Analytics Roundup – Updates from June 2025

AI agents are coming for your analytics

Anonymous user tracking has pretty much always relied on cookies, and cookies are associated with browsers, not people. This means that a person who browses your site from three different devices (and therefore browsers) shows up in your analytics as three users (mostly). 

Privacy regulations, ad blockers, firewalls, browser privacy controls and other factors make the problem even more difficult.

And it’s about to get a lot harder, and weirder.

I went to Salesforce Connections a few weeks ago and I have a lot of, uh, feelings about it, but they did make one thing abundantly clear: the future is agentic* and Salesforce is here to help.

While there was no shortage of hyperbole at the conference, I’m inclined to agree that agentic AI will be significantly more transformative than LLM chatbots, and that’s saying something. Google is touting it too, with AI Mode, and Claude and ChatGPT boast an ever-increasing range of agentic capabilities.

The tie-in to user tracking is the fact that these agents are going to be visiting websites, harvesting content, and doing stuff on people’s behalf. So, on the other side of that bot that is crawling your product detail pages or checking your appointment calendar could be a legitimate prospect in the midst of a customer journey. 

As I have delved into this, I have found myself wondering, “will people visit business websites at all in the future?” The idea that they won’t is not as farfetched as you might think, but even a gradual shift in that direction has major implications for tracking.

I’m still getting my head around this so I don’t have a definitive recommendation for what we should be doing differently, but broadly speaking it fits into my perspective that direct measurement and attribution is misleading at best, and that we should be relying more on inferential methods such as marketing mix modeling and incrementality to measure the impact of our website and marketing channels. The fact is that we can’t reliably track users, which means we can’t accurately attribute conversions using direct measurement.

*If you’re fuzzy on what “agentic” means, this IBM Technology video does a nice job of explaining it. In general, I like their AI videos a lot.

Product Updates

• Google Analytics consent settings hub enhanced with Tag Diagnostics
  Google has expanded the consent settings hub to include more detailed diagnostic information about how your tags are handling consent signals.
• ROI-Boosting Measurement: Google Leader Shares What’s New and Next in GA4, Dana DiTomaso interviewing Steve Ganem, Google’s head of product for Google Analytics
  Not exactly a product update, but Steve shares plans for this year and beyond. The presentation itself is about 24 minutes, followed by Q&A. A couple of things he talked about got me pretty excited:
  
  • Support for building dashboard reports within GA4, with a wide variety of chart types and a lot more customization than current GA4 reporting.
  • More integrations for importing metrics from other platforms and increased reporting support for imported data.
  • The assisted conversions report is coming back!
  • There is quite a bit more covered in the video.
• I tend to be a bit skeptical of roadmap presentations like the previous bullet, but GA just announced that they now support direct import of cost data from a wide variety of data sources, so some of what Steve talked about is already becoming true!

Workflow

• Enhanced Conversions & Consent Mode: How They Work, MeasureMinds Group
  A step-by-step example of how to set up an enhanced conversion for Google Ads. The article also explains how Consent Mode works with the user-provided data that is necessary for enhanced conversions to work.
  If you are wondering what happened to the automated setup, there’s also this: What Happened to Automatic Enhanced Conversions Setup in Google Tag Manager?
• Common Mistakes When Working With Click Identifiers, Jude Nwachukwu Onyejekwe
  Various tips on working with click identifiers (gclid, fbclid, and a host of others).
• How to query Google Search Console data in BigQuery, Julius Federovicius
  A basic intro to SQL in the context of analyzing Search Console data in BigQuery. If you are not familiar, you can set up a daily bulk export of data from Search Console to BigQuery. I love the depth of data it provides, and Julius includes a few sample queries to get you started. Beware of the anonymous queries he mentions – GSC actually hides the queries that drive about half of all clicks. Based on the long-tail distribution of query volume, safe to say that we can see far fewer than half of all unique queries. In spite of that limitation, it’s still one of the best sources of business intelligence we have at our disposal, IMO.

Attribution

• Last month, I included an article from Dan Taylor that compared website conversion rates for organic search traffic versus AI traffic. In the data he analyzed, organic generally performed better than LLM traffic. A few more data points below:
  
  • • Case Study: 6 Learnings About How Traffic from ChatGPT Converts, Seer Interactive – this case study is based on a single client, but found that ChatGPT traffic converted at about 9X the rate of Google organic. A sample size of one is not enough to say “how traffic from ChatGPT converts”, but it’s still a dramatic finding.
    • The Average LLM Visitor is Worth 4.4x the Average Traditional Organic Search Visitor, SEMrush – this article doesn’t provide any detail on where the 4.4X number comes from, so hard to know how trustworthy it is.
    • AI Visitors Visit Fewer Pages and Bounce More Often Than Traditional Search Visitors, ahrefs – does not cover conversion rates, with the exception of a few anecdotes, but is the most comprehensive analysis of the bunch. Also kudos for providing details on the data used for the analysis.
  • Last month I also mentioned that AI Mode impressions and clicks would be visible in Google Search Console. Technically that’s true, but so far they are lumped together with other clicks and impressions. Underwhelming.
  • Related to my musings above about AI agents obfuscating user tracking, this LinkedIn post from Dan Hinkley describes how to analyze web server log files with Screaming Frog to see when ChatGPT visits your site in response to a specific user’s prompt.

Ideas

• • The Rise of Systems of Consolidation Applications, Selina Wang
    A big thinker’s perspective on the evolution of how enterprise data is managed and activated. As I read the article, I imagined being able to easily access sales, CRM, inventory, finance and other data in the context of marketing impact analysis and optimization. It was a bit of a downer to return to the real world, where data silos are still the norm.
  • Vibe Analytics: When Everyone Becomes an Analyst (And Analysts Become Everything Else), Timo Dechau
    There are a lot of thought-provoking ideas here. One that stands out for me is the concept that “the data warehouse becomes less like a carefully curated museum and more like a well-organized storage room.” The reason being that AI can synthesize data and construct models on the fly, based on the specific business question(s) being asked.
  • Semantic Layers: The Right Idea, the Wrong House, Olga Berezovsky
    Before reading this, my context for understanding the value of having a semantic layer was the occasional, condescending comment, “oh, you don’t have a semantic layer??” Wanting to be part of the in-crowd, this naturally made me want one too. Olga has way more experience in this than I do, and provides a clear-eyed breakdown of the costs and benefits, as well as when it makes sense to implement (spoiler: rarely). You have to pay to get a full subscription, but if you install the Substack app, you can read the article for free.