Two weeks to calibrate. Then it runs every month, and you only approve.
Other companies in this market advertise how many AI agents they have. The number tells you nothing, because the next one along will always claim more. What matters is which parts run unattended, which parts a person checks, and who holds the final yes. All three are written down here.
Your site is one of five things deciding that sentence.
When an engine answers a buyer question about your category, it assembles that answer from several sources at once. Your own website is one of them, and usually the smallest. This is why an agency working only on your pages can do excellent work and move nothing.
01
Your website
Specifications, prices, service areas and whether any of it is written in a form an engine can read.
02
Third party roundups
Usually one dominant page per category, quoted by most engines, and often years out of date.
Directories and listings
Carrying your product range, categories and contact details, frequently wrong.
Reviews and forums
Where your description often comes from verbatim, including a complaint nobody answered in 2024.
Crawler access
Whether the six engine bots can reach you at all. The best page ever written is invisible behind a block.
In the categories we have measured, most of the sources an engine quotes are not the client’s own website. How that is measured is on the methodology page.
Six steps, every month, for as long as you stay.
Measure
Forty conversations across six engines on a schedule, not once a month when a report is due. A description that weakens on a Tuesday is caught on the Tuesday, not three weeks later when somebody opens a dashboard.
Runs unattended
Explain
Which source moved, which competitor gained, which model shipped that week. The cause and the proposed fix are ready before anyone opens a laptop. Then a senior person reads that diagnosis and agrees with it before any work starts, and if they disagree nothing gets drafted.
Unattended, then checked
Prioritise and draft
04
Approve
Nothing leaves in your name until you press the button. Reject it, edit it, or ask for a different angle. What you change is remembered as a rule, so month six writes closer to your voice than month one did.
Publish
Live on your site, sent to the third party source that needed correcting, or posted where the engines are actually reading. Forum and community replies are posted by a named person from their own account, never by a tool.
Verify
We fetch the URL afterwards and confirm the content is live, the structured data parses and the crawlers can reach it. Only then does it enter the work log, with the before and after attached, including the times it did nothing.
Runs unattended
The machine steps run overnight. The two that decide whether this works are human on purpose: the person who checks the diagnosis, and you.
Other companies count their agents. Here is a Tuesday instead.
An illustrative example of one working day on one account.
One row on one account, from the overnight measurement to the thing waiting for approval at lunchtime. Every step is stamped with who did it, machine or person, because that is the part you are actually buying.
Row: “are Kestrel panels any good?”
ILLUSTRATIVE
03:14
Re measured all forty conversations across six engines. 1,200 samples. Two descriptions had changed since last month.
Traced the change on this row to a forum thread from 2024 that three engines have started quoting again, and flagged the phrase costing the second turn.
Your account lead read that diagnosis before any work started, checked the thread, and agreed with it. If she had disagreed, nothing would have been drafted.
Drafted two things from your brand knowledge in your voice: a reply to the thread with current lead times, and a lead times block for your site.
Edited both. The draft said lead times from eight days. Your own rule says never claim under ten, so it was corrected before you ever saw it.
Two items waiting for your approval. Nothing publishes in your name until you press the button, and the forum reply will be posted by a person from their own account.
Six steps. Three ran unattended, two needed judgement and have a name attached, and the last one is yours. An agent count is a claim. This is a receipt, and you get one every day.
The first two weeks.
Agree the conversations
We propose forty from the vertical library, you tell us which ones a real buyer would actually ask, and the set is agreed in writing. We do not pick questions we already know you win.
Together
Learn your business
One conversation of about fifty minutes where we learn your products, your real prices, your certifications and how you talk. Everything written afterwards comes from that.
Lock the baseline
Measured and recorded before a single piece of work happens, so there is always something honest to compare against. Dated, versioned, and impossible to quietly move afterwards.
Hand over the board
On a call rather than by email, so you can ask what a row means while somebody is there to answer. After that the loop runs and your only job is the approve button.
Three days where we already run boards in your sector, about twelve where we are building the question set from scratch. Which one you are is published here.