Brief IA
›

SEO Agency: Ahrefs and Semrush Ramp Up Demonstrations

Models & LLM·Tom Levy·

SEO Agency: Ahrefs and Semrush Ramp Up Demonstrations

SEO Agency: Ahrefs and Semrush Ramp Up Demonstrations
Key Takeaways
1Ahrefs describes an agent that prepares drafts for five articles each month, with human approval before publication.
2Semrush presents eight SEO processes based on connected data.
3Commercial demonstrations do not prove either a traffic gain or the reliability of a standalone publication.
Why it matters : the agentic SEO can prepare or execute changes, while the editor remains responsible for the published content.

In 2026, demonstrations of AI agents applied to SEO are multiplying, particularly at Ahrefs and Semrush. However, these presentations do not demonstrate either an increase in traffic or the reliability of standalone publications. For publishers, permissions and result verification remain distinct from the announced capabilities.

The Publisher Remains Responsible for Published Content

Even when software prepares a modification, the responsibility for what appears remains with the publisher. A statistic lacking a primary source, a misinterpreted screenshot, or a fabricated address requires human verification, regardless of the provider's reputation. This control must also examine the possible consequences of the change.

The error can precede the writing: an agent may mix impressions and clicks, compare non-comparable periods, or mistake correlation for proof. If this inaccurate result is sent directly to another application, it can propagate throughout the chain. The risk increases when the agent is entrusted with long lists of pages and is given publication permission. External pages, exported files, or consulted documents may also contain hostile instructions.

One protection is to keep this content as data, without allowing it to dictate a change in permissions. Access can be limited to what is strictly necessary; each proposal should be accompanied by the relevant address, the consulted proof, the intended intervention, and a means to cancel it. Editorial verification focuses on accuracy, usefulness, and quality.

Google specifies that the involvement of AI alone does not render content non-compliant; its rules target the mass production of pages lacking added value, particularly to manipulate rankings. Declining a comparison fifteen times can increase conflicts between URLs, maintenance work, and information obsolescence. An agent should therefore be able to refrain from creating a page: consolidating two weak pieces of content or correcting an error may present more economic interest than an additional publication.

Ahrefs and Semrush Present Their Uses

At Ahrefs, the team describes a monthly data collection concerning five articles with tables. The agent then prepares working versions in WordPress and sends the preview addresses; a person must approve their publication. The company also presents uses for prioritizing technical issues, identifying declining content, and suggesting internal links. These examples pertain to its own tools, without constituting an independent measure of their performance.

Semrush outlines eight task sequences powered by connected data. Search Atlas and Quattr claim capabilities up to execution, while Contengi illustrates a pathway based on its connector, without establishing a universal benefit for SEO. Depending on the offering, the intervention can stop at advice, provide a pre-filled fix, modify the content management system, or trigger an automatic deployment. Product, integrations, and permissions determine the actual scope: detecting, preparing, and applying are not equivalent, nor is the presence or absence of a human veto.

The vocabulary itself covers distinct objects. The SEO Handbook primarily describes the use of agents to perform SEO work. First Page Sage and Promptwatch use similar terms to make a site selectable by user agents; Promptwatch refers to Agentic Search Optimization. This second approach is closer to optimization intended for generative engines, rather than executing SEO tasks.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Agents Need Data and Permissions

Having ChatGPT generate ten title proposals remains a writing aid. Agentic work can, however, combine the examination of Search Console, comparison of competing pages, and preparation of an update. Audits, keyword research, internal linking, or corrections are among its possible uses, depending on the available tools. The deliverable can be a report, a technical proposal, or a draft in WordPress.

Unlike a scheduled export at a fixed time, an agent can request additional data, abandon a hypothesis, and choose another verification. An orchestrator distributes the steps; reusable procedures, called "skills," specify the checks to be performed. This latitude does not ensure the correctness of reasoning: the instructions received and the permissions influence its behavior. Automation therefore does not guarantee either a correct explanation of a drop in audience or a better ranking.

An isolated model does not know which articles actually lost clicks the previous day. It must access authorized sources: Google Search Console, audience metrics, crawl data, position tracking, or the content management system inventory, the CMS. The Model Context Protocol, or MCP, and application programming interfaces, APIs, can facilitate these exchanges. Ahrefs presents the agentic environment, connectors, and procedures as complementary; Semrush also highlights MCP for its SEO research and analysis.

However, these connections do not guarantee that the data is exhaustive or that each operation is permitted; multiplying tools does not ensure accuracy either. Consulting performance does not require publication permission, and preparing a draft does not impose the ability to alter redirects or the robots.txt file. A cautious approach is to start with limited access to consultation before granting restricted, recorded, and revocable permissions.

A Limited Trial Distinguishes Proposals and Corrections

A trial protocol can focus on about fifty URLs, with a single task: duplicates in titles, broken links, or competition between pages. It includes an observation period, accessible sources, a query budget, and the format of justifications. Katama recommends documenting procedures and planning interruptions before any action.

In this protocol, the first pass formulates recommendations without touching the site. The comparison with a human audit serves to identify false positives before considering only reversible changes. Canonicals, deletions, redirects, noindex tags, and administrative privileges require more stringent approval. Backups, version history, and an identified responsible party are part of the precautions against an error that could affect a hundred URLs.

Another example, purely methodological and not measured, takes an editorial site with 2,000 URLs: the agent would search for pages losing impressions over several weeks. It would correlate publication dates, queries, potential errors, and similar content. Seasonality, indexing difficulties, the arrival of a competitor, or outdated information can explain a decline; this does not demonstrate a Google penalty.

The system could then gather the sources to verify, propose the passages to update, and highlight missing internal links. The writer would retain the verification of claims and the choice of changes. A follow-up after publication would be possible, but an improvement in metrics would not be sufficient to prove that the intervention caused it.

Measurement Begins with Confirmed Anomalies

Evaluation can start with the directly observable outputs of the agent: duration of diagnosis, anomalies actually confirmed, proposals retained, and corrections ultimately canceled. It can then examine impressions, clicks, conversions, and the number of indexed pages. The commercial promise of ongoing optimization does not mean that an effect on SEO is immediately measurable or attributable.

Since 2026, Search Console also provides complementary indicators in its reports on Google's generative experiences.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.