still come through ChatGPT.
Conversion rate, on GA4 and on WISLR.ai
Out of every 100 visitors an AI assistant sends, GA4 counts about 1 who buys and WISLR.ai counts about 2. Behind that disagreement the two tools tell the same story: an AI visitor buys at two to three times the rate of the average store visitor, and a ChatGPT visitor buys at about three times the rate of a visitor from Gemini, Claude, Perplexity, or Copilot.
Gemini, Claude, Perplexity, and Copilot are pooled because each one alone had too few orders to state a rate. The vertical line is GA4's rate for every visitor to the store in the same month.
| Out of every 100 visitors, how many bought? August 2026 | GA4 says | WISLR.ai says |
|---|---|---|
| Visitors sent by any AI assistant | 1.1 | 2.0 |
| Visitors sent by ChatGPT | 1.1 | 2.2 |
| Visitors sent by Gemini, Claude, Perplexity, or Copilot, pooled | 0.6 | 1.0 |
Read across a row and the tools disagree on the number: GA4 sees about 1 buyer per 100 AI visitors, WISLR.ai about 2. Read down a column and they agree on the pattern: AI visitors buy at about twice the rate of the average visitor in GA4’s own figures, and ChatGPT visitors buy at about three times the rate of visitors from the other four assistants on either tool. Those four are pooled because each one alone produced too few orders in the month to state a rate. WISLR.ai’s own store-wide baseline is lower than GA4’s, because the logs count visitors who declined consent or run blockers, so the comparison to the average visitor uses GA4’s figure, the conservative one.
Why GA4’s number is lower:
- GA4 missed about 4 in 10 of the store’s orders in August, AI or not. We compared GA4’s purchase count with the store’s own order records. GA4 had recorded 56% of the paid orders.
- AI orders went missing at the same rate. Of the orders WISLR.ai tied to an AI visit within the hour, GA4 had recorded 62%. That is the same loss as the store overall, so GA4 is not undercounting AI buyers in particular. It is undercounting buyers.
- Return visits explain a little more. A buyer who reads an answer, taps through, leaves, and comes back within the hour by typing the store’s web address starts a new GA4 session, which GA4 files under Direct. This adds a few points at most. Earlier in the analysis it looked like the main cause; the order-capture figure above is the main cause.
- The two tools count the same AI visitors, so their rates can be compared. GA4’s count of AI sessions and WISLR.ai’s count of AI visitor-days are within about 2% of each other for the month, because both require clear evidence of an assistant before counting a visit. Their counts of all store visitors are not close, so the store-wide baselines cannot be compared the same way.
ChatGPT tags every link in its answers, and both tools read the tag
Both gaps have one cause, and ChatGPT is the place to see it, because ChatGPT is the assistant GA4 and WISLR.ai agree on. It appends utm_source=chatgpt.com to every link in its answers. The tag travels in the URL, so it survives whatever the app does with the referrer. GA4 reads it the way it reads any campaign link and attributes the session to chatgpt.com. WISLR.ai reads the same tag from the request line in the log. In August 2026 the two counts were under 3% apart: about 895 and about 875 of every 1,000 AI assistant visitors WISLR.ai counted that month.
The same product link, tapped in ChatGPT
Everything after the question mark is the tag. ChatGPT adds it to every link in its answers, and both GA4 and WISLR.ai read it on arrival.
The same product link, tapped in the Gemini app
No tag, and the app sends no referrer either. GA4 files this visit under Direct. WISLR.ai still counts it, because the request names the app that sent it.
The tag is doing more of the work than most people assume. In the data, three quarters of GA4’s ChatGPT sessions in August arrived with no referrer at all and were attributed on the campaign tag alone. Take the tag away and GA4 would lose most of its ChatGPT row too.
What a tagged link is, and why it decides what GA4 can see
That tag deserves a plain description, since the rest of the piece turns on it. A tagged link carries a note about where it came from inside the address itself. When ChatGPT shows you a link to a store, the address ends with a short marker, ?utm_source=chatgpt.com, the same way a link in an email newsletter ends with ?utm_source=newsletter and a link in a paid ad ends with ?utm_source=google&utm_medium=cpc. The marker rides along when you tap, and any analytics tool reads it on arrival. Marketers have tagged links this way for twenty years; ChatGPT does it automatically for every link it produces.
A link in an email newsletter
The marketer added this tag by hand. GA4 files the visit under Email.
A link in a paid search ad
The ad platform added it. GA4 files the visit under Paid Search.
A link in a ChatGPT answer
ChatGPT adds it to every link it produces. GA4 files the visit under AI Assistant.
An untagged link is the bare address. The only other clue to where a visitor came from is the referrer, the page they were on when they tapped, which the browser sends along with the request. A link tapped on a web page usually carries a referrer. A link tapped inside a phone or desktop app usually does not, because there is no web page to refer from and the app does not fill one in.
GA4 attributes a visit from those two clues and nothing else. A tagged link is attributed. An untagged link with a referrer is attributed to the referring site. An untagged link with no referrer is Direct.
