Kevin Indig Kalicube Summit 2026 session

Kevin Indig on the Grounded Gate — Prompt Tracking, AI Visibility Measurement, and Why Trust Has Replaced Rank

ByBernadeth Brusola

Session identifier

  • Session: Grounded Gate – Prompt Tracking, AI Visibility Measurement, and Why Trust Has Replaced Rank
  • Day: 1 (June 24, 2026)
  • Speaker: Kevin Indig, organic growth advisor and author of Growth Memo
  • Framework concept: The Grounded Gate — Gate 8 of the AI Engine Pipeline (ARGDW), where the engine decides whether to pull a brand’s content into a live response or leave it in the knowledge store uncited
  • SEL article: The funnel query pathway: A framework for measuring AI visibility

Speaker bio

Kevin Indig is an organic growth advisor and the author of Growth Memo, a weekly research-driven newsletter on SEO and organic growth with more than 25,000 subscribers. He led SEO and Growth at Shopify, G2, and Atlassian, and now advises some of the world’s fastest-growing companies — including Meta, Reddit, Ramp, Bounce, Dropbox, Hims, Nextdoor, and Snapchat — on search, AI visibility, and strategic growth. Kevin has published more empirical studies on AI engine citation behaviour than almost any other practitioner working publicly. Growth Memo has become the high-signal reference for operators trying to understand what AI engines actually do versus what they are speculated to do.

Session’s place in the Kalicube Framework

Kevin’s session covers the Grounded Gate — Gate 8 of the AI Engine Pipeline (ARGDW: Annotated, Recruited, Grounded, Displayed, Won), coined by Jason Barnard in 2024 as the activate phase of the Kalicube Framework. The Grounded gate is where the engine decides whether to pull a brand’s content into a live response or leave it in the knowledge store uncited. A brand can be Recruited — known to the system — and still never Grounded, which means it never appears in an AI answer. Kevin’s session is the empirical outside account of that gate: how to measure whether you are being Grounded, how often, in what contexts, and what the data reveals about the gap between classic search ranking and AI citation. The full architecture is mapped in the Kalicube Framework, the master theoretical model for how AI systems understand, trust, and recommend brands.

Episode summary

Kevin Indig opens with the measurement framework argument that underpins everything that follows: the old click funnel is dead for non-advertising channels. Clicks still exist but are not trending in the right direction, and direct attribution has become inferential at best. What replaces it is a leading-and-lagging indicator approach — bot crawls, retrieval, relative mentions and citations on the leading side; self-report attribution at checkout and CRM tracking on the lagging side. Kevin’s framing throughout is that this is not a crisis but a necessary adaptation. The weather and credit scores are tracked reliably despite being probabilistic. Keyword rankings were never perfectly deterministic either. The industry worked through that inconsistency before; it will work through this one.

The prompt tracking methodology section was the session’s most immediately actionable contribution. Kevin’s framework builds from three foundations: personas and jobs-to-be-done, intent stages across the buyer journey, and prompts tracked three to five times consecutively to understand citation consistency. The persona layer is non-negotiable because LLMs are highly personalized — the same query returns meaningfully different results for a CMO evaluating a CRM and a sales rep using one daily. Intent staging matters because the prompt set size and structure should reflect where in the buyer journey the persona sits, from definitional awareness to near-purchase confidence. The three-to-five consecutive tracking finding was the session’s most striking data point: after three consecutive prompts on the same query, approximately 2.2 percent of citations remain consistent. Kevin’s instruction was to track weekly rather than daily — daily tracking is too noisy — and to use the consecutive-run methodology to understand what actually survives, then build strategy from the patterns that emerge.

The convergence with Jason’s Funnel Query Pathway was the session’s most structurally significant moment. Kevin arrived at cohort-intent-based persona tracking from his own data work; Jason arrived at it from a framework-first direction. Both arrived at the same diamond-shaped architecture: a bottom-of-funnel conversion query, middle-of-funnel queries that naturally precede it, and a top-of-funnel brand-mention layer tracking which brands come to mind for a given topic. Kevin confirmed the approach as a sound proxy, because there is no way to know exactly what people prompt — all tracking is inferential, and that is entirely fine as long as the methodology is consistent, defensible, and explainable to a management team.

