SEO or GEO audits fail without these 3 essentials

Running SEO and GEO/AEO audits is an excellent use case for AI, especially with recent models that have agentic capabilities. These models have extensive knowledge bases and can perform multistep processes, such as extracting webpages, reviewing data, and formulating recommendations.

But before running your SEO or GEO/AEO audit in Claude or ChatGPT, have you considered whether the model has everything it needs to provide a good response?

You might be shocked to discover that state-of-the-art models provide detailed recommendations without the most basic information, including Google SERPs, keyword volumes, or even the ability to fetch URL content. For example:

As CEO of an SEO/GEO agency in the B2B tech space, I receive AI-generated audits from clients and prospects daily. I call them “naive audits.” They appear detailed and impressive, but they quickly fall apart when you ask basic questions:

• What was this based on?
• Where did the AI get this data?
• What is the methodology?

I’m writing this article to help explain the gap between what you expect from an AI-based audit and what you get in practice. My goal isn’t to discourage you from using AI for these purposes — you definitely should. Rather, I propose a framework with three elements to help you get AI audits right, ensuring they’re rooted in reality and provide real value.

‘Naive’ SEO audits: What can go wrong?

Let’s take a simple example: generating SEO recommendations for an existing blog. This should be an easy task for an advanced language model.

I took a blog written by one of our clients about shortages in the flash storage industry. It’s a timely topic, and the article could probably attract traffic if optimized properly.

Here is the prompt I gave Claude Opus 4.7 (with Adaptive Reasoning).

Claude thinks for a bit and provides a detailed 1,600-word report with everything we should fix in the article. See the full response here. I’m showing the highlights below. Looks promising!

But here’s the first warning signal: Claude hints that it had to infer the current structure of the article. Let’s check that out. I asked whether it had actually read the article:

Surprise 1: Claude couldn’t actually read the article. It didn’t admit that until I asked. Instead, it relied on search snippets. The entire analysis was based on a few snippets rather than the full content, which means many of the findings were probably irrelevant.

Now let’s take it one step further. I asked Claude to come up with a main keyword, and it suggested [intelligent data tiering]. Does that keyword actually have search volume?

Surprise 2: Claude doesn’t have access to search volumes, and when thinking about it, it admits the keyword is unlikely to have volume. Here is what Semrush had to say.

So the results were based on “inferred” content of the page and a keyword that isn’t really searched by anyone.

But let’s take it one step further. Let’s say this is the right keyword. Did Claude actually have the top SERPs for this keyword?

Surprise 3: Even assuming the keyword is fine, Claude didn’t really have the top 10 results. Again, it “inferred” the top results from related searches.

I went out of my way to help the model. I grabbed the top 10 results and provided them. Then I asked whether it would actually read them.

Surprise 4: Even given the top nine URLs, Claude was only able to read five of them. The rest were blocked and not available for retrieval.

Fun fact: In our testing, AI chatbots typically retrieved only 30% to 40% of the URLs we provided due to various technical issues. This is why it’s important to pre-scrape URLs or use a specialized library to access page content.

What went wrong in this example?

Claude couldn’t fetch the page’s content. It generated a keyword with no search volume and based all the recommendations on it.

Even for this keyword, Claude didn’t have a good way to grab the top 10 SERPs. And even when I handed it the top 10 SERPs, it could only retrieve about half of them.

Two more points to think about:

• Imagine the prompt wouldn’t mention a main and subkeyword. The audit would just be a list of generic SEO guidelines, like adding a title and meta description, without even attempting to focus the page on an actual query. This is why methodology matters.
• The report is way too long and detailed. Even assuming it were valid, few writers would spend the time to read a 1,600-word analysis that’s longer than the post itself. Even if they read it, they probably wouldn’t have the time to implement it.

This example isn’t sensitive to the blog post or model used. Feel free to try it yourself on state-of-the-art models from OpenAI, Google, etc., with content from your own site.

