AI agents in SEO: What you need to know

2025-04-15 22:00:07

You’ve probably been hearing a lot about AI agents lately – whether in your workplace conversations or scrolling through your social feeds (hopefully both). 

While there’s no shortage of articles discussing their general benefits, there’s surprisingly little coverage on what they mean specifically for SEO – where their impact is not just significant, but amplified.

Before we dive into the two key reasons AI agents are so important for SEOs to understand (and yes, you’re probably already using them – even if you don’t realize it), let’s first get clear on what AI agents actually are.

What are AI agents?

At their core, AI agents are autonomous systems equipped with access to external tools, data, functions, and more. 

They operate with a clear understanding of an end goal and are provided with the resources needed to achieve it.

In some cases, they’re also given instructions on how to use those tools. In others, they’re left to figure it out on their own.

Rather than diving into a chart or technical diagram of a sample agenting system, I think a simpler – and surprisingly accurate – illustration can be found in one of nature’s most complex yet overlooked lifeforms: the humble ant.

Ant colony and AI agents

Imagine an ant colony: the queen, much like a master AI algorithm, sets the overarching goal. The worker ants – each equipped with their own specialized tools – are the individual agents tasked with specific functions.

Consider the parallels:

  • Queen = Agent operator: Directs and adjusts the overall strategy.
  • Worker ants = Sub-agents: Each has a specialized tool or function, whether it’s gathering data, analyzing content, or communicating findings.
  • Colony efficiency = System optimization: As ants work together, the system optimizes resources and information flow, mirroring how AI agents coordinate to achieve complex tasks.

The queen communicates the goal to each “tool,” which each ant then tries to accomplish. 

They return with their requested resource, communicate and assess their status, share information to accomplish their macro goal faster and report back. 

An overall status is reported to the queen, who communicates adjusted commands to her tools.

This is not all that different from an AI agent, other than being generally more sophisticated (though not as impressive to us, as it only sustains a species and doesn’t automatically make a stock trade 56 nanoseconds faster after catching a new trend and applying the sentiment as positive).

I’ll poorly parallel this to AI agents below.

But before I do that, let me answer why one of my assertions above is true. 

Why the impact of AI agents in SEO is multiplied many times over most other professions

I can’t think of an industry that won’t be touched by agents, at least indirectly. 

  • Lawyers will use agents to look up and summarize judgments and analyze loopholes used for their clients.
  • Software engineers will use them to assist in developing code and systems, referencing their internal docs, repos, and external knowledge.
  • Bakers will receive their ingredients through shippers coordinated using agents.
  • SEOs will use them as tools to do their jobs faster and better – as I’ll illustrate below.

On top of that, we also need to learn and adapt to marketing into agentic systems.

Generative engine optimization (GEO) entered the scene not that long ago. 

But what it is evolving into is something different — something far more powerful. 

Something that takes us past optimizing for an algorithm, even one driven by an LLM like AI Overviews or ChatGPT, and into optimizing for agents, their functions, and their tools.

We’re seeing this evolution in its toddler years right now, and if you’re on the ground floor, that’s a great place to be. 

While there are exceptions, for the most part, generative engines are performing a lot like search engines in their presentation of solutions.

  • The user enters a query.
  • The user receives a reply.
  • That reply might have a few links in it.

Sure, the system might check on the web for additional references outside of its current knowledge base, but nothing revolutionary. 

Again, it functions a lot like traditional search with a better user experience. 

I expect the next steps in this evolution will be gradual, as tools like Google and ChatGPT add new capabilities – such as the recently announced feature where an AI-driven system can call a store to gather additional information for you.

However, new pieces will gradually fall into place until we reach a point where providing your agent with insights into your goals or needs will trigger actions in ways we likely can’t fully understand yet.

Here’s a simple example.

You give the Google agent (for example) your goal, want, or need. 

Let’s say you need new shoes for a wedding. The agent can then:

  • Check your calendar for the wedding date.
  • Check the weather in that city on that date, or likely weather based on the time of year if specifics are unavailable.
  • Ask what you’ll be wearing.
  • Knowing your size, general style, and preferred brands and stores – source options that will arrive in time for the wedding.
  • Source and store a local backup, in case something goes wrong with the delivery or fit, to have that information ready in case it detects a problem.
  • Ask if you would like to see the options:
    • If yes, send them to a display of your choosing.
    • If not, move on to the next step.
  • Once the shoe is selected, complete the order.
  • Check what other common items might be needed for weddings, based on your status at it (guest, best person, bride or groom, etc.), and optionally send an email list of these to you if it doesn’t have evidence these are completed.