The user behavior study that closed the session was its most commercially decisive contribution. Kevin recruited approximately fifty non-technical, non-marketing people and gave them high-ticket purchase tasks — laptop, washing machine, insurance — across AI Mode and classic search. The finding: seventy-five percent of the time, people pick the first item on an AI shortlist. That is very similar to pre-AI search behavior where the top result gets the most clicks. But the only factor that overrides shortlist rank entirely is brand trust. When a participant encountered GEICO in an insurance shortlist they had not expected, they immediately chose it over whatever was ranked first — because they knew and trusted the brand. Kevin’s conclusion: trust is the most important commercial asset in the AI era, and it is the one variable the tracking methodology is ultimately trying to measure and strengthen.

Key quotes

“That old click funnel does not work anymore. I’m not saying clicks don’t exist — they do — but they’re not trending in the right direction. And so from a purely agentic perspective, we need to think about leading and lagging indicators that point towards business impact.” Kevin Indig, opening the session’s measurement framework argument”

“The fact that these are stochastic models should not lead to the conclusion that tracking them doesn’t make sense. We track a lot of probabilistic things — like the weather, credit scores. Not everything is perfectly deterministic.” Kevin Indig, on why AI volatility is not an argument against measurement”

“After three prompts, about 2.2 percent of citations remain consistent. Instead of saying none of this matters, none of this is relevant, we need to crack the nut and not shy away from it.” Kevin Indig, on the citation consistency finding that should reset how practitioners think about prompt tracking frequency”

“When people know and trust a brand in that shortlist, it doesn’t matter where it’s ranked — they will always prefer it.” Kevin Indig, on the user behavior finding that trust overrides shortlist rank entirely”

“I think links matter a lot less and mentions matter a lot more. And there’s a whole more or less black market around AI mentions that is somewhat similar to the black market we had for backlinks.” Kevin Indig, on the shift from link metrics to mention and co-occurrence signals in AI visibility”

Further reading

The funnel query pathway: A framework for measuring AI visibility, Article 14 of Jason Barnard at Kalicube’s Search Engine Land series on AI search architecture. The article maps the prompt tracking methodology that Kevin and Jason independently converge on in this session — persona-first, intent-staged, weekly rather than daily, tracking the path from top-of-funnel brand awareness to bottom-of-funnel conversion intent

Related sessions

Link to preview post

This session was previewed here before the Summit.

Link to speaker post

About Kevin Indig

Links to session clips

Trust Beats Rank: Kevin Indig’s Study on How People Actually Choose From AI Answers

Why the Old Click Funnel Is Dead: Kevin Indig on AI Attribution

Transcript for Kevin Indig on the Grounded Gate — Prompt Tracking, AI Visibility Measurement, and Why Trust Has Replaced Rank

Jason Barnard: Hi, and welcome to another session. This one is with Kevin Indig — one of my favourite people in the industry, super smart, and apparently with lots of data.

Kevin Indig: Jason, it’s always good to be here. Thanks for the compliments — I’ll come back.

Jason Barnard: Well, Claude was the one who told me you had a lot of data. Every time I had conversations with Claude about who to bring on this Summit to help me demonstrate the Kalicube Framework, your name just kept coming up. Claude was really insistent that you were an important person to have on it, because you’re going to bring a superb viewpoint to it.

Kevin Indig: It’s clearly the best model. What can I say?

Jason Barnard: So I’m looking today at grounding and display. I’ve talked to Krishna from Microsoft. I’ve talked to Michael King. We’ve got different perspectives on it. And it’s particularly the display part — how we’re measuring it and what we can do to improve our chances in a world where AI assistive engines and AI assistive agents are opaque. Let’s start there. How do you measure SEO performance in an AI world? Small question.

Kevin Indig: Small question, yes.

I think first of all, it’s critical to understand that the measurement approach is very different from what we’ve done before. In the SEO world, in the simplest sense, we looked at keyword ranks, traffic, maybe impressions, and then connected that to the business baseline. That old click funnel does not work anymore. I’m not saying clicks don’t exist — they do — but they’re not trending in the right direction. And so from a purely agentic perspective, we need to think about leading and lagging indicators that point towards business impact. But it’s all directional. There is no direct bridge anymore.