Ask for an audit and push back on the findings. You’ll be shocked at the responses.

Most importantly, this is a simplified example. Imagine what happens when you give AI an open-ended task to audit your entire site, review all the URLs, and help you grow traffic. I’ll leave it to your imagination.

How to do it right: Building a page audit agent

When we build an agent that provides SEO content recommendations, we build a self-sufficient agent that includes the following steps to make sure the output is robust:

• Pre-scrape the content of the page for analysis and provide full HTML to the model.
• Use a keyword research skill connected to a keyword tool to identify keywords with real volume related to the content, and allow the user to verify the keywords.
• Pull the top 10 URLs for the selected query from the same keyword tool.
• Pre-scrape the top 10 URLs and hand the full HTML to the model.
• Use an outline skill to build an ideal outline of an SEO article for the keyword and compare it to the actual blog outline.
• Instruct the model to provide minimal, bite-sized recommendations. Writers are busy, so we need to give them brief guidelines with minimal changes to align the post with the ideal structure.

Here’s the output of an agent we built for the same task:

This is accurate, insightful, and actionable for a human writer — only 350 words of concise, focused guidance.

A world of difference from the “naive audit,” which basically guessed its way through the task, provided invalid recommendations, and wrote a 1,600-word thesis explaining them.

‘Naive’ GEO/AEO audits: What can go right?

In the previous section, I showed how a simple SEO task, which should be easy for an AI model, can go horribly wrong. If this is the case with SEO, where there are established practices going back two decades and lots of authoritative material for AI to learn from, what would happen with a GEO/AEO task where there are no established practices?

Here are a few basic issues to consider when you ask a language model to audit your site for GEO/AEO:

• What material do AI engines use to learn about GEO/AEO? There’s a deluge of speculative information by so-called experts and, worse, AI-generated articles with wildly hallucinated info. Lily Ray called this the AI slop loop — AI generating info on how to optimize for AI, which is then regurgitated by AI.
• Are there studies based on data or experimentation? Most of the info about GEO/AEO, even content written by experts, isn’t backed by hard data, and there are very few recorded experiments. As an anecdote, I tried to find research proving the common notion that FAQs improve AI visibility. Guess what? There’s none.
• Some best practices are actually harmful. Many of the so-called “best practices” in this field can be very harmful to your organic presence, again referencing Ray: “Your GEO strategy might be destroying your SEO.” Some of them could even hurt your AI presence. Following AI-generated recommendations in this field can be dangerous.
• AI isn’t self-aware. It can’t tell you how to optimize for itself. This is a fallacy many people subscribe to: “I can ask Claude how to optimize my site for AI visibility, because it’s Claude, it knows what it likes.” But this is fundamentally wrong. AI engines aren’t self-aware. Claude needs a “Claude capabilities” skill because it doesn’t even know its basic features, not to mention its inner workings and how to optimize for it.

Everything these engines say is based on their world knowledge, which, in the case of GEO/AEO, is extremely lacking. If you’re going to the moon, ask someone who has actually been there, not someone who has read AI-generated content about it.

People like me, who are actively working on GEO/AEO projects with clients, are engaged in active experimentation and are learning by doing — what works and what doesn’t. We didn’t learn how to do GEO/AEO from blogs. Many of our unique learnings aren’t written anywhere, so AI engines don’t know them. You really need to become an expert, or engage with an expert, to succeed in this field.

Does this mean you can’t do GEO/AEO audits with AI?

You can. But you need to have a solid understanding of what actually works in GEO/AEO and what to avoid.

It’s simply dangerous to blindly follow AI-generated guidance without a strong understanding of the field. Once you have a sound methodology, you can use it to build an agent that performs effective audits at the page or site level.

AI is a great execution tool, but please don’t use it to learn how to optimize for AI engines.

