The Keyword

More clicks from less visibility

AI SEO case study: The Keyword

Case study ·

The problem

The Keyword is a marketing news site whose articles are written by agents: an aggregator finds the stories, an editor decides which are worth running, a researcher and a writer produce the piece, and an image generator illustrates it, up to one article a day. Publishing is the part that is solved. Being found is not.

A news page has a short, steep life. It arrives, it takes whatever search traffic it is going to take, and then it decays while everyone involved has moved on to the next day's story. The work that would earn it a second life is small and unglamorous: notice that a page collects impressions and almost no clicks, that a query sits just off the first page, that a title answers a question nobody asks. That work never competes successfully with tomorrow's publishing deadline.

There is a second front too. A growing share of the questions this site answers are now typed into an AI assistant rather than a search box, and an assistant quotes the page that states an answer plainly, which is not always the page that ranks highest.

What we built

An agent that works the site's search data once a week and makes the on-page changes that follow. Each run goes through the same steps.

  1. Step 1
    Pull the site's own search data

    Queries, pages, impressions, clicks and average position, straight from Search Console rather than from a third party's estimate.

  2. Step 2
    Compare the window with the one before

    A page that slipped, a query that started appearing, impressions that grew while clicks did not.

  3. Step 3
    Read the page as it stands

    The title, the description, the headings and the answer the page actually gives, before anything is rewritten.

  4. Step 4
    Rewrite what decides the click

    The title and the meta description first, then the heading or the paragraph that should answer the query plainly.

  5. Step 5
    Set aside what is not winnable

    A query a platform owns, a term an established brand takes whatever is written, a search whose intent the site does not serve.

  6. Step 6
    Publish into the site's own system

    The change goes into the content management system the articles already live in, under the rules it was given.

  7. Step 7
    Write the line in the ledger

    One dated line per page: what the data showed, what was done or deliberately not done, and the reason.

One run of the agent, from the data it pulls to the line it leaves behind.

Safeguards

  • It writes down what it will not touch. Of the 36 lines in the ledger so far, 12 are pages it examined and deliberately left alone, each with the reason written out. A set-aside is as auditable as an edit.
  • It changes presentation, not claims. Titles, descriptions and the way an answer is stated are its work. What the article asserts about the world is the newsroom's, and the agent does not edit it to suit a query.
  • It does not chase an unwinnable term. Queries a platform owns, searches that belong to an established brand, and intents the site does not serve are recorded and skipped rather than fought.
  • It writes for a reader first. A rewritten title has to be a sentence a person would write. A query stuffed into a line nobody would say is a stop, not a shortcut.
  • Everything it does is dated and in one file. The ledger is how the work is audited, and it is the only record either side relies on.

The rules are the approval. Work the rules cover, the agent carries out itself; anything turning on a judgement nobody has made is written down instead of guessed.

The result

Up to October 2026, month over month, visits to the site from AI tools were up 79%, and clicks from Google more than doubled on fewer impressions than the month before, so click-through rate went from 0.07% to 0.17%.

Those are site-wide figures for a site run end to end by agents, over one month. They are stated here beside the work, and no part of them is claimed as the effect of a single agent: nothing in the data separates one agent's rewrites from everything else that happened that month.

What is the agent's own, and auditable line by line, is the ledger: 36 dated lines from June 2026 to October 2026, of which 24 are work carried out across 18 distinct pages and 12 are set-asides with their reasons. A page actioned twice appears twice. Nobody on the site has had to find those pages by hand.

A line that became workactioned
What the data showed
Impressions, almost no clicks
What the agent did
Rewrote the title and description
What it left alone
The article itself
Recorded
Dated line in the ledger
A line that became a set-asideskipped
What the data showed
A query a platform owns
What the agent did
Nothing, deliberately
What it left alone
The whole page
Recorded
Dated line, with the reason
The two kinds of line the agent writes. A third of the ledger so far is the second kind.

Frequently asked questions

Can an AI agent do SEO work on a live site?

Yes, and this one does it once a week on a news site that publishes up to an article a day. It pulls the site's own Search Console data, picks the pages where a small change can move something, rewrites the title, description or on-page answer, publishes into the site's own system and records what it did.

Did the agent cause the increase in traffic?

The page does not claim that, and the data does not show it. The figures are site-wide for a site where agents also write the articles, over a single month. They are reported beside the work, and the only thing attributed to this agent is the ledger of what it changed and what it set aside.

What is the AEO part of the work?

Answer engine optimisation: being a page an AI assistant can read and quote. It is a different job from ranking, since an assistant lifts the page that states the answer plainly and early, with the specifics it needs to repeat. Visits arriving from those assistants are measured as their own figure, which is where the 79% comes from.

What does it refuse to do?

It does not change what an article claims, does not chase a term it cannot win, and does not stuff a query into a sentence a person would not write. Each of those is recorded as a set-aside with its reason, which is why a third of the ledger so far is pages it chose not to touch.

This is the agent behind our AI SEO agent. The agent most businesses start with is the AI receptionist, and how we work and what it costs is on one plain page.

Tell us what you'd like us to build.

Free Consultation