And my hot take on the meta-point here is that attribution for channels not related to advertising is basically dead or does not work. So we need to think about leading and lagging indicators.

Under leading, I would cluster things like bot crawls — is our site even crawled and visited? Things like retrieval — can bots actually retrieve our content? And then in the middle of that pipeline, I would look at things like mentions and citations, ideally in a relative context. I think absolute values are not helpful since there’s too much change and too much noise.

On the lagging side, that’s where we can think about methods like self-report attribution — meaning when people sign up or buy a product, you ask them: how have you heard about us? You can do this in more elaborate ways. And then maybe even for B2B businesses, some sort of CRM attribution tracking how customers or prospects found you. But it is very opaque, and there’s a lot of black-boxing going on. I’ll spare you the TED talk for now.

But even in the way we currently structure and design prompt tracking, there are significant gaps.

Jason Barnard: Number one: asking people how they found us is going to be the next thing we all need to do on the checkout form. Obviously you don’t want to create friction, but you do need to know. I’m asking people on sales calls: how did you find us? Most of the time when I ask who’s the best in the world at knowledge panels, they say: you. So Claude likes me almost as much as it likes you.

The second thing is: my series of articles on Search Engine Land — nineteen articles laying out the Kalicube Framework — started because Rand Fishkin in February published findings where he said something like two percent of responses to queries are the same. And his argument was: this is completely random, there’s no point in tracking it. My counter-argument was: if you search my name or Kalicube’s name on any of the models, it’s very consistent. And that’s because I’ve cleaned up my little corner of the internet. What’s your take?

Kevin Indig: The fact that these are stochastic models should not lead to the conclusion that tracking them doesn’t make sense. We track a lot of probabilistic things — like the weather, credit scores. Not everything is perfectly deterministic. And keyword tracking was never that perfectly deterministic either. There was personalization that Google did, especially on the local side depending on the device you were using. Rankings jumped around constantly if you tracked them daily or hourly.

The thing is: with keyword tracking and SEO, we agreed on a unified tracking method. There were not as many platforms as we have today in this AI world. So I think we’re going to work through this as an industry and find methods and approaches that work better. But we need higher rigidity in what we track and how.

Just to give you a glance at some of the gaps: right now, most prompt trackers just track a single version of a model like ChatGPT — but there are differences between model versions. There’s a difference when it comes to high or low reasoning. There are massive differences in response length, citations, and fan-out queries. There’s no common standard for fan-out queries. There’s just all these gaps that as an industry I think we can work through — and hopefully raise that 2.2 percent to maybe thirty, forty, fifty percent consistency.

I also expect a certain slowdown in the development pace and the shape of these models as the broader industry discovers that this is all pretty expensive, and compute constraints and so on come into play.

I agree with Rand’s broader point that this is a different thing than we had before. There’s a lot more randomness here, which means we need to be careful. But I think we can adapt to this and find models that help us make good choices over time and track their impact. We’re not confident today in direct attribution. So I’m not sure we’ll ever get to a point where we can with incredibly high confidence say exactly what the business impact is. I think it’s going to be much more like billboarding — where we can infer, we can get to some confidence on business impact, but we’re never going to know for sure.

Jason Barnard: I’m actually very grateful to Rand for having published that, because it did set me off on a nineteen-week, one-article-per-week run on Search Engine Land. And it was a real discovery for me as I worked through it.

At one point I thought: I’ve got a degree in economics. And micro to macroeconomics is a great parallel. We’ve gone from microeconomics to macro. You don’t measure inflation by pretending you know the price of everything. I go downstairs to buy my loaf of bread — it isn’t going to be exactly three percent more expensive because inflation is at three percent. So it’s the basket of goods idea. How would you approach that? I’ve got my own approach, but how would you approach it?

Kevin Indig: I love that analogy. The way I approach inflation measurement in this case is there are a few ways you can make prompt tracking better.

First: LLMs are highly personalized. I just published a new article about the improved memory states in ChatGPT — the way it collects memories and structures them. So it’s much more personalized, which means we need to operate on a persona basis. One of the first things I do with most of my clients is think about: what are the key personas we’re going after? What are their characteristics? And most importantly, what are the jobs they’re trying to do?