Each assistant reaches a site from four or five surfaces, and only one of them tags the link
Where a link is tapped decides what travels with it, and assistants are used in more places than most people picture. Every assistant in the grid ships as a website, a phone app, and in most cases a desktop app, and several are being built into a browser or an operating system. The surface decides what your site receives.
Website
The assistant open in a browser tab, on any device. Links tapped here carry a referrer.
Phone app
The iPhone or Android app. Links open in the app's own browser and carry no referrer.
Desktop app
A separate program on a Mac or PC. Same behavior as the phone app: no referrer.
Built in
Part of a browser or the operating system, so a link can open with no app visible at all.
A surface is any place a person can use the assistant, and each one sends your site something different when a link is tapped.
| Where you can use it | ChatGPT | Gemini | Claude | Perplexity | Copilot |
|---|---|---|---|---|---|
| Website, in any browser | ✓ | ✓ | ✓ | ✓ | ✓ |
| Phone app, iPhone and Android | ✓ | ✓ | ✓ | ✓ | ✓ |
| Desktop app, Mac and Windows | ✓ | ✓ | ✓ | Windows | |
| Built into a browser | its own browser | Chrome | Chrome extension | its own browser | Edge sidebar |
| Built into the phone or the OS | Android assistant | Windows | |||
| Tags every link so GA4 can attribute it | ✓ every link | no | no | some links | some links |
Read the grid by row:
- Website and phone app. Where most people meet an assistant: a browser tab on any device, or the app on their phone.
- Desktop app. A separate program on a Mac or PC. Every assistant except Gemini ships one.
- Built into a browser or the OS. The newest surfaces, where a link can open without any app being visible. ChatGPT and Perplexity each ship their own browser, Gemini is built into Android and Chrome, Claude runs as a Chrome extension, and Copilot sits in Windows and in the Edge sidebar.
- Tags every link. The row that decides what GA4 sees.
What each assistant sends along with a tapped link:
- ChatGPT tags the links in its answers. GA4 and WISLR.ai count its visitors within 3% of each other.
- Perplexity and Copilot tag some links and not others.
- Gemini and Claude tag none, and their apps send no referrer, so a visit from either app reaches GA4 with nothing to attribute. In August, fewer than 1 in 100 visits from those two apps carried anything GA4 could attribute, and the apps are the surfaces most people use the assistants on.
Two details behind the grid:
- The web versions of all five send a referrer, which GA4 reads. The apps are where attribution breaks: a phone or desktop app opening a link sends no referrer, and without a campaign tag GA4 has nothing.
- The ChatGPT app also fetches a thumbnail for every link it shows in a chat, and that reaches your site as one image request with no page view. Both instruments count those as an impression, not a visit. Removing them changed nothing in the Gemini or Claude rows.
The surface rows are as each vendor publishes them in September 2026. The tag row is what the analysis measured, and only for the phone and desktop apps; the browser and OS integrations are too new to have a measured row.
GA4 attributes only what arrives in the URL or the referrer
GA4 decides where a visit came from with two clues: the campaign tag in the link, and the referrer the browser sends. A visit with neither goes into Direct, next to bookmarks and typed web addresses.
The Gemini and Claude apps send neither clue, so almost every visit from those apps lands in Direct. The small numbers GA4 does show for Gemini and Claude come from their web versions, which send a referrer. WISLR.ai counts the app visits because each request names the app that sent it, and the edge records that name. That is the whole gap: GA4 records about 1 in 3 Gemini visitors, 1 in 22 Claude visitors, and 1 in 4 from Perplexity and from Copilot, whose apps tag only some links.
The chart below follows those app visits week by week since June: the part of the AI channel GA4 files as Direct. The trend is the point rather than the size. Gemini and Claude are small next to ChatGPT today, and both lines run higher in August than in June.
Both lines start at 100 in June, so the chart shows growth rather than counts. The Gemini visits come from phones, and some of those visitors reach the cart.
The chart says visits rather than people for one reason. Most of the Claude line comes from the desktop app, and our team has not yet confirmed how much of it is a person browsing and how much is the app loading a link in the background. The Gemini line is people on their phones, and it makes up about half of the two lines together.
Close the gap on your site, and capture the 40% of AI orders GA4 misses
GA4 sees about 6 of every 10 orders that follow an AI visit. The other 4 are booked to Direct or Organic, and nothing inside GA4 can move them back, because the signal that would identify them is one GA4 drops on arrival.
WISLR.ai reads the logs your CDN already writes, needs no code on the page, and is free to start:
- Connect your site and the report is live in minutes: the assistants sending you visitors, the visits GA4 files under Direct, and the orders that followed an AI visit within the hour.
- Every count in this piece came from that report, on the same kind of logs.
- The results will differ by site and by month, as the assistants change how they tag links and what their apps send. A tag added by Gemini tomorrow would move its row toward the ChatGPT line, and the report would show it the same day.
Anyone with GA4 and their own server logs can also run the comparison by hand: one month of GA4 sessions by source against the same month of logs, counting the visitors whose requests carry an assistant’s tag, referrer, or app name. On the assistants that tag their links the two counts should sit within a few percent. The assistants that do not are where the gap opens, and the request is the only place their source survives. Try WISLR.ai free.