Crawler hits, conversation citations, and referral sessions are three different things. Most AI visibility reports treat them as one. Here’s the framework that separates them.
TL;DR
When someone says their AI referral traffic tripled, that number could mean crawler volume, conversation citations, or actual human sessions — and those three figures can point in completely opposite directions. The four AI visibility states (Invisible / Stage 0 / Mentioned / Cited) assign each metric to the correct stage of the funnel, so you know whether you’re measuring a precondition or an outcome.
Only one state drives referral sessions. The others are diagnostic signals — useful for understanding what’s blocking you, not for reporting success.
Crawler hits, conversation citations, and referral sessions are three different metrics that respond to different interventions on different timescales — reporting them in the same breath, as most AI visibility reports do, makes the numbers impossible to act on. The four-state visibility model separates them and maps each to a specific GEO Stack layer.
The AI Referral Traffic Numbers That Don’t Mean the Same Thing
A post circulating in April 2026 reported three figures in the same breath: 32,479 ChatGPT conversations had cited a product’s pages in the past 30 days, up 381%. The site had logged 917,000 AI crawler hits over the same period. And AI referral traffic had tripled in recent weeks.
These are impressive-sounding numbers. They are also three completely different measurements, at three different stages of the AI visibility funnel, with three different implications for what to do next.
The 917,000 crawler hits are a server log figure — bots visiting the site to index content. The 32,479 conversations are a platform-level citation count — queries in which ChatGPT retrieved and linked the domain. The referral traffic is a GA4 session figure — humans who followed a citation link to the site. The first is a precondition. The second is an outcome measure. The third is the downstream commercial result of the second.
Conflating these three figures doesn’t just muddy reporting — it misdirects intervention.
If crawler hits are high but citation rate is low, the bottleneck is extractability and entity clarity, not crawl access. Adding more content to increase crawler volume will not move the needle.
The GEO Lab uses a four-state visibility model to formally separate these signals. Each state corresponds to a distinct position in the funnel, with a distinct diagnostic implication and a distinct intervention path.
The Four AI Visibility States
Every page, on every AI platform, sits in one of four states at any given time. The states are mutually exclusive and ordered — you cannot be Cited without passing through the earlier stages first.
Establishing which state a page is in before any intervention is the purpose of Month 0 baseline measurement — without that starting point, the state transition you observe cannot be attributed to a specific content or structural change.
The AI system has no access to the page — either because the crawler has not indexed it, the domain lacks sufficient authority for the platform’s retrieval threshold, or the content is blocked. No retrieval means no possibility of citation.
The crawler has visited the page and the domain appears in the platform’s index, but the content is not being selected for inclusion in responses. The page exists to the system — it is not participating in answers. This is the most common state for new or low-authority domains.
The AI response refers to the site or entity by name in the answer text, but does not include a source link. No session is driven. In GEO Lab’s E002 experiment, Gemini mentioned thegeolab.net in 21.2% of responses and cited it zero times across the same query set. Mention rate is not citation rate.
The AI response includes the domain as a clickable attributed source. A human can follow the link. This is the only state that produces referral sessions in GA4. Everything before this is diagnostic infrastructure — not a traffic outcome.
The critical distinction is between Stage 0 and Mentioned. Both produce zero referral traffic. But they have different causes and require different responses. Stage 0 indicates a retrieval problem — the content is indexed but not selected. Mentioned indicates a citation-conversion problem — the content is being retrieved and used, but the platform is not attributing it with a link.
What Each Metric Actually Measures
Mapping common AI visibility metrics to the four-state model makes the diagnostic logic concrete.
| Metric | What it measures | Visibility state |
|---|---|---|
| AI crawler hits (server log) | Whether platform bots are accessing your pages | Invisible → Stage 0 transition |
| Citation rate (query checks) | Proportion of relevant queries where the domain is linked in the AI response | Stage 0 → Cited transition |
| Mention rate (query checks) | Proportion of relevant queries where the domain is named but not linked | Mentioned state — not a traffic signal |
| Conversation citations (platform data) | Total query instances in which the platform cited the domain over a period | Cited state — volume measure |
| AI referral sessions (GA4) | Human sessions originating from an AI platform citation link | Cited state — downstream outcome |
The relationship between conversation citations and referral sessions is not 1:1 and not fixed. A citation in a query that is asked by thousands of users in a day drives more sessions than the same citation in a query asked by three. Reporting raw citation count without query volume context produces a number that cannot be compared across time periods or topics.
Where This Model Comes From
The four-state framework was developed during E016 — The GEO Lab’s noise floor measurement experiment, which ran five consecutive days of citation checks against the same 30-query set with no content changes. The purpose was to establish baseline citation rate variance independent of any content variable.
The E016 noise floor data also underpins the ten citation measurement variables post — the same experiment that produced the four-state model also quantified how much each variable contributes to citation rate variance.
On Day 1, the data immediately separated into two distinct populations. Tier 1 queries — those about proprietary GEO Lab concepts like the GEO Stack, Retrieval Probability, and Extractability — produced citation rates between 30% and 60%. Tier 2 queries — category and commercial terms like “generative engine optimisation tools” or “AI search optimisation” — produced citation rates of 0% across all five days.
