👋 Hi, it’s Kyle and welcome to Growth Unhinged, my weekly newsletter exploring the hidden playbooks behind the fastest-growing startups.

I was initially skeptical AI engines would be a major growth channel. After all, it sends a laughably small amount of referral traffic. I’m now convinced it’s real (15% of new subscribers tell me they first heard about Growth Unhinged from an LLM) yet it remains woefully misunderstood.

To help, I turned to AI search expert Kevin Indig who advises leaders at Meta, Ramp, hims, Upwork, and others on organic growth. Kevin also writes the brilliant Growth Memo newsletter, which is full of research-backed insights on AI-driven search. Up today: Kevin’s framework for measuring the impact of AI search the right way.

There is no doubt AI transforms how people buy software.

Naturally, companies do their best to be visible in AI. But no one is sure about how to correctly measure the impact of answer engine optimization (AEO).

Over 40% of participants in a Growth Memo survey of 599 marketers said the lack of reliable measurement tools and attribution is their #1 AEO challenge.

The way most teams go about it right now is by measuring pipeline from referral clicks. But that’s like valuing a Super Bowl ad by QR-code scans. It under-attributes the impact.

By the end of this article, you’ll have a clear framework for measuring the impact of AEO based on how I track AEO impact with companies like Airbnb, Asana or Xero.

Users don’t click AI citations.

Pew research observed 900 U.S. adults and found only ~1% click on citations inside AI Overviews. A ChatGPT leak confirms a similar CTR of 0.69%.

On top of that, 70.6% of AI-referred traffic lands as "Direct" in GA4 with the referrer stripped. Even best-effort custom channel groupings recover only 50-70% of it.

Last-click attribution for AEO makes no sense. Focusing on traffic is the wrong call.

What about AI citations themselves? Well, those come with challenges, too:

But we know being mentioned in AI answers and at the top of shortlists truly matters. In a user behavior study, we found that users pick the first results ~75% of the time they encounter a shortlist of products!

So, the three traps teams need to avoid are:

The fix is a measurement model marketing already invented once, for exactly this problem.

Related read: Traffic is no longer a reliable growth metric

The system I use with companies like Airbnb, Asana or Xero is a ladder of leading and lagging indicators, as Andy Grove suggests in High Output Management.

Why a ladder? Because when revenue attribution is lagging, you want to know whether you’re on track or not as quickly as possible.

The ladder reflects my Retrieved → Cited → Trusted framework and divides each part into three rungs: leading indicators → quality guardrails → lagging indicators.

Before any of this work, you need a stable input:

Let’s unpack each layer of the AI visibility ladder. I’ll illustrate it with a fictional example: CartDesk, an AI help desk for Shopify brands targeting CX VPs at eCommerce brands with 20-200 support agents with prompts like “best help desk for Shopify brands with high ticket volume”.

Leading indicators tell you whether AI can find you. An answer engine can't cite a page it never crawled, so retrieval is where AEO starts.

Three signals tell you whether you're in the running.

For CartDesk, the picture is healthy:

The decision this rung drives is keep, increase, or shift the AEO work. Rising crawls and citation share mean the pages are doing their job. Flat or falling means retrieval is broken before any downstream metric gets a chance to move.

Quality guardrails exist because of one uncomfortable fact: you don't write the answer. The model decides how to describe you and where you land and how you're framed decides the sale.

You can't set that message, but you can measure it and fix the inputs it reads.

Three metrics turn portrayal into something you can watch.

Separate "mentioned" from "recommended" as you go. Getting named in a paragraph and getting put on the shortlist are different outcomes.

CartDesk looks strong on two of the three quality guardrails:

When the model mis-frames you, you fix what it reads: the positioning, the product pages, the comparison content, the PR. For CartDesk, the action items would be to ship a stronger returns workflow page and update the comparison pages so the model has something to connect the attribute to.

Related read: Is AI talking about your product in the right way?

Lagging indicators are where AEO finally touches revenue. Someone reads about you in ChatGPT, types your name into Google a week later, and books a demo.

You recover that demand by asking and by listening.

CartDesk turns this into a number a CFO would read:

For CartDesk, the monthly call writes itself. Shortlist position climbed from 18% to 32% and AI-sourced opportunities went from 6 to 14, so shift more content and PR budget into Shopify-specific proof points.

This rung drives the budget. When self-reported opportunities, win rate, and pipeline move in the same direction as your leading and quality metrics, you have a board narrative: the work at the top of the ladder is showing up at the bottom.

Related read: How to turn ChatGPT into your best pipeline source

Citation share, share of voice, and shortlist position all come out of a prompt tracker.

This is what AI visibility tools like AirOps, Profound, Semrush, and Peec are built to do: you load your frozen prompt set, they run it across the models on a schedule, and they parse each answer for cited domains, brand mentions, and the order brands get listed in. If you'd rather own the pipeline, the same thing runs as a script. Hit each model's API or services like Dataforseo or Oxylabs with your prompt set on a schedule, then parse the responses for URLs and brand names.

Sentiment and attribute match are a classification step on top of that raw output.

Sentiment is how the answer talks about you, positive, neutral, or negative, and most trackers score it for you. Attribute match is the one you define yourself. Write down the five or six attributes you want to own, then check whether the model connects you to them. An LLM handles this well: feed it the answer text and your attribute list, and ask it to tag each attribute as present, absent, or wrong. Run that across every answer in the set and you have a number that moves week to week.

For a century, we ran billboards, TV spots, and PR with no way to trace a click, and we measured them with proxies and a "how did you hear about us" field instead.

AEO is the same job with one difference. You used to write the ad — a billboard says exactly what you paid for. An AI answer says whatever the model decides. It writes your pitch, picks who you sit beside, and changes its mind without telling you. You can't control that. You can only measure how you show up and feed the model better inputs.

We spent 20 years getting so good at performance marketing that we forgot the rest of it. The half you can measure to the click is real, and it works. But the marketing that builds a brand has always been the half you can't trace: the billboard, the keynote, the Super Bowl ad nobody scans a QR code from. AEO is that half, now wearing a citation.

Wanamaker said half his ad money was wasted and never knew which half. He spent it anyway, and built one of the great brands of his century. That's the bet brand advertising has always asked you to make, and AI search just brought it back.

The companies that win inside the answer will be the ones who saw the demand was real, built the ladder to track it, and kept investing while their competitors waited for proof that was never going to load.

The click isn't coming back. The demand already did.

Other things to click on: