There is no Search Console for AI prompts, because ChatGPT, Claude, and Perplexity never send you the words a person typed. When someone clicks a link inside an AI answer, your analytics receives the referrer, say chatgpt.com, and nothing else. The prompt that produced the recommendation stays sealed inside the chat. The single most valuable signal in twenty years of search marketing, the actual query, has quietly gone dark.
That is not a tooling gap someone will patch next quarter. It is a structural property of how these products work, and it changes what "tracking" can even mean. Below is why the prompt is invisible, why no traditional SEO tool can recover it, and the three-layer stack that gets you as close to a Search Console for AI as the ecosystem currently allows.
Terms in this piece
New to the acronyms? Each term below links to our field guide with plain-English definitions and how YouGotRanked scores them.
Search engines tell you the query. AI assistants do not.
For two decades, the deal was simple. Someone searched, Google recorded the query, and Search Console handed it back to you: this exact phrase earned this many impressions, at this position, with this click-through rate. You could see demand, rank against it, and measure the result. Keyword research existed because the keywords were observable.
AI assistants broke that deal. The conversation is private, the retrieval happens server side, and the link you get cited in carries none of the prompt. So the question every marketer now asks, "which prompts are sending us traffic from ChatGPT?", has no honest answer from your analytics. The words never left the chat window.
Any vendor that claims to show you the real prompts behind your AI referral traffic is inferring them, not observing them. It is worth being precise about this, because the honest version of the answer is more useful than the marketing version.
Why the prompt is invisible: three walls
The prompt does not leak for three separate reasons, and each one blocks a different workaround.
Framework
Three walls between you and the prompt
1. The referrer carries the domain, not the query. A link clicked inside an AI answer passes at most the source host, such as chatgpt.com or perplexity.ai. There is no query string with the prompt, because the assistant writes your canonical URL into the answer, unchanged.
2. The apps often pass no referrer at all. Traffic from the ChatGPT and Gemini mobile and desktop apps frequently arrives with no referrer, so it lands in analytics as Direct. Your AI referral count is therefore an undercount, not a ceiling.
3. You cannot tag a citation you did not create. UTM tracking works because you add parameters to links you control. When an assistant cites your page, it uses your plain URL, so there is nothing to append a utm_source to. The tagging playbook that solved campaign attribution does not apply here.
Put together, these mean the prompt is not merely hard to get. It is not transmitted at all, and the usual recovery tricks, referrer parsing and UTM tagging, are structurally unavailable.
What each surface actually reveals
"AI search" is not one thing, and the surfaces differ sharply in what they expose. Treating them as identical is how teams end up chasing data that does not exist.
How much each surface reveals about the query
Bars show how much query-level insight each surface hands back. Only Google surfaces the real terms; the conversational assistants reveal almost nothing beyond that a session came from them.
Source: YouGotRanked analysis
The practical breakdown:
- Google AI Overview. This is the exception worth knowing. AI Overviews are still Google Search, so the queries that trigger them appear in Search Console like any other query. If you want real terms behind an AI surface today, filter your Search Console data. You already have it.
- Bing and Copilot. Copilot runs on Bing, and Bing Webmaster Tools surfaces some query data, so there is partial visibility here.
- ChatGPT, Claude, Perplexity, standalone Gemini. No prompt. At best a referrer, often not even that. This is the majority of the conversational surface, and it is close to opaque.
The uncomfortable implication: the biggest, fastest-growing part of AI discovery is precisely the part where the query is unknowable.
Why traditional SEO tools cannot close this gap
The instinct is to reach for SEMrush, Ahrefs, Moz, or a rank tracker. They cannot help here, and it is worth understanding why, because it is not a feature they forgot to build.
Traditional SEO tools are built around a keyword database and a crawler. They tell you what ranks in Google for a known list of keywords. That entire model assumes the query is observable and the ranking is a public, crawlable list. AI answers break both assumptions: the prompt is private, and the answer is generated on the fly from sources the tool cannot see being chosen. A keyword rank tracker pointed at ChatGPT has nothing to crawl and no query to track. Adding an "AI" tab to a keyword tool does not change the underlying fact that the thing it measures, keyword rankings, is not what an AI answer is made of. This is why purpose-built AI brand monitoring software samples the answers directly instead of trying to reconstruct them from rankings.
