Hey Reader!
I recently replaced the WordPress search on Analytics Playbook with an AI chatbot, built using Botpress. I named it Coach, trained it on all of my content and courses, and now I have a new analytics problem to solve: how do I know if people who talk to Coach are more likely to actually buy something (or even just engage more!) than people who don't?
It’s possible but it also requires treating the chatbot as a new kind of event, not just a replacement widget you dropped on the site. And that's where a lot of feature rollouts disappear from an analytics perspective. Something new gets added and then the analytics setup, which was working fine before, now has a gap because the new thing isn’t just a replacement, it’s an enhancement.
With these kinds of changes you can’t just use the same event, now with a different data source. In my case, on-site search is tracked with the view_search_results event. Sure, I could use that event to track these conversations but the content of the conversation is likely going to be over 100 characters, so it’s going to get truncated. Plus, I’d want to know how many messages someone has, what kinds of answers they got, and so on.
What I’ll do instead is record the Botpress chat ID in GA4 and then I’ll create a connection between the Botpress data and the GA4 data, likely using BigQuery. I have the luxury of using my own time to set this stuff up, but when it comes to your own work, you likely have resource and/or budget constraints and can’t just decide to fool around in BigQuery for funsies.
These kinds of new features require some pre-launch thinking on how you’re going to track it properly, so make sure to get involved in that conversation as early as possible to get the budget and/or time you need to track these features correctly from the start (which is not what I am doing, oops).
How to Handle Google Ads During Your Slow Season (Without Losing Ground)
Sammy just published a guide on how to handle Google Ads during your slow season, and it covers the question you’ve probably heard before: should I pause? The short answer is usually no. Pausing means losing the learning your campaign has accumulated, which means that the campaigns have to learn again when you come back. Sammy recommends reducing gradually to keep you in the auction and to keep the algorithm learning. Plus other great advice that would probably make sense to PPC people, but not me.
If you have seasonal clients, send this one along!
Articles Worth Your Time
———•
|
It's Time to Quantify the "Halo Effect"
Nick LeRoy walks through a situation I recognize: a client with a high volume blog, one million+ organic clicks a year, more than a million dollars in directly attributable revenue, and they get hit by AI Overviews. As a result, clicks drop nearly 50% and directly attributed revenue drops 25%. Now leadership is asking whether the investment is still worth it.
Nick's answer is to introduce a supplementary metric he calls assisted conversion revenue. First-touch influence (users whose first site visit was through the blog, who then converted later), and any-touch influence (purchases where a blog page appeared anywhere in the journey). He's not claiming these as official KPIs, but they tell a much stronger story than last-click alone, and they make the business case in terms leadership can actually act on.
This piece is a great how-to for anyone trying to have that conversation with leadership right now.
AI Traffic vs AI Citations: What Clicks and Cited Pages Show About the AI Search Journey
Aleyda Solis always brings serious research to her pieces, which I really appreciate. In this article, she analyzed April 2026 Semrush data across 40 sites in four verticals and surfaced a distinction that a lot of people are conflating: AI traffic (where the click lands) and AI citations (which pages AI systems reference in answers) are measuring completely different layers of user behaviour.
Of course, organic search is still larger than AI traffic, we know that. But what’s useful is the data showing the pages AI systems use to build their answers and the pages people click to from AI answers are largely different pages.
If you're only tracking AI referral traffic, you're missing most of AI's actual influence on visibility. The takeaway is that you need to track AI presence, AI citations, and AI referral traffic separately, and segment each by page type before drawing any conclusions.
The AI Convergence Problem
In this piece, Mark Williams-Cook makes two great points: first, LLMs fail badly at novel problems, and second, LLMs are quietly pulling everyone toward the mean.
Shared training data plus shared incentives plus fast iteration loops equals marketing that is indistinguishable from everyone else's. He analyzed Hansard (UK Parliament transcripts) from 2007 to 2025 and tracked the spike in ChatGPT-typical phrases—"I rise to speak," "underscores," "bustling"—after ChatGPT launched. The graph is not subtle.
What to do? He suggests using LLMs for commodity work where being average doesn't matter but don’t use them for anything where differentiation is the whole point. In those cases, treat the LLM's first answer as a baseline to deliberately diverge from.
Where You Can Find Me
———•
|
BrightonSEO San Diego
I’m going to be busy this September: I’m teaching a workshop (“Question-Driven Reporting with Data Studio and AI”) and giving a talk (“Fewer graphs, better answers: dashboards that get used”). See you there?
That's it for this edition of The Huddle. As always, if you have questions or want to share what you're working on, just hit reply!