AetherCutChangelogHow it worksPricingOpen the editor

Home / Changelog / v1.9.13

v1.9.13

AI-search attribution: ChatGPT/Perplexity/Claude/Gemini referral tracking

AetherCut v1.9.13 shipped on Jun 3, 2026.

What's new in v1.9.13?

  • We now capture referrals from every major AI assistant: ChatGPT, Perplexity, Claude, Gemini, Copilot, You.com, Meta AI, Phind — so when they cite AetherCut, the resulting traffic is attributed.
  • Central classify_referrer() taxonomy in routes/attribution.py — single source of truth for all implicit-source detection across email signup, Google OAuth, and page-view paths.
  • New /api/admin/attribution/ai-search dashboard endpoint returns {chatgpt: {page_views, signups, paid, revenue_cents}, ...} for a configurable lookback window (default 30 days, clamped 1-365)
  • Also expanded social-referrer detection: Twitter/X, LinkedIn, Hacker News, Product Hunt, YouTube — previously only Reddit was recovered when UTM params were missing.
  • 6 new pytest cases covering classifier mapping, page-view auto-tagging, rollup math, admin auth, and window clamping. All 32 backend tests pass across the critical suites.

How attribution actually gets recovered when AI assistants don’t send clean UTMs

The useful part of this release is not just that more sources are recognized, but that attribution now survives the way AI-assistant traffic usually arrives in the real world: inconsistently. In practice, an assistant may open a page inside its own browser wrapper, hand off to a normal browser, strip campaign parameters, or send only a minimal referrer. That creates a familiar analytics failure mode where visits appear as direct traffic even though the user first discovered the page through an answer in ChatGPT, Perplexity, Claude, Gemini, or another assistant. The new attribution path closes that gap by treating source detection as a classification problem instead of depending on one fragile signal. If a user lands on a page after clicking a cited result, the system can infer the assistant family from the request context and then carry that source through subsequent page-view, signup, and payment events.

A concrete example makes the behavior clearer. Say AetherCut publishes a page about subtitle cleanup, and a user asks Perplexity for tools that can remove filler words from transcripts. Perplexity cites the page, the user clicks through, reads for a minute, creates an account with Google OAuth, and later upgrades. In many stacks, only the final signup or billing event gets stored, and by then the original source is gone unless a UTM survived every redirect. Here, the classifier runs consistently across the page-view and auth flows, so the first identifiable source can still be attached even if explicit campaign tags never existed. The same pattern matters for ChatGPT and Gemini especially, because traffic from answer engines often shows stronger intent than generic search visits but is harder to measure using legacy channel rules. With the new dashboard endpoint, that journey stops collapsing into “direct” and starts showing up under the assistant that initiated it.

There are a few implementation mistakes teams usually make when they try to patch this after the fact. The first is scattering source-detection logic across multiple code paths, which leads to subtle disagreements between what the landing-page tracker saw and what the signup handler records. Centralizing classification avoids that drift. The second is overfitting detection to a single hostname or referrer string and assuming those values are stable forever; assistant products change browser behavior, embed flows, and redirect chains often enough that attribution logic has to be maintainable, not just clever. The third is counting only visits and forgetting revenue rollups, which makes the data interesting but not decision-ready. A source table that can answer “which assistant produced paid conversions in the last 30 days” is materially more useful than one that only shows top-of-funnel clicks. Finally, teams often forget to clamp reporting windows, which sounds minor until an internal dashboard query turns into an accidental full-history aggregation.

If you are adopting this release and want to sanity-check it quickly, the simplest approach is to test one assistant at a time with an end-to-end path instead of only inspecting raw requests. Click through from an actual cited answer, confirm the first page view is attributed to the expected assistant family, then complete a signup and verify the same source persists in downstream reporting. After that, compare the AI-search numbers with your pre-existing direct-traffic baseline over a few days. You should expect some “direct” volume to shrink as previously unattributed assistant visits get reclassified. It is also worth checking pages that historically earned mentions in social channels, because the expanded non-UTM social detection can shift traffic out of direct at the same time. That combination is what makes the release more than a dashboard addition: it improves the integrity of the acquisition record across the full path from discovery to revenue.

How do I get this update?

Open the editor and you have it. AetherCut runs in Chrome, Edge, Safari and Firefox. There is nothing to install, and you can edit without an account. Open the editor or see how it works.

What shipped just before and after?

See all 111 releases · Follow the RSS feed