You cannot see the exact ChatGPT prompts behind a referral visit to your App Store page or your site: standard analytics have no direct, reliable native method for it. AI platforms do not pass the individual prompt or keyword in the referrer, the browser signal that says where a visit came from, the way Google Search Console passes queries.
An App Store owner put the gap in a public post about OpenSEO: "what questions users ask ChatGPT that bring them to my App Store page." The owner had tried OpenSEO, and it could not do that. The same post said ChatGPT brought $150 in sales that month, enough to cover a paid ChatGPT subscription.
The split for that owner, and for anyone counting the same line, is what stays hidden, which checks recover the visit, when a shared chat recovers the prompt, and when a simulated prompt is enough.
What You Cannot Learn About Which Questions Drove Visits
AI referral traffic is a visit that starts when someone clicks a link inside an assistant's answer, from ChatGPT, Claude, Perplexity, Gemini, or Copilot, as TapClicks defines it in its 2026 tracking guide.
TapClicks' guide says AI apps frequently strip referrer headers, so a URL typed from memory and a click inside a chat answer look the same in analytics. A referrer that still arrives, such as chatgpt.com, names the app the click left. The question the person typed stays in the chat.
Why Referral Traffic from ChatGPT Shows up as Direct
Most AI referral traffic is invisible in a default report.
TapClicks, citing Loamly's State of AI Traffic 2026, describes an analysis of over 446,000 website visits in which 70.6% of AI-driven visits arrived with no referrer header and were classified as Direct in Google Analytics 4, the property that sorts visits into channels. The other 29.4% carried a recognizable AI referrer. A standard dashboard therefore misattributes the majority.
TapClicks' baseline check is the Direct trendline. If Direct sessions are growing faster than brand campaign spend or offline brand awareness would explain, the guide treats some of that growth as AI referrals. The visit can still be marked AI-sourced.
How to Fix Referral Attribution When GA4 Says Direct
TapClicks' 2026 tracking guide lays out five checks, and each one recovers a different slice.
Build a dedicated AI-referral channel group in Google Analytics 4. A channel group is a rule that sorts visits into named buckets. Use regex, pattern-matching rules, on known AI referrer domains and on user-agent strings, the name a browser or app sends with a request. TapClicks names the current core set: chatgpt.com, claude.ai, perplexity.ai, gemini.google.com, and copilot.microsoft.com. Check that list against the referrers you are actually seeing before you publish the group, because the domains change. This catches visits that still present one of those domains or user-agent strings.
Tag every link you control. UTM parameters are small tags on a URL (source, medium, and campaign) so a click carries its context with it. Add the same tags to pricing pages, comparison pages, and product docs that an assistant might cite. TapClicks says this recovers the slice where someone clicks a tagged link. A link copied out of the answer and pasted into the browser arrives without those tags, and that copy-paste path is common with AI answers.
Segment Direct by landing page. AI-referred sessions tend to land on a specific deep page, a pricing page, a comparison page, or a particular how-to, and they arrive already knowing what they want. Filter Direct sessions by landing page and by behavior to flag likely AI-sourced visits the channel group missed. A cluster of those visits on a pricing comparison flags that page as the one they wanted.
Cross-check server logs. Server or CDN logs, the raw request records kept by your host or by a content delivery network, sit outside the JavaScript pageview tag. Some AI crawlers and some in-app browsers do not reliably run that tag, so the log is where that bot and referral activity shows up. For traffic that arrives with no referrer at all, the majority, TapClicks calls server or CDN log analysis the most reliable way to close the gap.
Report AI traffic as its own row, beside Direct and beside Organic (unpaid search), so its conversion can be read on its own.
When a Shared Chat Recovers the Prompt
The five checks mark the visit. The sentence the person typed shows up when they hand over the chat.
Someone running a demo funnel described that recovery in a public post. When a demo arrived referred by Claude or ChatGPT, they offered free credits if the visitor exported the chat that sent them, then read every prompt and every source in those chats. They kept asking people who signed up through an assistant, and used that pile of their own chats to decide which wording to repeat.
That recovers prompts from people who agree to share the chat. Ask a buyer who replies for the question they put to the assistant and for the answer that linked to the app. Everyone else stays in the channel group, the tags, the landing-page filter, and the logs.
Which Prompts Simulation Can Suggest
Simulated prompts are estimates inferred from test queries.
Visibility suites, including BrightEdge with AI Catalyst, Profound, Semrush, and Lantern, simulate thousands of conversational queries to estimate which prompts, topics, and categories trigger a citation or a brand mention for a domain. Treat the list as likely triggering prompts: what to test and what to write next. The visits you isolated are what already arrived.
Citation presence, whether a brand showed up in an answer at all, is a different record from which visit arrived. Write to the questions the landing pages are already answering, and tag the links you control.
The next draft needs the phrases Google suggests for that question and the results page Google is showing for it now. qequals returns the suggest strings for that question and, on a search call, the live results page for the same query.
Why a Small Share Still Drives Revenue
TapClicks' summary of Loamly's benchmarks reports AI referrals as a small share of total traffic, well under 1% of global traffic for most sites even after roughly 700% year-over-year growth.
TapClicks reports that dark AI traffic in the Loamly dataset, visits that arrived without a usable referrer, converted at 10.21% versus 2.46% for non-AI traffic, a 4.1x gap. TapClicks also reports a Microsoft Clarity study of 1,200 publisher sites in which AI-referred visitors signed up at 1.66% versus 0.15% for organic search, an 11x difference, with 27% lower bounce and 70% longer time on site. TapClicks gives the reason those rates run high: the visitor already did the research inside the chat before clicking. The App Store owner's $150 month is what a small count of that kind can look like on one product.
When Simulation Suffices and When Only Visits Count
Simulated prompts are enough when you are deciding what to publish or which page to tighten. If the suites point at the same comparison questions and your logs show AI visits landing on those pages, write to those questions and tag the links.
When you judge what paid, read the AI row against checkout or App Store sales, and keep that row on the report so the next small month has a channel beside the revenue.
