Attributing AI Traffic: Our Measurement Stack
How we connect AI crawler requests, citations, identified referral clicks, and conversions without overstating what referrer data can prove.
AI attribution is a chain with four separate events: a crawler fetches a page, an answer cites it, a person clicks, and that visitor converts. No single chart proves the whole sequence. Each event has a different collection method, and the final commercial outcome belongs in the client's analytics or CRM.
We build the stack in that order. It gives us a practical diagnosis when performance stalls. No crawl evidence points toward discovery or access. Crawls without citations move the investigation toward content and source selection. Citations without identified clicks raise questions about the answer context and the reason to visit. Clicks without conversion belong in the landing experience, offer, analytics setup, or sales process.
Layer one: crawler evidence at the edge
JavaScript analytics does not measure crawler requests that never execute the tracking script. We collect AI bot activity from the CDN, edge, or another supported server-side log source. The record should contain the request time, path, user agent or classified crawler, status, and enough verification information to assess whether the request is genuine.
Training bots, search index bots, and live citation fetchers are kept separate. Their requests have different implications. A training crawl is not proof that a page can appear in current search answers. A live fetch says the provider requested the page, but it still does not prove the final answer cited it.
Promptwatch's AI crawler insights guide documents the crawler, path, status, and date views available in its logs. It also describes provider IP checks where published ranges exist. We use those records as observed access evidence and keep the identity limitation visible for crawlers that cannot be verified the same way.
Layer two: citations in monitored answers
Citation monitoring records when an answer links to a URL. We map cited URLs to canonical pages, prompt groups, model, market, and observation date. This layer tells us whether the client's own page, an offsite mention, a Reddit thread, or a video supplied the source.
A citation is not the same as a brand mention. A model can use a client's documentation to support a general fact without recommending the client. It can also mention the brand while citing another publisher. Reporting both events prevents a strong citation count from being presented as recommendation share.
Page-level comparison matters here. We line up crawler activity and citations for the same URL and period, while allowing for delay and other retrieval paths. We do not claim that a particular request caused a later answer. The sequence is evidence for diagnosis, not a deterministic attribution model.
Layer three: identified AI referral clicks
Visitor measurement uses the referrer the browser sends when someone clicks from an AI platform. We install the relevant script directly or through the client's tag manager, test it, exclude internal activity where the system supports that, and compare its landing-page records with the broader analytics platform.
Promptwatch's visitor analytics documentation separates total site visits from visits attributed to listed AI referrer domains. That distinction belongs in the report. Total traffic should never be relabeled as AI traffic merely because the same script observed it.
Referrer-attributed clicks are a floor. Some visits arrive without a usable referrer. AI answers may also influence someone who later returns directly or through branded search. We acknowledge those paths, but we do not calculate a multiplier for unobserved AI influence. The data does not support one.
Landing pages make the click layer useful. If a cited article receives identified AI visits, we can inspect what those visitors do next. If a frequently cited page receives few recorded clicks, the answer may satisfy the question without requiring a visit, or the citation may be presented weakly. Those are hypotheses to examine, not automatic conclusions.
Layer four: conversion and revenue
Promptwatch directly tracks the first three parts of the chain: crawler activity, citations, and identified referral clicks. According to its AI traffic attribution guide, conversion and revenue should be joined in the organization's own analytics or CRM.
We configure an AI referral segment from the same domains used at the click layer, then inspect conversion events by landing page and date. The definitions come from the client. A lead is only a lead if its analytics or CRM rules say so, and revenue is only attributed according to the model the business has approved.
For B2B, that may require connecting a session or form event with later CRM stages. For commerce, it may involve transaction records in the analytics or store platform. We do not move revenue into a visibility dashboard by hand or assign value to a mention that produced no identifiable session.
Consistency matters more than picking an impressive attribution model. We document lookback settings, channel rules, internal traffic exclusions, timezone, currency, and any consent-related data loss. If the client changes one of those settings, the report marks the break rather than comparing incompatible periods.
Reporting the joined chain
Our working view is page based. For each priority URL, it shows successful crawler activity, observed citations, identified AI referral visits, and conversions from the client's system. We add publication and revision dates so the reader can see what changed. Totals remain available, but pages reveal where the chain breaks.
We also keep model and prompt context. A visit from an AI domain usually does not disclose the exact prompt that led to the click. A nearby monitored citation can support analysis, yet it does not prove that the visitor saw that monitored answer. We state this plainly in client reports.
Teams that already report through Looker Studio can use the Promptwatch connector to bring visibility, citation, prompt, sentiment, and crawler data into existing dashboards. Analytics and CRM sources still need their own connectors and governance. A shared dashboard is a presentation layer, not evidence that every row has user-level attribution.
We run client programs on Promptwatch because it places the first three measurement layers close together. Our stack keeps the fourth with the systems that own it. The monthly review then has a concrete order: find the earliest missing event for each priority page, assign that problem to the responsible team, and preserve the measurement limits in the client record.