The Tier 2 result is not a content quality failure. It is a domain authority failure at the retrieval stage. The content exists. The crawlers visit it. The platform does not select it for competitive queries because the domain does not yet have the authority signals required to compete in that query space. That is Stage 0 visibility — not Invisible (crawlers are active), not Cited (no link is produced), just retrieved and set aside.
The noise floor experiment is what makes the four-state model falsifiable rather than theoretical.
Without baseline variance data, any change in citation rate could be noise. E016 established that Tier 1 citation rate holds stable across days (±8% variance), while Tier 2 holds stable at 0%. That makes interventions on Tier 2 content testable — if a change produces a non-zero result on a Tier 2 query, that is a signal worth examining.
Applying the Model to Reported Results
Return to the original figures: 917k crawler hits, 32,479 conversation citations, 3x referral traffic growth.
The four-state model connects to the GEO Stack layers: crawled but not cited is a Layer 2 extractability failure; cited but not driving referral traffic is a Layer 3 authority or query-match issue that content changes alone cannot fix.
Mapped to the four-state model, these read differently. The 917k crawler hits confirm the domain is not Invisible to the platforms generating those logs — GPTBot, PerplexityBot, and similar. That is the Invisible-to-Stage-0 transition confirmed. It says nothing about whether those crawled pages are being cited.
The 32,479 conversation citations confirm the domain has moved past Stage 0 on ChatGPT for some query set. The 381% growth figure is interesting but uninterpretable without knowing the query distribution — whether the growth came from a single high-volume query, a broad set of lower-volume queries, or a change in the platform’s retrieval behaviour for a topic category.
The 3x referral traffic growth is the only downstream outcome measure. If this figure is tracked to specific citation sources — which pages, which query types, which platforms — it becomes actionable. Without that breakdown, it is a success metric with no replication path.
None of this makes the numbers wrong. It makes them incomplete as a reporting framework. The missing layer is always the same: which visibility state is each metric measuring, and what is the bottleneck between the current state and the next one?
The Measurement Protocol at The GEO Lab
The GEO Lab separates these signals into three distinct measurement tracks, each running on a different cadence and producing a different type of output.
The citation rate track uses the standard probe query set — ten query types across Tier 1 and Tier 2, run weekly during active experiments and bi-weekly during baseline monitoring.
Crawler access (weekly, server log review): Which AI crawler bots hit which pages, at what frequency, and whether any pages are missing from the crawler logs despite being indexed by Googlebot. A page present in GSC but absent from AI crawler logs is Invisible to that platform regardless of its traditional search performance. The AI Visibility Diagnostics Console surfaces this discrepancy as a crawl parity failure.
Citation rate (per experiment, query-based checks): 10 queries per experiment run, across Perplexity, ChatGPT, and Google AI Overviews, with a fixed query set and documented noise floor from E016. Citation rate is recorded as a percentage per platform, not blended. A blended figure hides platform-specific behaviour — Perplexity and ChatGPT have different retrieval mechanisms and different domain authority thresholds.
Referral sessions (monthly, GA4 source/medium breakdown): AI referral traffic tracked by platform and landing page. The landing page dimension is the link between the citation rate measurement and the session outcome — it identifies which cited pages are converting crawler access into human visits.
Running all three in parallel is what makes it possible to isolate the bottleneck. High crawler hits and low citation rate: extractability or entity clarity problem. High citation rate and low referral sessions: query volume or competitive displacement problem. High referral sessions without a corresponding citation rate increase: a platform behaviour change worth investigating.
Frequently Asked Questions
What is the difference between AI crawler hits and AI referral traffic?
AI crawler hits are server log entries from bots like GPTBot, PerplexityBot, and ClaudeBot visiting your pages to index content. AI referral traffic is human sessions arriving at your site from an AI platform after a citation. A page can receive thousands of crawler hits and zero referral sessions — crawling is a precondition for citation, not a guarantee of it.
What is the difference between a citation and a mention in AI search?
A citation is a named source link in the AI response — the platform retrieved the page and attributed the content with a clickable link. A mention is when the AI refers to a site or entity by name without linking. Mentions do not drive referral sessions. In GEO Lab experiments, Gemini mentioned thegeolab.net in 21.2% of responses but cited it zero times across the same query set.
What are the four AI visibility states?
Invisible — the platform does not retrieve the domain at all. Stage 0 — the platform retrieves the domain but does not include it in the answer. Mentioned — the AI names the domain in its response without linking. Cited — the AI includes a named source link. Only the Cited state drives referral sessions. The earlier states are diagnostic — they identify what is blocking the transition to citation.
Why does conflating AI visibility metrics lead to bad decisions?
Each metric sits at a different stage of the AI visibility funnel. Optimising for crawler volume when the bottleneck is citation rate wastes effort. Celebrating referral session growth without knowing which queries drove it makes the result non-repeatable. Treating a mention as a citation overstates actual traffic impact. The four-state model assigns each metric to its correct stage so interventions target the actual constraint.