So to give an example: if you’re a CMO at a software company evaluating a new CRM to buy, you evaluate that from a different perspective than the sales reps who are using that CRM every day. So we need personas. Start with three to five, but you can technically develop hundreds of synthetic personas. It’s a very important one in the tracking world. First: who are we talking to? What are the jobs to be done?

Second: let’s talk about intent — specifically buyer intent. Within every persona, I usually like to look at three to five different stages of intent, from just trying to understand what something is — like “what is a CRM” — all the way down to “should I buy Salesforce?” I’m very close to a decision, just trying to gain confidence to make it. And between those are obviously several more steps. So personas and jobs to be done.

And then you need the right number of prompts. I also like tracking prompts three to five times consecutively to understand what the retention is across the parts of the answer — from citations to mentions and so on. And track weekly instead of daily. I think daily prompt tracking is way too noisy — we’re averaging it out anyway. But if you run the same prompt a couple of times consecutively, you better understand the consistency of the AI answer. These are a couple of ways I make AI inflation tracking much more accurate.

Jason Barnard: I’m quite stunned, because we’ve come to more or less the same conclusion sitting in our different parts of Europe. I’ve called mine the Funnel Query Pathway, and it’s cohort intent mapping. I have different funnels, each with a cohort intent combination at the bottom. I figure out the persona by asking: what is the ideal question they would ask just before they convert? Whether or not people actually ask it doesn’t matter — it’s what’s the ideal query? Then I move up the funnel and ask: what’s the query before that? And then at the top of the funnel, within this particular topic with this cohort intent, which brands are top of mind? So I’ve got a kind of diamond shape, and I measure that going down using brand mentions, prominence, and sentiment.

Kevin Indig: Yeah, it makes perfect sense. Because again, at the end of the day, it’s all a proxy — and that’s totally fine. There’s no way to know exactly what people prompt.

Jason Barnard: And I agree with you about weekly. People are tracking daily, tracking hourly, and as you say, it just creates noise. But I love your idea of tracking three times in a row to see the consistency. Sorry — I cut you off. Please finish what you were saying.

Kevin Indig: No, you’re fine. I love that we riff on this. But yeah — three to five times consecutively. I did a large data study, and we found that after three prompts, about 2.2 percent of citations remain consistent. So there’s obviously a lot of nuance depending on topic type and prompt type. But just to give people an idea of how much variance there is — and it is high. But instead of saying none of this matters, none of this is relevant, we need to crack the nut and not shy away from it.

Jason Barnard: And a couple of things come from that. Number one: Claude was right, you do have a lot of data. Number two: getting a link — being in the grounded set and getting a citation — is something people are quite obsessed about. How valuable are they in truth?

Kevin Indig: I did a large analysis to understand this. I looked at hundreds of thousands of query-page pairs and looked at all the common link metrics from SEMrush to understand whether there’s at least a correlation between link metrics and visibility in terms of AI mentions. And there is. The strongest correlation I found — I don’t remember the exact value, but it was above 0.5, so fairly strong — was with SEMrush’s aggregate link quality score. So you can say that high-quality links have a fairly robust correlation with AI visibility.

Now, I do think there’s a bit of a red herring here, though. I think it’s not necessarily the links themselves — I think it’s much more the mentions that matter. We know that mentions of brands on authoritative sites matter for LLMs. And I’d argue there’s more than just a plain mention: there’s probably also the context, the sentiment, and the attributes they’re connected to. You’re the expert in the field of entities, Jason, so I won’t preach to the choir. But I think links matter a lot less and mentions matter a lot more.

This is something Google has sort of included for years. But LLMs really care about this. And there is a whole more or less black market around AI mentions that is somewhat similar to the black market we had for backlinks. It’s going to be very interesting to see if LLMs or search engines take action against that. But right now I can tell you that companies are willing to spend four to five hundred thousand dollars a year for the right mention in the right place. This is not a small thing. I’m not sure it’s worth that amount of money, but there’s something there.

Jason Barnard: And the engines — OpenAI, Perplexity — don’t have the same means to analyze to the extent that Google and Microsoft Bing can. So there must be a weak spot there for them, given they don’t have the financial resources or infrastructure to analyze to the same extent that Google and Bing do. Is that a fair point or am I exaggerating?

