AI Traffic Decoder:
attribute the visits AI answers send you

Most AI visits reach your analytics stripped of the context that would tell you where they came from.

When someone follows a link inside the ChatGPT app, the visit often lands with no referrer at all and files itself as direct traffic. Perplexity and Google's AI surfaces usually pass a referring domain; in-app browsers, copied links and some mobile flows usually do not. So the raw view in GA4 undercounts AI, and the missing share hides inside direct and unassigned traffic. Decoding does not invent traffic that was not there; it moves visits that already convert into a channel you can report, compare against paid and organic, and defend in a budget conversation.

Decoding runs as a pipeline that reads two data planes — your analytics and your raw request logs — then reconciles them.

First, a source-and-medium rule set folds the known assistant hosts into one AI channel in GA4. Those hosts — chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and their kin — otherwise scatter across “referral” and “organic” until you group them. Second, server-side referrer logs catch the same hits at the edge, before a client-side tag can drop them on a slow or blocked load. Third, the decoder classifies every request by user agent into human, retrieval agent, training crawler and other bot, so machine traffic never sits inside the human numbers. Fourth, it recovers AI-assisted human sessions that lost their referrer, using the signals that survive the hand-off: an intact referrer where one exists, tracking parameters some assistants append, landing-page patterns only AI answers link to, and timing that follows a retrieval fetch of the same URL. Fifth, it persists the classified log from day one, because edge providers keep raw request analytics only briefly — commonly around a day at lower tiers — after which the history is gone. Finally, it joins three timelines per URL — when retrieval agents fetched the page, when it appeared as a cited source, and when AI-assisted sessions landed on it — reported monthly next to your other channels. Change a page, then watch whether its referral pattern shifts.

What the decoder measures for each engine

Each engine passes referrers its own way, so each gets its own source rules and its own reading rather than a blended average.

ChatGPT links from the app frequently carry no referrer and file as direct, so recovery leans on log matching and landing-page patterns. Perplexity usually passes a perplexity.ai referrer and attaches numbered citations to almost every answer, which makes it the most self-labelling engine of the set. Gemini generally passes a gemini.google.com referrer that must be kept apart from ordinary Google organic. Copilot passes a copilot.microsoft.com referrer and skews toward work-context queries. Google AI Mode and AI Overviews sit on the google.com search surface, so their referrals can look like plain organic and need landing-pattern rules. Grok arrives from the X context, and Meta AI runs mostly in-app, where a referrer is frequently absent. Coverage is weighted to where your own audience asks, so a B2B site heavy on ChatGPT and Copilot reads differently from a consumer brand heavy on Google AI Mode. The aim is per-engine truth, not a single number that hides which engine is doing the work.

The metrics it reports, and what each one means

Every figure on the decoder means one specific thing, and the definition matters more than the total.

The headline reading is AI-assisted sessions — human visits recovered from an AI answer and moved out of Direct into a named AI channel — split by engine as per-engine referral share so you see which assistants send people rather than a lump sum. Direct-recovery share sizes the problem you were previously blind to, and landing-page distribution shows which pages AI visitors actually arrive on, rarely the set marketing expects. On the machine side, fetch rate by page — how often retrieval agents pull a URL — is the leading indicator: a pricing page an engine starts fetching daily is about to be cited on pricing questions. Outcomes use assisted conversions, reported apart from last-click so the number is comparable to every other channel, and every reading carries a confidence tier (all defined in the glossary below).

An engine fetching your page to answer a question now is not the same event as a corpus being built for later, and neither is a human click.

Two families of machine traffic reach your logs, and merging them misleads. Retrieval agents — OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User and ClaudeBot — fetch a page so an engine can answer a live question or ground a citation; a rising fetch rate here is a near-term signal that a page is entering answers. Training crawlers — GPTBot, CCBot, Google-Extended and Applebot-Extended — fetch to build or refresh a model's corpus, a slower and different motion. Neither is a person: counting either as a session inflates page views and depresses conversion rate on exactly the documentation, pricing and comparison pages agents favour most. The decoder reads user-agent and IP patterns to file all machine activity in its own columns, keeps the two agent families distinct, and leaves the human referral numbers clean — so crawler demand becomes a forward-looking metric rather than noise in the visitor count.

No tool can prove that one sentence in one answer caused one click, and Asva does not pretend otherwise.

