At 6:14 on a Tuesday morning an SEO agent finished scanning Denver for plumbers. It had read the top results for 25 searches, noted which of the 140 businesses it found were paying for ads, pulled six months of traffic history on each, asked an AI assistant to recommend a plumber and recorded who it named, and scored every business on how badly search was costing it. The owner of the agency that runs it woke up to a list of eleven businesses, ranked, with a proof page and a draft email for each. Nobody had typed anything.<\/p>\n
That is the useful definition of an SEO agent. Not “an AI that does SEO”, which means nothing, but a piece of software that carries a whole job from a one-line instruction to a result a human can act on, making its own decisions about what to look at next. This article is about what that job looks like hour by hour, what an agent is genuinely good at, and the two places where it should hand the work back to you.<\/p>\n
The first thing an SEO agent does is decide who to look at. The instruction it starts from is small: a niche and a place. “Plumbers, Denver.” “Dentists, Manchester.” “CRM software, worldwide.” Everything after that is its own work.<\/p>\n
It begins by writing the searches a real buyer would type for that niche. This matters more than it sounds. A template that bolts “near me” onto every phrase produces nonsense for a national product brand, and one that bolts “services” onto a product niche produces phrases nobody types. A decent agent reads the niche first, decides whether it is a local service, a product, software or a professional practice, and writes its keyword set in that vocabulary. For Denver plumbers that is “emergency plumber denver” and “water heater replacement denver”. For car chargers in the UK it is “buy iphone car charger” and “fast iphone car charger uk”, with no city at all.<\/p>\n
Then it reads the results. For every search it records who ranks, at what position, and who is paying for the ads above them. Two lists fall out of this: the businesses that appear organically, and the businesses that are buying clicks. The overlap between those lists is the first thing worth selling to, and it is the kind of pattern a person would take a day to spot and an agent spots in a minute.<\/p>\n
With a candidate list, the agent visits every business it found. This is the part of the job that used to be a junior’s entire week. For each site it answers a fixed set of questions:<\/p>\n
Each answer becomes a signal with a strength. A dentist paying for ads on the keyword she ranks eleventh for is a stronger signal than a dentist with a slow homepage, because the first one has already proven, with money, that she cares about the keyword. The agent weighs those and produces one score per business: how broken it is, whether it can pay, and whether you can reach the person who decides.<\/p>\n
The agent is not being clever here. It is being thorough in a way that does not get tired at business number ninety. That is most of the value.<\/p>\n
A list of scores is not something you can send to anyone. The next thing the agent builds is a proof page per business: a single page that shows the owner their own numbers. The keyword they rank eleventh for and what position one takes. The traffic line going down since March. The AI answer that recommends three competitors and not them. The estimate of what that costs each month at typical customer values.<\/p>\n
This is the artifact that changes close rates. A template proposal, even a good one, closes somewhere around 15 to 25 percent. Build the same proposal from the prospect’s own data and the numbers move to 50 to 65 percent, because the first thing the owner reads is something they can check and find to be true. The agent’s job is to make that the default rather than the thing you do only for the biggest prospect of the month.<\/p>\n
The agent then drafts the opener. Four sentences. The finding, with the number in it. What it costs every month it stays broken. A link to the proof page. A small, specific ask. It does not write “I noticed your website could use some help”, because every business owner deletes forty of those a week.<\/p>\n
It does not send the email. That is a choice, and it is the right one. You read the draft, you change the ask to something that suits your calendar, and you send it. The agent has removed the two hours of research that usually stand between “this looks like a good prospect” and “I have written to them”. It has not removed you.<\/p>\n
There are two places an SEO agent should stop and wait for a person.<\/p>\n
The first is the call. Some agencies run an AI caller for the very first touch, and it works better than most people expect for one narrow job: confirming the owner is the right person, mentioning the one number from the proof page, and booking fifteen minutes. It is not the sales call. The sales call is where a plumber tells you his real problem is that his son-in-law built the website and nobody is allowed to touch it, and that conversation needs a human who can laugh at the right moment.<\/p>\n
The second is judgement about the client. An agent can tell you a business is losing 4,200 dollars a month to a competitor. It cannot tell you that the owner is retiring in a year, or that the practice is already talking to another agency, or that the “decline” is because they closed a second location on purpose. The agent’s findings are the opening of a conversation, not the conclusion of one.<\/p>\n
Everything else, it does not need you for. Rescanning the same city next month. Noticing that a prospect who ignored you in March has since lost another eight positions, and telling you that is worth a second email. Watching the AI answer for “best dentist in Austin” and flagging the week your client finally appears in it. Following up, which is where most of the pipeline dies: around 80 percent of sales need five or more touches, and 44 percent of sellers stop after one. An agent does not stop after one.<\/p>\n
The most recent change is where the agent lives. It used to be a dashboard you logged into. Increasingly it is something you talk to from the tool you already have open. SellSEO exposes its agent over MCP, which means you can sit in Claude or ChatGPT and type “find me twenty roofers in Phoenix who are paying for ads on keywords they nearly rank for, and draft the proposal for the top three”, then go and make a coffee. The scan runs, the proof pages build, the drafts appear in the conversation.<\/p>\n
That is what “SEO agent” should mean when you read it in a tool’s marketing. Not a chatbot that knows what a canonical tag is. A worker that starts from a one-line brief, reads a whole market, produces evidence and a draft, and knows which two moments to hand back to you. If the tool you are looking at cannot show you the eleven businesses it found this morning and why, it is a glossary with a chat window.<\/p>\n