Kevin Indig: I think there are more and more signals pointing at OpenAI building a web cache — essentially just indexing the web. And I do have the impression that reasoning is able to filter out a lot of manipulative brand mention insertion. So I’m not sure how big a problem it really is. Spam is obviously happening and working to some degree, but I’m not confident it’s as effective as a lot of people make it out to be.

Again, it goes back to the right measuring methodology — how you define your prompts and what you’re looking at. But in the grand scheme of things, I have yet in my own usage to come across spammy results or clearly manipulative answers. I’ve been scammed on Google Search not too long ago, but when it comes to LLMs, I’m not sure how effective that really is. At the end of the day, we need more data to give a better answer to that.

Jason Barnard: The advantage they have over old search results is that there’s an actual analysis and evaluation layer — waiting to figure out whether something is probably true or not. And Krishna was talking about the grounding and how the follow-up questions are a stunning piece of engineering — almost quasi-human, I think I’m paraphrasing rather than quoting — the idea that the engine is designed to try to help you move down the funnel. For him, that’s a huge piece of engineering.

Kevin Indig: Yeah, I couldn’t agree more. It’s massive, and it’s probably only possible for certain companies at a certain scale, with a certain data repository and a certain level of usage. So if you’re Perplexity — and it seems that Perplexity has fallen significantly behind ChatGPT or Gemini — you’re getting a lot less reinforcement learning from usage. The more people use an LLM and ask follow-up questions, accept them, give a thumbs up, et cetera, the easier it’s going to be for those LLM makers to improve the system over time. So there is a real moat, a real network effect — similar to how Google uses clicks and user engagements to make search results better. The same system exists in AI based on how users react to your initial answer. And the data advantage there is pretty clear.

That’s why I think there’s so much more advancement in coding — coding is much more deterministic than everything else, and it’s much easier to say whether the output was good or bad.

Jason Barnard: And you make the point that Google has a huge data advantage. It also has the entire funnel plus the post-click data, and it’s the only one that can actually measure the entire customer journey.

Kevin Indig: I think they’ve won the race at this point. I’ll be honest, Jason — I left enough traces across the web, but when ChatGPT came out at the end of 2022, I wasn’t sure Google would be able to respond. But they truly have, in a fascinating and outstanding fashion. And I think now it’s pretty clear that they’re owning the consumer space — with all the implications and challenges that brings for the web.

You have this whole economy that basically grew on classic web search results and traffic. I don’t think that can continue, even if Google adds more links to their AI answers. The shift is kind of out of the bottle here. And that brings us back to the earlier point about why improving the way we measure our brand presence and the biases that models have about our brands is so critical.

Jason Barnard: And brand is the foundation — my favourite topic in the world.

Kevin Indig: Yeah, it is the foundation for sure.

Jason Barnard: So I’m not going to say search engines, assistive engines, and assistive agents are all going to live together in some neat hierarchy. Search isn’t going to die — it will continue to exist because people like to window shop. And people will want an agent to do the job for them because they’re willing to delegate. I’m looking at the Delegation Boundary: where do we as human beings delegate — from allowing the machine to choose the ten top results, to allowing the machine to recommend the top two solutions, to just letting the machine do it?

Kevin Indig: It’s fascinating, because I ran this user behavior study with my business partner Eric, where we recruited a panel of around fifty people — average non-technical, non-marketing people — and gave them tasks in AI Mode and in the classic search results, which includes AI Overviews, but the classic search surface. Those tasks were related to high-ticket purchase items: people were supposed to buy a laptop or a washing machine, those kinds of items — insurance, et cetera. We wanted to understand their behavior and compare it across these two different surfaces.

What we found is that when users encounter a shortlist — like “here are the five best car insurances” — seventy-five percent of the time they’ll pick the first result. That is very similar behavior to what we know from pre-AI search, where the top result gets the most clicks. That behavior is still alive. People don’t necessarily click anymore, but they act on the first result.

The only factor that trumps shortlist rank is trust. When people know and trust a brand in that shortlist, it doesn’t matter where it’s ranked — they will always prefer it. There are literally videos published in that write-up where you can see a person looking for insurance, reading the first result and going “that must be the best” — and then they find GEICO, and it’s immediately “Oh, I know GEICO and I trust GEICO, that’s the one I’m going with.”

