All work

Search run by agents, approved by people

An architecture and interiors studio in Singapore with three search pitches on the table and a site that couldn't measure a lead.

/work/design-studio-searchCame in under strain

System
Website, search accounts, AI agents and scripts
Ask
Paid search to bring in new projects, run in-house
Symptom
Analytics never loaded. No enquiry recorded as a conversion.

The piece A ratchet of ninety teeth held by one hot pawl: ninety days in which every change waits for a person's yes.

  • SGD 570tools and agents, six months, planned
  • 5agentseach with a job and limit
  • 90daysevery change waits for a yes

The ask

Run paid search for new projects. Three firms had pitched and the advice was to hire none and run it in-house.

The real problem

The site couldn't measure anything: analytics never loaded for visitors who declined cookies and no enquiry was recorded as a conversion.

Paid search on top of that would have spent the budget and learned nothing.

Who does the standard work

  • Idle
  • Live
  • Strain
  • Fail

BeforeAnalytics that never loaded, forms that recorded nothing and an agency package that kept the logins and would have spent blind.

The setupScripts fill one sheet, agents draft and propose and nothing reaches the account, site or a client until a person says yes.

What we built

People decide, agents draft, build, test, watch and report, and scripts move the data. Free scripts in the ads account export search terms, pacing and broken pages overnight into one sheet.

Every Monday a search agent reads the week and proposes negatives, pauses and bid moves. Four more agents rebuild the site, draft pages, watch uptime and triage enquiries.

How we made it hold

Every agent has a written limit: the build agent never merges, the director approves each page and the studio's coordinator sends every reply.

For ninety days every change the search agent proposes waits for a person and those decisions become its rules.

Backstage The decisions behind the build, and the hardest problem

Decisions

  1. Scripts move the data into one sheet.

    Nobody copies a report by hand and the agents and the people read the same numbers.

  2. Agents propose, a person approves.

    Every change waits for a yes for the first ninety days and those decisions become the rules.

  3. A written limit for every agent.

    The build agent never merges or touches the domain and clients only ever hear from people.

Ruled out

  • Paid search before the site and tracking were fixed.

    Analytics never loaded and no enquiry was recorded, so the spend would have taught nothing.

The hardest problem

Handing the standard work to agents without letting anything reach the public unchecked. Nothing an agent drafts or proposes reaches a client, the site or the ads account without a named person's approval.

The difference

  • Every account is in the studio's name and no vendor keeps a login.
  • Nothing an agent produces reaches a client or the public without a named person's approval.
  • Free scripts and low-cost tools leave nearly all of the budget for media.

Who did what

m-anant
discovery, vendor advice, strategy, operating model
Weave
the rebuilt site and its design system
Forge
agents, scripts and the site build
Next case A booking that follows the data Booking portal for four beauty and spa brands, Singapore.