What Have the Authority Sites Tested Lately?

On 2026-10-12, Ahrefs published “How ChatGPT Does Query Fan-Out (15 Million Queries Studied).” The study shows that when a user asks one question, the model breaks it into several sub-queries behind the scenes to retrieve from. For Agent-type tools this is both good news and hard news: the good news is that it can help you record and cover multiple query surfaces; the hard news is that the boundaries of those sub-queries keep shifting.

Another Ahrefs piece, “Google Doesn’t Punish AI Content; It Punishes Bad Content” (2026-07-06), looked at 331,000 pages and concluded that what gets dealt with is bad content, not the method of generation. This is directly relevant to Agents — what an Agent does best is produce at scale, and production at scale is what most easily turns out pages with no new information.

Semrush’s “How to build an AI visibility dashboard” (2026-10-12) takes the observation route: first turn visibility into metrics you can actually look at. Put the three together and my judgment is this: an Agent belongs at the two ends of execution and record-keeping, not at the judgment end. If you want to build up a view of the tooling first, see How to Build an SEO Tool Stack.

Strip Away the Packaging: What Exactly Is an SEO Agent?

An ordinary AI conversation is you ask one question and it answers. An Agent adds two things: first, tools — for example, it can read your website, run searches, and write files; second, a loop — it can decide on its own what to do next and look at the results afterward. That is why it can run a dozen steps in a row without you watching every one.

The real difference this makes is “it can do the work,” not “it is smarter.” It will not suddenly understand your industry; it just automates the repeated clicking that used to fall to a person. Many products on the market called SEO Agents are, in practice, a pipeline that strings together keywords, content, and monitoring. Whether one is worth buying depends on how many steps in that pipeline you were going to do by hand anyway.

My own way of using it is pretty unglamorous: treat the Agent as an intern who works fast and remembers well but knows nothing about the business. Hand it the tidying, comparing, and drafting; do not hand it setting direction or drawing conclusions. Once that positioning is clear, it is hard for marketing copy to sweep you off course.

Which Four Things Can It Really Do for You?

Of these four, I think the most valuable are comparing and recording. They are dull, repetitive, and error-prone — exactly the machine’s home turf. Drafting comes third, because no matter how fast the first draft is, the facts still have to come from you. For how the relevant tools divide the work, see How AI SEO Tools Divide Their Uses.

If you are not yet at the stage of using an Agent, the most ordinary AI chat can handle most of these four things too, you just have to feed it step by step by hand. A tool only strings the steps together; it is not a different species.

Which Three Things Can It Not Do for You?

Where to Focus

Which category to work in, which kinds of keywords to go after, which pages to rescue first — these decisions rest on your judgment about capacity, lead times, profit, and customers. An Agent cannot see any of that; it can only see web pages. We tore down 12 competitors across 53 keywords and arrived at “the number of winnable keywords with no strong site in the top ten is 0.” Behind that conclusion is an understanding of how buyers make purchasing decisions, not something a web page can compute.

First-Hand Facts

Why rework happens, what share gets inspected, how peak-season delays are handled — only you know these. If the Agent does not know them, what it writes is a generic template. In How to Write Answer Paragraphs That Can Be Cited I keep making the point that the facts must be supplied by a person.

Business Judgment and Responsibility for Results

Which inquiry is worth following up, which customer is a bad customer, whether this investment is worth continuing — these are human work. A tool can give you more data, but it cannot take on responsibility for you. This is also why I have always objected to putting a “visibility score” straight into a report; see How to Choose an AI Visibility Tool for where the boundary lies.

There is also a practical issue: the quality of an Agent’s output depends on the quality of the input you give it. Give it a messy website and it will give you a messy list; give it a clear fact table and only then can it write a usable first draft. So whether or not to use an Agent comes down, in the end, to who holds information others do not — which is the same thing as the content-layer judgment in the earlier pieces. One more warning: connecting an Agent means handing over access to your website or data. Before signing, get three things clear: what it can read, whether it will change pages automatically, and whether changes can be rolled back. If you cannot get those three answered clearly, the time you save may turn into rework.

What Our Own Tests Warn Us About

In our T0-to-T1 panel, the prompts are fixed and each round has 72 answers. What actually moved strict mentions from 3/72 to 6/72 was not that we sped up content production, but that we filled in several pages of answers carrying specific numbers and conditions. Service-provider mentions went from 0 to 2, and both of those appeared on “how to choose” type questions that call for a shortlist of candidates.

Looked at the other way, the overlap of cited domains between the two rounds was only 90/322. If a tool promises that you will “stably occupy AI answers,” this data set is the counterexample. The source pool keeps changing on its own; whoever supplies facts that are usable right now gets a chance. A tool can raise how often you appear, but it cannot guarantee you get picked.

The canary data is also worth keeping next to any tool-purchasing decision: 7 pages reached 7/7 indexed only at D+14, and only 2/7 had impressions. Between output and results stand several gates, and a tool mainly acts before the first gate. The full record is in Test Research and Method Notes.

A Low-Cost Implementation Checklist

Three steps. Run them first with free or existing AI tools; there is no rush to buy.

1. Have the Agent do one tidying job: export your whole site’s page titles, descriptions, and update dates into one table, and look through it yourself.
2. Then do one comparison: have it find duplicate-content and outdated pages, output a list, and you still make the decisions.
3. Only then try drafting: pick one page, give it three real facts you have written, let it organize the language, and check the facts sentence by sentence yourself.

After these three steps you will know whether you need a paid tool. When choosing one, also look at What Generative Engine Optimization Tools Are For and Competitor and Keyword Research, and put the tool into the process rather than at the very front of it.

The Questions Bosses Ask Most Often

Can an SEO Agent replace outsourced SEO?

It cannot replace judgment or fact supply. It can speed up execution work, but the main direction to push, the real experience inside the content, and the judgment about inquiry quality still have to come from a person.

Is there a risk in using an Agent to generate content at scale?

Yes. Production at scale most easily turns out pages that are duplicative and carry no new information, and Ahrefs’ 2026-07-06 study of 331,000 pages points to exactly this kind of content being dealt with.

What is the first thing a small foreign-trade team should use an Agent for?

I suggest starting with tidying work: copy existing page information into a table, find duplicate and outdated content, and generate rewrite drafts for old pages, rather than mass-creating new pages first.

Can an Agent help me track AI citations?

It can help. It can run records against your fixed prompts and compile them into a table, but spot-checking and judging whether the descriptions are accurate still require a person.