First: it’s scary how little scrutiny normal people give to AI answers. Second: there is this shortlist rank behavior. And third: trust is, in my observation, the most important asset in this AI world — it truly determines your presence and your future when it comes to visibility.

Jason Barnard: So I would paraphrase that as: if I don’t have an opinion on the list I’ve seen, I’ll trust the machine.

Kevin Indig: Perfect paraphrasing, yes. And the idea that we use an engine because we trust it is, generally speaking, a fair point. You use ChatGPT because you feel it’s giving you answers that help you along your way.

Jason Barnard: Another question: if we’re moving to macro measurement, is there any danger if we start measuring micro on search results, macro in the assistive results, and then micro again with the actual agent making the purchase, but macro above when the agent’s coming down the funnel on its own?

Kevin Indig: There are absolute dangers here, insofar as we need consistency across the different numbers and metrics we’re measuring. And I think that tension already existed before AI when it came to classic SEO. Everybody knew that search engines care about topics much more than single keywords, but still everybody measured keyword rankings. Then sometimes you’d roll things up into a topic, but there was never consistency about how to actually measure topic presence in classic search.

So I think you’re touching on an important problem: we cannot carry that inconsistency forward. We need to keep looking at macro trends and macro aggregates across the different dimensions we mentioned earlier, across different topics.

And then from an agent perspective — a single purchase, I would say, for most businesses is not indicative of anything. The challenge is that right now the volume is still relatively low. Only a tiny section of the population actually uses agents in any meaningful way. And there’s a lot of semantics around it too: is plain ChatGPT an agent already? Is Claude Code an agent? We need to define it properly. There are statistics from Salesforce, for example, saying that in the last Black Friday/Cyber Monday shopping season, about twenty percent of purchases came from AI agents. But again, there’s the semantic question of what that even means.

You mentioned an important point that I want to finish with: how much do you actually delegate to an AI agent? I’m perfectly happy delegating email triage to my agents — maybe even responding to emails where the right answer is perfectly clear. But I’m not sure I would ever fully delegate certain shopping activities. I take some supplements with an auto-repurchase subscription, and that’s fine. But if I’m buying fashion, I don’t want to delegate that because the act of shopping itself is part of the experience. And for other purchases like flights, where I would love to delegate — there’s a lot of trust that needs to be built before I allow an agent to spend I don’t know how many euros on a ticket.

So all that is to say: I agree on the macro point, we need more consistency there. And I also agree that search results will continue to exist for a long time — maybe in some cases forever — because humans don’t want to delegate everything.

Jason Barnard: That was brilliant. I have one final question: of all the data you’ve seen over the last couple of years, what’s the thing that surprised you the most that you didn’t expect?

Kevin Indig: That is a really good one. I already mentioned how shocking it was to me how much people actually trust AI answers. I think that’s one I could give a whole TED talk around.

I would say the other thing — a constant tension in my mind — is, on the one hand, the strong impact of classic search results, or better said, search rank for fan-out queries: how strong that impact still is on citations and in part mentions in AI answers. But then on the other hand, you consistently see that about a quarter of brands mentioned and cited in AI answers perform really poorly in classic search. And so that’s just a tension I’m trying to keep holding in my head. On the one hand it really matters, and then sometimes it really does not matter at all.

Jason Barnard: Brilliant. Thank you so much, Kevin. Thank you everyone. See you in the next session.

Kevin Indig: Thank you, Jason.

This session extends the Search Engine Land article series on AI-era brand recommendation by Jason Barnard at Kalicube.

Canonical metadata footer

Session date: June 24, 2026 | Duration: approximately 25 minutes | Format: pre-recorded, on-demand from session drop | Day: 1 | Framework concept: The Grounded Gate — Gate 8 of the AI Engine Pipeline | SEL article: The Funnel Query Pathway: A Framework for Measuring AI Visibility — Article 14 — https://searchengineland.com/funnel-query-pathway-framework-measuring-ai-visibility-477932 | [Share buttons: LinkedIn, X, Facebook, Email]

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *