We recently ran a live, 2-day intensive AEO cohort.

People loved it, and since the next live cohort isn’t until October, we decided to make the course available on demand.

So, if you want to understand how AI search works in 2026 and what you can do to improve your visibility across AI search engines, you can now go through the same material at your own pace.

You’ll get access to the recordings, resources, and frameworks from the live cohort.

The course is $499 and already has a 5-star rating from people who joined live.

If you want to get up to speed with AEO without waiting until October, you can buy it here (use ‘growthwaves10‘ for a 10% discount):

Watch On-Demand

Most teams tracking AI search focus on one question: are we mentioned?

They open ChatGPT or Perplexity, type “best project management tool,” and count how often they appear.

Roughly speaking, that’s visibility tracking.

Sure, it matters. But it answers only one question. (The easy one!)

Here’s a harder one:

How does AI describe you? When a buyer asks about specific features or integration depth, what does the model say about your product?

That is perception. And most teams are not measuring it.

Read on if you want to get an edge.

Two prompt types

There are two types of prompts worth tracking in AI search.

1. Topical Prompts

They help you test whether you show up. They cover your category and use case.

“What is the best email marketing platform for e-commerce?” is a topical prompt. You track these to check if your brand appears, how often, etc.

2. Evaluation Prompts

They help you understand how you are being perceived.

They dig into specific capabilities and limitations buyers consider. Is AI’s perception of your product accurate?

Most teams only run topical prompts.

That tells you if you are in the room.

Evaluation prompts tell you what the room says about you.

Here’s how evaluation prompts work

Every evaluation prompt sits inside a three-level hierarchy:

A Topic is a broad area.

For email marketing, that could be deliverability or automation.

For a CRM, it might be pipeline management or reporting.

An Evaluation Attribute is a specific, measurable criterion inside that topic.

Under deliverability, the attributes could be inbox placement, bounce management, authentication support or sender reputation tools.

For each attribute, you write one Evaluation Prompt using a consistent template:

Evaluate {Brand} on the {platform/product/solution}’s ability to {specific evaluation criterion}, {desired outcome or business impact}.

The template stays the same every time. What changes is the attribute and the outcome.

(Yes, it reads like a rubric. That is the point.)

An example

Say you sell email marketing software and deliverability is the top buyer concern. Here is how you break that single topic into seven evaluation attributes, each with its own prompt.

Run each prompt across ChatGPT, Perplexity, Google’s AI Overviews, etc. Do this for every attribute that shapes a buying decision.

The result: a clear read on whether AI describes your product the way you need it described.

Things to keep in mind

Before I leave you, here are a few things to keep in mind when it comes to Evaluation Prompts:

1. Evaluation prompts measure perception, not visibility. This is about how AI describes you, not whether it mentions you.

2. Audit the cited sources alongside every answer. Every URL AI attaches to a response is either something you can fix or a content gap you should fill.

3. Track sentiment and sources together. Perception is the symptom. The sources AI pulls from are the cause. (Fix the source, fix the perception.)

4. Expect platform variance. Different engines pull from different retrieval sources. The same prompt will produce different answers depending on the platform.

5. Negative answers are an opportunity. Every “doesn’t support Y” or “limited capability in Z” points to a specific source you can correct or content you should create.

Let’s wrap this one.

Final Thoughts

Most AI search measurement stops at counting. Mentions, Share of Voice, rankings, etc.

Useful numbers. But they are the floor.

Evaluation prompts go deeper.

They reveal how AI search engines perceive your product on the specific criteria your buyers weigh.

When that perception is wrong, they show you which sources to fix.

Keep that in mind next time you are building your AI search tracking system.

Thank you for reading today’s note, and see you again next week.

Research Disclaimers and Limitations

GrowthWaves and its author are not sponsored by or compensated by any company mentioned in this note. This is independent editorial analysis and does not constitute investment, financial, or legal advice. The author may have relationships with, work with, or hold equity in companies referenced; however, no content in this piece was influenced, commissioned, or incentivized by any such relationship. AI tools were used as a research assistant in the preparation of this piece. All claims are sourced and linked throughout.

Sources

• SparkToro, “NEW Research: AIs are highly inconsistent when recommending brands or products”

• GrowthWaves, “Perception Deviation: The most important metric you’re not tracking”

I’ve been using Ahrefs for years. That hasn’t changed.

What has changed is how we track visibility.

Now, I know there are a lot of tools popping up for AI search. (We counted 200+ in a report we published.)

But, I’m only using two: one we’ve built internally (and know its limits), and Brand Radar by Ahrefs.

It’s the only external tool I trust for AI search tracking.

You can track your brand’s visibility across AI platforms like ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews, and compare it against competitors with clean, useful data visualizations.

Plus, you can track citations, your AI share of voice, and much more.

If AI search matters to you (it should), go check it out:

Check it out here

Most people think content has an AI problem.

They read studies like this one from Ahrefs, which found that 74.2% of new web pages contain AI-generated content, and they conclude:

There is too much content, and most of it is bad. (Even more so with AI.)

They are half right!

There is too much content, yes. But the quality of that content is not the real issue.

The real problem is that human consumption capacity has a ceiling, and we passed it years ago.

AI just made it impossible to ignore.

The math that breaks content marketing

Let’s try to frame this.

Content creation grew steadily, roughly linearly, for most of the 2010s.

Companies invested in blogs and content programs, and agencies scaled article production to meet demand.

Growth was real but manageable.

Then ChatGPT launched in November 2022…

Within a year, generative AI adoption jumped across organizations.

According to some studies, in 2024, the number of AI-generated articles surpassed that of human-written ones.

And with that, the consumption curve changed.

Content supply went exponential.

While human attention didn’t.

It is declining.

Mark Schaefer saw this coming in 2014.

He called it Content Shock: the point where content supply exceeds the audience’s ability to consume it.

His argument was simple.

There are only 24 hours in a day.

Even with growing internet usage and multiplying media formats, there is a hard ceiling on human attention.

He was right!

And twelve years later, AI has turned Content Shock from a theoretical concern into an observable crisis.

The second consumer of the internet

Here is where it gets interesting.

In June 2026, Cloudflare CEO Matthew Prince posted on X that, for the first time, bot traffic surpassed human traffic.

Here’s what the data looks like as of writing:

The crossover came a bit early. (Based on Prince’s own predictions.)

And let’s not forget the evolution of bots and search engine crawlers:

Agents

In plain English, an agent is a piece of software that reads and synthesizes content on behalf of users, then acts on what it finds.

According to a 2025 report by HUMAN Security, traffic from AI agents and agentic browsers grew 7,851% year over year.

That growth rate means agents went from a rounding error to a real consumer of web content in ~3 years.

This changes the consumption equation.

There are now two consumers of online content: humans and agents.

1. Humans browse, scan, read (sometimes), and make decisions.

2. Agents crawl, parse, summarize and recommend.

The question most content teams have not asked: which consumer are we creating for?

The consumption gap

No hard data behind this, but here’s how I think of the consumption gap:

The red/pink line is content creation.

Linear and (somewhat) manageable until late 2022, then exponential.

The blue line is human consumption capacity: growing slowly through the 2010s, then flattening and now declining.

The green dashed line is agent consumption: near zero until 2023, then climbing steeply.

The shaded area between the red line and everything below it is what I am calling the consumption gap.

Content that gets published and never gets consumed by anyone or anything.

That gap was narrow in 2014, when Schaefer wrote about Content Shock.

Today it is enormous, and it is growing every month.

The uncomfortable question

Let’s go philosophical for a minute.

If most new web pages contain AI-generated content, and human attention is declining, and bots now account for ~59% of web traffic, then a very large percentage of content being published right now will never be read by a human.

Not because it is bad. (Even though most content is indeed very bad.)

But because there are not enough human hours in the day to consume it.

So why is it being created?

I see four answers (plus a bonus one):

1. To justify the existence of in-house content teams. Content teams need to produce content to show they are doing their jobs. Output becomes a proxy for value. This is the worst reason to publish, and it is more common than anyone admits.

2. To justify agency retainers. Agencies bill for output. More articles = more deliverables. The incentive structure rewards volume regardless of whether anyone reads the work.

3. To feed AI agents, not humans. This reason is new, and it is legitimate. If your content is designed to be parsed by AI engines and recommended through ChatGPT, Perplexity, Google AI Overviews, or Claude, then you are optimizing for agents, not human readers.

4. To serve a specific audience with high-value content. The only answer that works for human consumption. Content designed for a defined audience at a specific stage of awareness, with a clear point of view and genuine utility. And a CLEAR distribution plan!

And now the bonus one:

It worked, and to a certain extent, it still does.

Marketers got away with mediocre content for years.

Rest assured, many of them still do, since major distribution channels such as Google and AI search continue to reward mediocre content.

The same doesn’t apply for social media, where the bar is (in most cases) higher, and the feedback loops are shorter.

In any event, there are answers.

Obviously, some of them may make some of us feel uncomfortable.

Still, they’re true, and regardless of what your answer is, you need to make a decision.

Two strategies, one decision

Companies need to make a choice:

Are they creating content for humans, for agents, or for both?

If you are writing for humans, then volume might be your enemy, unless you can maintain super high quality at scale.

But generally speaking, fewer pieces could result in higher quality, a stronger point of view, and better distribution.

That attention ceiling means you are competing for a shrinking share of a fixed resource.

If the audience is agents, the playbook is different.

Agents do not care about prose quality and whether you have strong opinions.

They care about information density, crawlability, and whether you can help them form an answer without having to consume the whole piece.

Most companies obviously need both.

Which means that we need more balanced content programs…

What a human reader values, narrative and a strong point of view, is invisible to an agent.

What an agent values, structure and machine-readability, puts a human to sleep.

Keep that in mind next time you’re having an internal discussion about content strategy.

Final Thoughts

Mark Schaefer warned us in 2014.

Not sure if he could predict ChatGPT and the paradigm shift that AI brought first and foremost into the world of bits.

Regardless, there’s so much content being created on a single day that can be consumed by humans and agents alike.

Obviously, different reasons.

Agents have computational limitations, and humans have finite time and attention spans.

So…

Content does not have an AI problem. Content has a consumption capacity problem.

The question for content teams and marketers is simple: is anyone actually consuming what you are creating? And if the answer is “mostly agents,” the follow-up: are you designing for that?

Thank you for reading today’s note, and see you again next week.

Research Disclaimers and Limitations

GrowthWaves and its author are not sponsored by or compensated by any company mentioned in this note. This is independent editorial analysis and does not constitute investment, financial, or legal advice. The author may have relationships with, work with, or hold equity in companies referenced; however, no content in this piece was influenced, commissioned, or incentivized by any such relationship. AI tools were used as a research assistant in the preparation of this piece. All claims are sourced and linked throughout.

Sources

• Cloudflare Radar, “Bot vs. Human Traffic” — real-time bot vs. human traffic split

• HUMAN Security, “2026 State of AI Traffic & Cyberthreat Benchmarks” — agentic AI traffic grew 7,851% YoY; bots account for 57.4% of web requests

• Mark Schaefer, “Content Shock” — the original 2014 essay coining the term

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more

Read more