This is why the honest solution is a different instrument, not a new tab on an old one. Measuring AI visibility means testing prompts, sampling answers for stability, and reading which sources get cited, none of which a rankings-and-backlinks platform was designed to do.
The closest thing to a Search Console for AI: a three-layer stack
You cannot get the real prompts, so the goal shifts from observing demand to triangulating it. Three layers, each doing a different job, get you as close as the ecosystem allows.
Framework
The three-layer AI visibility stack
1. Search Console, for real Google queries. This is your only source of true query data, and it now includes AI Overview appearances. Use it for the terms Google can still see.
2. GA4 landing pages by AI source, for topic-level intent. You cannot see the prompt, but you can see which page the AI sent the visitor to. If ChatGPT drops someone on your Profound comparison, the prompt was about comparing Profound. The landing page is your strongest proxy for what was asked.
3. Prompt testing, for the prompts you can never observe. Since real prompts are sealed, you run the prompts yourself: a broad set of buyer-intent questions across every engine, measuring which ones name you and which name competitors. You sample the prompt space instead of waiting for it to report.
Layer three is the substitute for the missing console. Rather than passively receiving queries, you actively probe the space of questions your buyers ask and measure your presence in the answers. It is the inverse of Search Console: instead of "here are the terms that found you," it is "here are the terms you show up for, tested on purpose." You can run this yourself with a free AI visibility checker before committing to a program.
To read the GA4 layer directly, and this is how you track ChatGPT referral traffic in practice: in GA4, open Explore, build a free-form report with Session source and Landing page as dimensions and Sessions as the metric, then filter the source to chatgpt, perplexity, gemini, claude, or copilot. That gives you the topics AI assistants are actually sending traffic to, which is the closest observed signal to intent you will get from the conversational surfaces. What it will not give you is the prompt itself, so there is still no way to see the prompts users type in ChatGPT, only the pages those prompts sent them to.
How YouGotRanked gets you closest
If Search Console is the tool that told you what people typed into Google, YouGotRanked is the closest thing that exists for what people ask AI. It cannot read a private conversation, because nothing can, but it does the next best thing, and it does it completely: it reconstructs the demand by testing the prompts your buyers actually use, across every engine, then shows exactly where you stand in the answers. Traditional SEO tools cover none of this. Most AI monitors run some prompts and stop at a dashboard. YouGotRanked holds all three layers natively: it connects your Search Console and GA4 directly, runs the prompt testing itself, and puts the whole stack on one timeline, per engine and against your competitors, with a plan you can act on.
Prompt testing is the workable substitute, and doing it well is harder than sending one prompt to one model. This is where the approach matters more than the label.
Framework
What complete prompt testing requires
Breadth of prompts. One prompt is anecdote. We run 200+ real buyer-intent prompts across organic, brand-anchored, and competitive clusters, so the picture reflects how your category is actually asked about, not a single lucky phrasing.
Multiple samples per prompt. LLM answers vary run to run, so a single query is noise. We sample each prompt repeatedly, which is the only way the number is stable enough to trust and to act on.
Every engine, scored separately. ChatGPT, Claude, Gemini, and Google AI Overview source answers differently, so a blended score hides the truth. We score each provider on its own.
Citations, not just mentions. We read which sources each engine quotes, so you learn not only whether you appear but who is winning the answer and where the citations come from.
Tied back to real traffic. We connect Google Search Console and GA4, so the prompt-testing layer sits next to the real query and referral data, and visibility can be expressed in paid-media dollars rather than a mention rate alone.
That last point is the one competitors and traditional tools skip. The three layers only become a Search Console substitute when they sit together: real Google queries from GSC, real AI referral topics from GA4, and the tested prompt space from the brand monitor, on one screen, for one brand, against your competitors. Prompt testing on its own is a data feed. Combined with the observable layers and a prioritised action plan, it becomes a decision tool.
What that looks like inside the platform
The exhibits below are the actual charts the YouGotRanked dashboard draws, rendered here with illustrative sample data so a customer's numbers stay private. This is the visual trail the platform builds: the tested prompt space, then your real Search Console and GA4 data on the same timeline.
Prompt visibility, scored per engine
Share of tested prompts where the brand is named, scored separately for each engine. A single blended number would hide the spread, which is where the real story usually is.
Source: Illustrative sample, YouGotRanked dashboard
Where you appear, by prompt cluster
Presence across the prompt clusters we test. Brands almost always win their own name and lose the organic buying questions, which is exactly where new demand is decided.