---

The CaML framework: 3 things AI needs to provide a useful audit

Meet CaML, my new recruit. He’s a friendly AI agent that gets a lot of stuff done for us. But the important thing is he has everything he needs to succeed, abbreviated as C, M, and L:

• Context/da.
• Methodology.
• Human in the loop.

I think the camel metaphor is fitting because a camel is self-sufficient — you can send it into the desert for weeks, and it has everything it needs to survive. Send a donkey into the desert and it will die.

• A naive audit is like a donkey — it fails because it doesn’t have the basic things it needs for its desert mission.
• A properly built AI agent is like a camel, which is completely self-sufficient and can survive and thrive on the hard road.

C: Context/data

The first thing your SEO/GEO agent needs is context and data to make the right decisions. Most of the problems in our example fell into this category.

Make sure your agent has:

• The basic info required to analyze: Typically, crawl data and webpage content. Don’t ask the agent to get this stuff on its own. Prepare it yourself and provide it, or build an MCP server that lets the agent grab exactly what it needs.
• Relevant SEO metrics: SERPs, keyword volumes, keyword ranks, clicks, impressions, sessions — anything relevant to the task at hand. Connect the relevant tooling to your AI via MCP (quite easy nowadays) and instruct it to use relevant tools for each task.
• Relevant GEO/AEO metrics: Solutions like Profound, Semrush AIO, or Ahrefs Brand Radar provide critical info like visibility of your brand and competitors for important prompts.
• Operational data: Do you have an SEO/GEO audit task board or ticketing system? Let the AI agent access it. This way it will know what you’re already working on, which fixes you were unable to perform for whatever reason, and avoid suggesting the same tasks again.
• Business context: It’s useful to tell the agent, as part of its instructions, any important context you would share with a junior employee. For example, the size of your organization, approval process for changes, or technical infrastructure of your website.

M: Methodology

There are many ways to do SEO, and even more approaches to GEO/AEO. Don’t let the agent randomly pick its methodology. Make sure you have a solid understanding of how to perform the task at hand.

If you’re not an xEO expert, follow real experts online, read their stuff, and understand the best approach to carrying it out. If you’re just getting started, take a good SEO course or read one of the classic books (my favorite is “The Art of SEO”) to get your bearings.

Once you have your methodology down, guide the agent on how to do the work:

• Define the work process: For example, for a page audit, read the content, find a main keyword, approve it with the user, read the top 10 results, then proceed with recommendations.
• Specify data sources and decision support: Tell the agent specifically what data it should use and how to use it to make decisions. For example, in our page audit agent we tell the agent to first brainstorm keywords on its own, then check search volumes, then find more variations (same process a human SEO would follow).
• Think about the end user: An AI agent doesn’t operate in a vacuum. It typically provides recommendations a human will need to review, approve, and implement. Think about these humans — what would make the recommendations useful and actionable for them?
• Update the process as algorithms change: SEO, and especially GEO/AEO, aren’t static. Whenever Google releases a Core Update or a new AI model launches, learn about the changes, experiment if possible, and update your agent’s technique. Otherwise, it will quickly become irrelevant.
• Create guardrails: Agents can go berserk. Carefully limit the agent’s operation so it can’t do any damage. I recommend never letting the agent update your site. An SEO agent should generate recommendations, and web or content teams are then responsible for implementing them. If they use AI for implementation, fine, but this is their separate responsibility. Also, think about token usage — an improperly designed agent could run up thousands of dollars in costs.