Imagining this world, I have a couple of questions for you:

  • How do you attribute that to Google?
  • Was it their crawler that surfaced the information to them? What kind of optimization does that take with LLMs?
  • Was it a product feed through Google Merchant Center?
  • Did they use an operator to navigate your site to get to it? Is there optimization you need to apply to filters to simplify that?
  • If you sell umbrellas, how do you ensure you’re part of those emailed suggestions from earlier in the event that it’s going to rain.
  • Oh, and how do you even get attribution for that?

This simple example highlights the immense complexity of what lies ahead. 

New technologies will emerge that companies and teams will need to adopt and optimize. 

Additionally, with the development of new protocols like Anthropic’s Model Context Protocol (MCP), adding your store’s feed to a marketplace – or even creating your own tools for other agents to use – will become much easier. 

This opens the door to greater distribution, though it may come with challenges like difficult attribution and untested effectiveness. 

The question is: 

  • Do you really want to wait and see if your competitors dive in first, or will you seize the opportunity now?

While I can’t predict the exact shape of the marketing world in the next two weeks, let alone a year from now, I can confidently say that we’ve already entered the agentic era. 

The rate of adoption and development in this space is unlike anything I’ve seen in over two decades of online marketing.

It’s even more disruptive than the changes brought on Google’s Panda and Penguin updates.

Dig deeper: From search to AI agents – The future of digital experiences

SEOs and GEOs use agentic AI, too

And on the other side of the coin, we also have SEOs using their own agentic systems.

As an example, I’ll share an agenting system I created to help generate article outlines for authors at Weights & Biases. 

What started as a simple replacement for a script I had previously written for the same task has since evolved. 

I’ll also highlight a few upcoming expansions to better illustrate the potential of AI agents.

This agentic system begins by asking the user for five things:

  • The primary phrase they are hoping to rank for with an article.
  • Any secondary terms.
  • The type of article they were writing.
  • The title (if they have one in mind).
  • The author.

It uses this information to inform the other agents within the system what to do and what data to access.

I’ve created several agents and data sources for the agent to access. 

The main ones (including a few still being finished after some testing) are:

A search agent

This agent has access to Google search and removes social platforms, which tend to block our web scrapers.

An analysis agent

This agent does a few things:

  • Extracts the entities from the pages using Google’s Natural Language API.
  • Summarizes content.
  • Extracts questions from the content.

I’ll likely separate these into their own agents as I expand the capabilities, but combining them works well in the current iteration.

A data store of examples

For each author, I created a folder with 10 markdown files that include:

  • The inputs they provided (primary phrase, secondary terms, title, etc.).
  • The outlines generated by the system.
  • The final outlines I handed off after manual editing.
  • The first paragraphs from the published articles, based on my criteria for how section intros should read.

This collection trains the agentic system to understand each author’s preferred structure and tone. It also helps suggest first paragraphs that align with their writing style.

I log all of this – inputs, extracted entities, questions, and outlines – to W&B Weave to monitor performance and guide improvements.

An outline agent

This agent takes in the information from the user, the search results, entities, questions, and summaries and generates an article outline.

Coming soon

Some agents I’m adding in presently are:

  • A keyword agent that will have access to the Google Ads API to get additional keyword ideas and search volumes.
  • A social listening agent that will monitor social channels for trending topics and auto-generate and outline when one crosses a threshold of likely importance.
  • A Slack/email agent: When an article outline is generated automatically, the agentic system will inform me – including a list of notable people talking about the topic and a summary.
  • A competitor agent that will check to see if known competitors are ranking for the content and send them to me with the outline.

I’m sure there’s more to come. (I considered waiting until everything was finished before writing this, but new ideas keep popping up, and this article would never get written.)

You should (and can) build agents too

I’m not alone in developing agents, and while some SEO tools claim to be agentic, I haven’t found any worth paying for yet. 

The real benefit of building agents is that they help me understand the environment I’m marketing in. 

If you want to try developing one, I’ve used obot.ai, which is simple and great for creating basic, useful agents for various tasks.

Big thanks to Marc Sirkin, CEO of Third Door Media, for introducing me to it. 

At the very least, it’ll give you a feel for how agents work, which is a big advantage over competitors who don’t understand what’s happening behind the scenes.