AI search, from first answer to real revenue.

Practically: ChatGPT (chat + search modes), Perplexity, Claude (chat + research), Gemini (chat + AI Mode), Microsoft Copilot, Google AI Overviews, plus the long tail of smaller engines (DeepSeek, Phind, You.com). They each answer queries by synthesizing across sources rather than listing them. From an attribution perspective, they all share the same fundamental problem — they often strip the referrer, send traffic in patterns GA4 misclassifies, and route the visit through architectures classic analytics was not designed for.

You stop relying on the GA4 default channel groups and add a custom AI-engine classifier — either via GA4 Custom Channel Groups (Admin → Data display) or, more reliably, via server-side detection that inspects the referrer, user-agent, and landing-page pattern. The dark-ai-traffic-ga4 article in the overview section above is the diagnostic walkthrough, and the chatgpt-referral-analytics-guide is the deep-dive on the recovery path. Attrifast does this classification automatically on the server side; if you do not want to roll your own, that is the wedge.

In our 200-site Stripe-connected cohort, ChatGPT sends the highest absolute volume of paid conversions. Perplexity sends the highest conversion rate per visit but at lower volume. Claude is the smallest volume but with the highest order value in B2B-skewed properties. Gemini sits between Perplexity and ChatGPT on conversion rate. AI Overviews citations show up as Google referrals (not as a distinct AI engine) and convert similarly to other Google organic traffic. The chatgpt-vs-perplexity-vs-claude-traffic-quality article in the measurement section has the per-engine breakdown.

Yes, materially, in our 200-site cohort — AI-sourced traffic converted at roughly 3-4x the rate of generic search traffic on average, with significant variance by vertical. The honest interpretation: AI traffic skews toward higher-intent queries (the user has already had a conversation with the model and arrived with a specific question), and the population doing AI search trends toward early-adopter / higher-income segments. Some of the conversion lift is intent quality, some is demographic. Read the chatgpt-vs-google-traffic-quality piece for the full data and methodology caveats.

Two surfaces with two different lever sets. For the live-retrieval surface (ChatGPT search, browse mode, AI Overviews, Perplexity), the levers are structural — schema markup, FAQ blocks, direct-answer formatting at the top of the page, primary-source citations in the body, freshness signals. For the training-corpus surface (no-browse model answers, default model recommendations), the levers are authority-based and slow — Wikipedia presence, Reddit mentions, consistent entity data, third-party citations from authoritative publishers. The /aeo and /geo hubs cover the optimization playbooks; this hub focuses primarily on the tracking and measurement layer.

Traditional SEO optimizes for blue-link rankings; AI search optimization additionally optimizes for being the source the model summarizes. A page can rank #2 in Google for a query and get cited in 80% of ChatGPT answers for that query — or rank #1 and be cited in 0%. The two states are correlated but not identical. The structural signals that earn AI citations (schema, direct-answer blocks, primary citations) overlap with classic on-page SEO but are not the same set. Most teams now run both as a unified practice.

Mostly yes, with engine-specific quirks. GPTBot (ChatGPT training crawler) and OAI-SearchBot (ChatGPT live search) obey robots.txt directives. ClaudeBot (Anthropic training) obeys robots.txt. PerplexityBot historically had some gray-area incidents; current behavior is mostly compliant. Google-Extended is the directive for Gemini training opt-out. llms.txt is a newer, voluntary convention — adoption is low (~7% of public SaaS sites in Q1 2026) and not all crawlers consume it. The llms-txt-vs-robots-txt piece is the full reference.