Attrifast Research
Original AI search benchmarks — traffic, citations, and revenue.
Eight original studies on AI traffic and AI citations, drawn from the Attrifast 200-site Stripe-connected cohort and our quarterly prompt-tracking studies: AI search performance data down to the engine level, and content performance metrics down to the cited page. Every study names its sample size, exposes its methodology, and admits its limitations — numbers other operators can replicate and check.
7-day free trial · $0 due today · from $9.99/mo · cancel anytime
If you want one piece of context for everything that follows, read the 2026 AI Search Revenue Benchmark first — the per-engine RPV table is the data point most operators are missing.
Headline findings: AI search performance at a glance
Every figure below comes from the same cohort: 200 Stripe-connected sites, opt-in and anonymized, June 2025 to May 2026, with each session joined to a settled payment through a first-party visitor ID.
| Finding | Value | Study |
|---|---|---|
| Share of GA4 "Direct" traffic that is actually AI-referred | 34% | Dark AI traffic in GA4 |
| Revenue per visitor, cohort-blended, by engine | Perplexity $1.42 · Claude $1.18 · ChatGPT $0.87 · Gemini $0.41 · AI Overviews $0.29 | The 2026 AI Search Revenue Benchmark |
| Revenue per visitor on B2B SaaS specifically — Claude leads | $1.94 (vs a $1.18 cohort blend) | The 2026 AI Search Revenue Benchmark |
| Conversion rate, AI-referred sessions vs Google organic | 2.7% vs 1.4% | The 2026 AI Search Revenue Benchmark |
| Median under-count of AI-referred revenue on default GA4 settings | 64% low | Dark AI traffic in GA4 |
The studies
- 200 Stripe-connected sitesThe 2026 AI Search Revenue BenchmarkReal data from 200 Stripe-connected sites — per-engine RPV, conversion rate, and ROI vs paid search. This is the piece I send when a CFO asks 'is this channel real?'
- 200 sites, ~12M sessionsAI Traffic Conversion Rate Benchmarks 2026What good looks like by channel, vertical, and AI engine — methodology disclosed, confidence intervals named, and the verticals where AI traffic underperforms are not hidden.
- 1,200 prompts × 4 enginesAI Citation Rates by Industry 2026A 1,200-prompt study across ChatGPT, Claude, Gemini, and Perplexity. Citation rate by industry, with the categories where AI engines disagree dramatically about who to recommend.
- 200 sites segmented by ARRAI Citation Rate Benchmarks by Brand SizeStartup vs growth vs enterprise — the 200-site analysis showing citation rate by stage, and the structural pattern that lets smaller brands outpunch their domain authority.
- 200 sites, per-engine splitChatGPT vs Perplexity vs Claude Traffic QualityWhich AI engine sends the best visits — 200-site cohort study with per-engine RPV, conversion rate, and order-value breakdown. The volume-vs-quality trade-off named explicitly.
- 200 sites, paired channelsChatGPT vs Google Traffic QualityData from 200 sites — the conversion gap between AI-sourced and search-sourced visits. Includes the population-skew caveat most "AI traffic is great" claims gloss over.
- 200 sites, full distributionHow Much Traffic Comes From ChatGPT in 2026?The Attrifast 200-site cohort benchmark — what real ChatGPT referral volume looks like for SMB SaaS and e-commerce. With the percentile distribution, not just the average.
- 150 prompts × 3 runs eachWhich Brands Does ChatGPT Recommend in 2026?150-prompt, 450-run study across categories — the patterns in who ChatGPT names and why. Includes the categories where ChatGPT's recommendations are wildly inconsistent run-to-run.
A note on methodology
The Attrifast cohort is 200 Stripe-connected sites running our tracker across SaaS, e-commerce, indie products, and content businesses. The cohort is opted-in and anonymized for research use; no individual site is identifiable in any published number. Sample sizes named in each study reflect the subset of the cohort that met inclusion criteria for that specific question.
Prompt-tracking studies (citation rates, brand recommendations) are run by issuing the same prompt set against the public-facing chat interfaces of each engine, with at minimum 3 runs per prompt to capture run-to-run variance. We do not use vendor APIs unless explicitly noted, because the public chat interfaces are what real users see.
Where a study reports a limitation, the limitation is real and the number is the best we can produce honestly. If you find a methodological gap we missed, the contact form goes directly to me.
Your site, anonymized, in the next benchmark.
Attrifast customers are automatically included (anonymized, opt-out in dashboard) in the cohort we use for these studies. Two-minute tracker install, $9.99/mo, no engineering ticket required.
- One script tag and a Stripe key — live in minutes
- First-party measurement without third-party cookies
- Recognized AI referrals joined to Stripe payments
7-day free trial · $0 due today · then $9.99/mo · cancel anytime