Source: Illustrative sample, YouGotRanked dashboard
Google Search performance, from your connected Search Console
Layer one of the stack, live on the dashboard: daily clicks and impressions from your connected Search Console, each on its own labelled axis. On the live chart, hovering a day also shows CTR and average position, and executed fixes are marked on the timeline.
Source: Illustrative sample, YouGotRanked dashboard (GSC integration)
Site sessions and AI referrals, from your connected GA4
Layer two: all sessions with the AI-referral slice broken out, from your connected GA4. The orange line is traffic arriving from ChatGPT, Perplexity, Gemini, and Copilot, the closest observable signal to AI demand.
Source: Illustrative sample, YouGotRanked dashboard (GA4 integration)
Top AI referrers, sessions by source
Which assistants actually send the traffic. Paired with the tested prompts above, you see both what you appear in and what it drives, the two halves of the stack on one screen.
Source: Illustrative sample, YouGotRanked dashboard (GA4 integration)
Re-scanning on a cadence turns each of these into a trend rather than a snapshot, with executed fixes marked against the movement, so you can read a rising line against the work that caused it.
AI visibility, scan over scan
The dashboard's headline trend: GEO score, AI visibility, and share of voice, one point per scan on a 0 to 100 axis. On the live chart, executed fixes are marked on this timeline so an uptick can be read against the work that produced it.
Source: Illustrative sample, YouGotRanked dashboard
That is the reporting loop a keyword console gave you for Google, rebuilt for a world where the keyword itself is unobservable.
To restate the stack plainly, because this is the answer to the question that starts most of these conversations:
- GSC gives you real queries, on Google surfaces only.
- GA4 landing pages by AI source give you which topics drive real AI referral clicks.
- Prompt testing gives you which prompts you appear in, since the real prompts are unknowable.
YouGotRanked is not the third layer with the other two left as homework. It holds all three: connect your Search Console and GA4 once, and the dashboard puts your real queries, your AI-referral traffic, and the tested prompt space on one timeline, per engine, against your competitors. That combination is as close to a Search Console for AI as the ecosystem allows, and no rankings tool can assemble it.
See which prompts name you, and which name competitors
Run a free analysis across ChatGPT, Claude, Gemini, and Google AI Overview. We test the prompt space you can never observe, score you per engine, and show who gets cited.
Analyze My Brand for Free →
FAQ
Can I see the prompts users type in ChatGPT to find my site? No. ChatGPT does not pass the prompt to your analytics. A click from a ChatGPT answer arrives with, at most, the referrer chatgpt.com, and often no referrer at all from the apps. The prompt stays inside the conversation, so there is no way to observe the real words a user typed. The workable substitute is to test prompts yourself across the engines and measure which ones surface your brand.
Is there a Search Console for ChatGPT or Perplexity? Not in the sense of real user queries. Google Search Console shows the queries behind Google Search and Google AI Overview appearances, and Bing Webmaster Tools shows some Copilot-related query data. ChatGPT, Claude, and standalone Perplexity and Gemini reveal no prompt data at all, so the closest equivalent is proactive prompt testing combined with GA4 landing-page analysis.
How do I track ChatGPT referral traffic in GA4? Filter your sessions by source for AI hosts such as chatgpt.com, perplexity.ai, gemini.google.com, and copilot.microsoft.com, and pair that with the landing page to infer topic-level intent. Note two limits: app traffic often shows up as Direct, so you undercount, and you cannot add UTM parameters to a citation the assistant writes, so tracking ChatGPT referral traffic relies on referrer detection rather than tagging.
Why can't SEMrush or Ahrefs show me AI prompts? Because they are built around a keyword database and a crawler, which assume the query is observable and the ranking is a public list. AI answers are generated privately from sources chosen on the fly, so there is no query to track and nothing to crawl. Measuring AI visibility requires testing prompts and reading citations, which is a different instrument than a rankings tool.
What is the closest thing to keyword tracking for AI? A three-layer stack: Search Console for real Google queries, GA4 landing pages by AI source for the topics driving real AI referral clicks, and prompt testing for the prompts you can never observe. YouGotRanked assembles all three for one brand, scored per engine and benchmarked against competitors.
Go deeper: What Is Generative Engine Optimization? · What Is Answer Engine Optimization? · 4 Signals That Make Content Citable · Compare AI visibility tools