Referrers almost never carry the prompt a reader typed, so exact one-to-one attribution is not available on this data. What decoding can do is correlate signals you can check and label how firmly each session's origin is known. A referrer-confirmed session carried a known assistant host in the referrer; a parameter session arrived with a tracking parameter an assistant appended; a pattern-inferred session had no referrer but matched a landing page and timing that line up with a retrieval fetch of the same URL. Reporting the tier alongside the count is what makes the number usable: leadership can weight a referrer-confirmed figure differently from an inferred one, and a report that states its own limits is one a team can act on. The honest promise is a strong, testable pattern with its confidence shown — not a false claim of precision.

Who uses it, and the question each role brings

The same readings answer a different question depending on who is looking, so the decoder is built to serve four roles at once.

The SEO or AEO analyst asks whether the optimisation work is moving anything, and reads per-engine referral share and fetch rate to see which pages pull AI visitors and which engine favours which page. The content or brand team asks which pages to keep answer-ready, and uses landing-page distribution plus the citation join to find where a cited page sends people who then bounce — the edit the data prompts is the value, not the chart. The agency operator runs the same questions across many client brands and needs the split reported per brand, not pooled, so a client-ready summary assembles cleanly. The executive or founder asks one thing — “is AI search sending us anything?” — and needs a headline number with its confidence stated, not an anecdote about a prospect who mentioned ChatGPT on a call.

How it connects to the rest of Asva, including the MCP server

Being cited is upstream; being visited is the result you can bank, and the two are best read side by side.

The decoder is the impact layer above Asva's visibility stack. Pair it with citation intelligence to see which of your pages engines actually quote, and with the brand visibility tracker to see how often you appear across engines in the first place; AI search monitoring is the continuous run behind the trend line, and prompt intelligence shows the questions buyers ask that produce those answers. The citation side is where Asva's own tracked numbers live: across the 12,849 citations Asva recorded in the 30 days to 18 September 2026, third-party media supplied 51.86% of the sources and company-owned pages 40.29%, and reddit.com was the single most-cited domain at 439 citations (source: Asva AEO MCP, 30-day window to 2026-09-18) — the pages sending you AI visitors are often not your own, which is why Reddit AI tracking sits in the same stack. That whole workspace is also queryable as an MCP server: the tracker behind these features is exposed as OAuth-authenticated MCP tools at mcp.asvaai.com/mcp, documented at /docs/mcp, so an analyst can ask an assistant for share of voice, citations or competitor rank in the thread where they draft the fix. All of it feeds answer engine optimization and ties to outcomes through the measurement chapter of AEO-101.

How it differs from a rank tracker or a social listening tool

A rank tracker watches positions on a results page; a listening tool watches mentions on the social web; neither reconciles AI answers with your own traffic.

The distinction is a different data source and a different job. A generic rank tracker measures where a URL sits for a keyword in classic search; it has nothing to say about an answer that never showed a ranked list, or about a visit that arrived with no keyword attached. A social listening tool counts where your name appears across social and news; it reads public mentions, not the referral column in your analytics, and cannot tell a retrieval-agent fetch from a human click. The decoder works on two planes those tools do not touch at once — your GA4 sessions and your raw server logs — and its unit of work is the AI-assisted visit and the machine fetch behind it, kept apart and tied back to the specific page and engine. Where it overlaps with Asva's own tools the division is clean: visibility trackers count whether you appear, the decoder counts whether that appearance sent anyone.

Metrics glossary

AI-assisted session

A human website visit recovered as originating from an AI answer and moved out of Direct into a named AI channel.

AI channel

The GA4 source/medium grouping that folds known assistant hosts (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com) into one reportable channel.

Per-engine referral share

The split of AI-assisted sessions by the engine that sent them.

Direct-recovery share

The portion of former "Direct" traffic that decoding could reclassify as AI-assisted.

Landing-page distribution

Which pages AI visitors arrive on, ranked.

Retrieval-agent fetch

A request from an engine's answer-time bot (e.g. OAI-SearchBot, PerplexityBot, ClaudeBot) reading a page to answer a live question or ground a citation.

Training-crawler fetch

A request from a corpus-building crawler (e.g. GPTBot, CCBot, Google-Extended, Applebot-Extended).

Fetch rate by page

How often retrieval agents fetch a given URL over a period; a leading indicator of upcoming citations.

Assisted conversion

A conversion where an AI-assisted session was on the path but not the last click, reported separately from last-click.

Confidence tier

The label for how a session's origin was established: referrer-confirmed, parameter, or pattern-inferred.

Frequently asked questions