SEO agent: what AI agents can and cannot automate

An SEO agent is an artificial intelligence system that works with a site’s data and completes a whole task without anyone dictating each step: review the site, compare it with what people search for, work out what is missing and leave it ready for a person to approve.

A traditional SEO tool delivers a report and waits for someone to read it. An agent reads that result, decides what to do with it and keeps going until the brief is finished.

Anthropic, the company behind Claude, draws the distinction in its guide “Building effective agents”, published on 19 December 2024: in an automated workflow, whoever programs it decides every step in advance; in an agent, the system itself decides the steps and which tools to use.

Hiring an agency and using an agent are not the same thing either, even though one search covers both. An agency brings judgment and answers for the result; an agent brings the capacity to execute. On DoGeo projects the two work together, and what gets published is signed off by a person.

Three traits set an agent apart from classic automation:

• It works towards a goal rather than a closed list of instructions: it is told what result is expected and chooses how to get there.
• It uses tools on its own: it reviews pages, reads search data and proposes changes.
• It leaves a trail: everything it proposes can be read, accepted or rejected before it reaches your site.

An SEO agent works from four sources:

• The site itself: which pages exist, what each one says and how they link to each other.
• Google Search Console: which searches each page appears for, how often and with how many clicks, and whether Google has it indexed.
• Web analytics: who arrives, from where and what they do next.
• AI assistants: whether they name the brand when asked a buyer’s questions, and which sources they rely on.

Google Search Console is queried through its API, and its limits set the pace of the work. URL inspection allows 2,000 queries a day and 600 a minute per property, according to the Search Console API usage limits page, last updated on 28 August 2025. An agent checking the status of tens of thousands of URLs has to spread the work over several days and start with the pages that have changed.

Performance data has a limit of its own: each API query returns at most 25,000 rows, and 1,000 when no other number is set, according to the Search Analytics reference last updated on 11 August 2026. An agent that does not request results in batches works from a sample without knowing it.

Anthropic gives the underlying reason in “Building effective agents”: an agent’s autonomy means higher costs and errors that compound from one step to the next, which is why it recommends extensive testing and guardrails before letting one work alone. In SEO, the most useful guardrail is that no page ships until a person has approved it.

HeyTalent has 600 pages indexed across several languages, published by its content system, according to HeyTalent’s own data. In Google, its clicks went from 69 in March 2026 to 679 in August, 9.8 times as many, rising every month in that stretch, according to HeyTalent’s Google Search Console.

In Google’s AI-generated answers, the week of 8 to 14 September 2026 brought 9,500 impressions, 6.4 times the first week measured, according to the generative AI report in HeyTalent’s Google Search Console. These are times the site appeared, not clicks or citations, and the report is still in beta.

In AI assistants, the AI performance report in HeyTalent’s Bing Webmaster Tools counted 14,200 citations over three months, averaging 19 cited pages. And the first customer came early: HeyTalent took just under a month and a half to close the first one who said they had found it through ChatGPT.

These figures come with their limits. Appearances grow much faster than clicks, which fits answers that people read without leaving the search page, but does not prove it. And a publishing system explains the volume of pages, not the commercial result of each one.

Across DoGeo’s projects publishing content at scale, between 60% and 70% of the pages Google knows about end up indexed, that is, eligible to appear in its results. That is a reference range from two projects rather than a market standard, and it does not mean the rest are bad pages: Google leaves pages out for many reasons.

Appearances are shared out even more unevenly. At Kulturo, according to Google Search Console, the calculators account for 54.9% of per-page impressions in Google’s AI-generated answers. At HeyTalent, according to Google Search Console, a single page answering a pricing question accounts for 29.8% of per-page impressions, and the top three for 56.1%.

That split is a correlation and not a rule: pages carrying information nobody else has concentrating the appearances does not prove that publishing them is the cause. What does follow is the reverse: filling a site with pages that repeat what already exists will not change the split, and it will fill Google with pages nobody has a reason to cite.

Publishing without review also carries a risk with Google. Its spam policies for web search, last updated on 28 August 2026, define scaled content abuse as generating many pages primarily to manipulate search rankings, and their first example is “Using generative AI tools or other similar tools to generate many pages without adding value for users”.

For Google, what counts is purpose and value, not the tool: its guidance on using generative AI content, last updated on 10 December 2025, asks for accuracy, quality and relevance, including in the titles and descriptions people see in the results.

1. AI assistants must be able to read it. Many sites block assistants’ crawlers, such as OpenAI’s GPTBot or Anthropic’s ClaudeBot, without knowing it, from robots.txt or from the CDN’s firewall, while letting Googlebot through: they show up in Google and no assistant cites them. DoGeo has a free tool that checks this without asking for an email.
2. The text must be in the HTML the server sends. If part of the content only appears once JavaScript runs in the browser, some assistants never see it, because they read the server’s first response and do not wait for the page to finish assembling.
3. There must be something of your own to say. An agent cannot invent a measurement: the only thing competitors cannot copy is the data your company has measured and can publish.
4. Someone must review and sign off. Without that person, a system of agents does not speed up publishing; it speeds up publishing mistakes.