Webclat / AI Visibility

Verifying AI conversion tracking: the runtime check, platform by platform

A dashboard that shows conversions is not proof the pixel works - it might be showing you the platform's own model of what should have happened. The only real proof is watching the actual request leave the browser and land with a success status. This is the house differentiator: we check the wire, not the dashboard.

In short

Verifying an AI conversion pixel means opening the browser's network panel, filtering by the platform's own request host, and confirming the expected configuration-then-event request pair fires with a success status, only under the right consent state, and exactly once per real event. The specific hosts and status codes differ per platform - documented on each platform's own page in this cluster - but the method is identical everywhere, and that consistency is what makes it teachable and auditable.

Why the dashboard isn't enough

A platform's own ads-manager reporting reflects what its collection endpoint received and how its attribution model chose to count it - both useful, neither a substitute for confirming the request actually left the browser correctly, with the payload it should have, under the consent state it should have. A tag can be "reporting conversions" in a dashboard while still firing on pages it should not, sending more data than intended, or double-counting against a server-side event with no deduplication. The dashboard tells you the platform's story about the data; the network panel tells you what actually happened.

The method, in five steps

  1. Open DevTools → Network, and clear it. Reload the page from a cold state so you catch the very first request the pixel makes, not just whatever fired after you started watching.
  2. Filter by the platform's request host - for example bzr.openai.com for OpenAI, the relevant UET or Meta domains for those platforms. Filtering by host, not by request name, catches requests you did not know to expect.
  3. Confirm the configuration or SDK-load request fires on page load, with a success status. This is the pixel announcing it is present and correctly configured - its absence is the single clearest sign of a broken or blocked install.
  4. Perform the tracked action (add to cart, checkout, purchase - whatever the pixel is meant to capture) and confirm an event request fires, with the expected status, containing the event type you triggered.
  5. Repeat under each consent state your site supports - consent granted and consent declined, at minimum. A pixel that fires identically in both states has no consent gate at all, whatever the installation checklist claims.

What healthy looks like, by platform

PlatformHost to filterHealthy pattern
OpenAI (ChatGPT Ads)bzrcdn.openai.com, bzr.openai.comConfig GET (200) on load, events POST (202) per tracked action
Google (AI Mode / Gemini Shopping)Existing Google Ads / GA4 tag hostsStandard gtag conversion event, plus an Enhanced Conversions hashed-data payload on the same event
Perplexity (Buy with Pro)None to verify on the wire - no merchant-facing pixel foundVerification here means confirming the order-tag value appears correctly in your own order-management system
Microsoft Copilot (Copilot Checkout)Existing UET tag hostUET tag fires on the checkout-confirmation page; CAPI call as a separate server-side check if enrolled in the pilot
Meta (AI-enriched Pixel + CAPI)Existing Meta Pixel hostPixel event fires with enrichment fields populated; matching server-side CAPI event with a shared event ID for deduplication

The four failure modes worth naming

  • Missing configuration request. The pixel is not installed, or something (a consent tool, an ad blocker, a content-security policy) is blocking it before it can load.
  • Event request with no prior configuration request. The SDK is firing without having initialized correctly - a misconfiguration, not a working pixel that happens to look incomplete.
  • Firing regardless of consent state. The most consequential failure - not a measurement gap but a compliance one, and the exact scenario the illustrative story in our pixel-installation use case is built around.
  • Duplicated events with no shared identifier. Common wherever both a browser pixel and a server-side conversions API report the same action - inflates every downstream number until deduplication is confirmed working, not just configured.
The house differentiator, stated plainly

Any agency can point a client at a platform's help article. What we do differently is confirm the request on the wire before a number goes into a board report - the same standard we apply to our own estate's crawler and answer-tracking instruments, extended to every vendor conversion tag we install.

Common questions

What does runtime verification of a conversion pixel actually mean?

Watching the pixel's real network requests fire in a browser session - not reading the vendor's install documentation and assuming it behaves as described. A confirmed request with the right status code on the wire is evidence; an install checklist marked complete is not.

Does the same verification method work across every AI platform?

The method is the same - network panel, filter by the platform's request host, confirm the expected request pair fires with the expected status - but the specific hosts, request shapes, and status codes to look for differ per platform, which is why each platform's page in this cluster documents its own.

What does a healthy install look like on the wire, in general?

A configuration or SDK-load request on page load, followed by an event request on the tracked action, both returning a success status (200 or 202 are common), firing only when consent allows it, and appearing exactly once per real event - not zero times, not duplicated.

What is the single most common failure mode?

An event request firing with no prior configuration request - meaning the SDK is sending data without having loaded correctly first - or a consent-gated tag firing on every page load regardless of the visitor's actual consent state, which was the exact failure this discipline exists to catch.

Get every AI-platform pixel confirmed on the wire, not just in a dashboard.

Network-panel verification per platform, consent-state testing, and a written record of what was actually observed.

Verify My AI Conversion Tracking