Explain a drop in Google Analytics 4 conversions, from Claude or ChatGPT
Compare two Google Analytics 4 periods, isolate the channel, page or device pulling conversions down, rule out false signals: one conversation with your AI, no GA seat per member.
Published on October 31, 2026
In short
- With Persyst’s Google Analytics connector, your assistant compares two periods in one call and isolates the channel, landing page or device behind the drop.
- It starts from the key events actually configured on the property, and flags data that is still provisional or hidden by Google before drawing conclusions.
- Everything is read-only, on the GA4 properties opened to the member, with no export and no user added to Google Analytics.
Our demo requests dropped this month. Where does the drop come from?
The “generate_lead” key event goes from 186 to 131 (-30%) over the last 28 days ending the day before yesterday. Paid Search accounts for two thirds of the gap: 41 fewer conversions with sessions nearly flat, so traffic still arrives but converts less. On that channel, the drop is concentrated on mobile and on the /demo landing page, whose mobile conversion rate goes from 4.1% to 1.9%. I would check the /demo form on mobile.
Sources consulted
- Google Analytics · key events
- Google Analytics · 28-day comparison by channel
- Google Analytics · landing pages by device
The problem: fewer conversions, but why?
The Google Analytics acquisition report shows conversions going down. It does not say where the gap comes from. To find out, you compare two periods, then drill down channel by channel, landing page by landing page, device by device, while staying wary of the last few days that are still incomplete.
An AI assistant does this in one conversation, as long as it can read the GA4 property.
What the assistant does, step by step
With Persyst’s Google Analytics connector, the question “Where does the conversion drop come from?” runs through this reasoning:
- List the key events of the property (
list_ga_key_events). The assistant counts the conversions that actually exist, and says when the one you are looking for is not set up. - Compare two periods by channel (
run_ga_report, two date ranges in a single call). Sessions and key events side by side: falling traffic is not handled like a falling conversion rate. - Drill into the channel at fault, by landing page and by device, in the same multi-dimension report.
- Look for a custom dimension when needed (
get_ga_metadata): form type, offer, site language.
The answer separates the likely causes: fewer visits, visits that convert less, or a measurement issue.
Asking the right question
- Name the key event if you track several: “demo requests”, “purchases”.
- Specify the periods: “the last 28 days against the previous 28, ending the day before yesterday”.
- Ask why, not just for the number: the assistant then chains the breakdown by channel, page and device.
Three questions to try
- “Which landing pages have the most sessions and the lowest conversion rate this month?”
- “Compare mobile and desktop engagement rate on /blog pages over the last three months.”
- “How many people are on the site right now, and on which pages?”
For Google search, the same reasoning applies to clicks with Search Console.
Setup
- An administrator connects Google Analytics under Data sources, with a Google account that has access to the property. Persyst only requests read access; the Google Analytics guide covers the steps.
- They open the property to the marketing team, an agency or a single person.
- Each member adds Persyst to their assistant (getting started) and asks their question.
Limits to know
- Standard reports and realtime, not explorations (funnels, path analysis, cohorts).
- 500 rows per call; beyond that, the assistant pages through.
- Google’s thresholds and sampling apply as in the interface; the assistant flags them.
- No Search Console or Google Ads data in this connector: a GA4 session is not a Google click.
Questions
Frequently asked questions
Do I need access to the Google Analytics property?
No. Someone with access to the property connects it once in Persyst. Authorized members then query it from their AI, without being added as users in Google Analytics.
Why leave yesterday out of the comparison?
Data from the last 24 to 48 hours is not final yet in GA4. A drop on those days is often an artifact. End the period two days before today; the assistant is told not to read a drop there as a real decline anyway.
The rows do not add up to the total. Is that normal?
Yes, when Google applies thresholds: it hides low-volume rows to protect anonymity. The assistant is told and compares rows with each other rather than with the total.
Can the assistant fix conversion tracking?
No. The connector only has Google Analytics’ read scope. It can tell you an event is not set up as a key event; the fix happens in Google Analytics or Google Tag Manager, by someone with access.