L: Human in the loop

This is possibly the most important element. Even with today’s super-advanced models, you can’t trust them to make the right decision every time. They will make mistakes, miss relevant context, hallucinate, and often run into technical issues. You need a human in the loop (HITL) to validate every agent decision

What is needed for human in the loop:

• Make your agent explainable: The agent should not only spit out recommendations. It should explain to the user how it arrived at those recommendations. The explanation must be brief — this is critical, otherwise humans will ignore it and move on. Agents can churn out tasks at scale, and no one has time to read pages of analysis.
• Create a process for reviewing tasks at scale: For example, we have a central recommendations board in Asana. Agents create tasks directly on this board, they are reviewed by an expert, escalated to a senior expert if needed, and only then handed over to customers.
• The human must have relevant expertise: If the task is related to content, the person reviewing it should be an editor or writer. If it’s related to SEO, it probably should be an SEO expert. If you’re a small operation or a one-person team, make sure the people reviewing tasks are reasonably knowledgeable about what the agents are doing.
• Use the feedback loop to improve the agent: When you review tasks, you will see recurring issues or edge cases where the agent fails or gets things wrong. Go back to the agent’s instructions and tweak them to solve the problem. It’s important to do this carefully. If you overadjust, you might create other side effects.

Tip: I recommend this prompt engineering course by Andrew Ng (old but still relevant), which covers the basics of tuning AI engine instructions.

Beyond the basics: What SEO pros can add to an AI-generated audit

Let’s finally address the elephant in the room: If you can run complex SEO/GEO/AIO audits with AI, why do you need an SEO professional on your team, or why should you hire a consultant or agency?

In light of the previous discussion, I think it should be obvious that SEO pros can add tremendous value in an agentic AI environment.

Strategy, direction, and guidance

It starts with building the agent and reviewing its output, but goes far beyond that.

Which agents should you build? What should you be focusing on? What are the main problems preventing your site from growing in traffic or AI visibility?

Having an expert on your project can guide these decisions and help you understand what can move the needle. Then AI can help execute.

Think of an SEO expert as the North Star of AI workflows: identifying growth issues, devising solutions, and designing AI systems that can execute them.

Unique analysis

SEO and GEO/AEO are highly dynamic and getting exponentially more complex. Google updates its algorithms, and AI systems launch and update models daily.

As an agency, we have to constantly innovate and find new techniques to help our clients. This means reading studies, carrying out our own studies based on customer data, and performing experiments to see what works.

This is where real SEO professionals shine. The cliché that says “AI will free up humans to do more strategic work” is really true in this case.

Instead of crunching spreadsheets and reviewing URLs until their eyes bleed, SEO experts can come up with new breakthrough techniques that will help you lead your industry.

For example, based on the results of the last Google Core Update, we’re developing a new AI Overview optimization technique. When we have it, we’ll build it into our next generation of agents, use them to optimize our clients’ content at scale, and help them get more traffic from AI Overview citations. No AI system can provide this value, at least for now.

Measuring results and updating AI workflows

The most important part of search engine optimization, and the same holds true for GEO/AEO, is measurement and analytics.

After you’ve done all the things — written content, made technical fixes, published links, or brand mentions — you need to measure whether it actually works. Did those tactics generate traffic or visibility? What worked and what didn’t?

AI can help with this analysis. But anyone with experience in this field knows that analytics and measurement are hard. It’s difficult to get the right data, make sense of it, and draw conclusions.

Most organizations have “dashboard blindness” — they stare at graphs and make decisions without really understanding what the data means and whether it’s even valid at all.

Going forward, we will need SEO experts to measure results: to collect the right data, analyze it, present it, and make the hard call. Did it work or not? What should we do differently?

Then they can build those hard-earned lessons into AI workflows to ensure they really provide value.

Building an agent-first SEO organization

As I write these words, our SEO/GEO agency is undergoing a major transformation into an agent-first enterprise. We’re building a platform of 60-plus AI agents that perform all major SEO/GEO tasks.

Our team will use these agents to provide more value to clients, and we’ll hand them over to clients so they can do their own analysis and optimization.

I’m confident this is the future of SEO agencies. Our role in providing strategy and expert guidance hasn’t changed.

But instead of doing the heavy lifting and manual work, we’ll take over the complex job of building AI systems that can do the work at scale. We’ll review their output, maintain and optimize it, and make sure they’re really delivering results for clients.

Hit me up if you want to join us on this exciting journey.