How the desk works

Every pitch, reply, statement, review response, edit request and report is a proposal until a named human approves it. The approver sees the recipient, the full text, every source, the risk flags the fact-checker raised and the diff since the last version. One click approves. Any edit after that click makes a new version that needs a new click. Approvals expire after 72 hours, because a pitch that was right on Tuesday may not be right on Friday.

Two named approvers are required for anything about a crisis, litigation, a regulator, money, health, politics or a private individual. A kill switch pauses all sending for one client or for everyone, and the agents check it before every outbound call.

Who approves depends on the plan. On Debut, you do, with our editor spot-checking behind you. On Buzz and above, your lead does, and she does not delegate it to the agents. On Situation Room, nothing public moves without two people.

A claim is a number, a date, an award, a customer name or a quote. In the editor, every claim renders as a chip that links to its receipt: a document you gave us, with a version hash, or a page we fetched, with the time we fetched it. A claim without a receipt renders as Unverified and the approve button stays grey until it is fixed or cut. Not a warning. A block.

Quotes are only ever attributed from a quote bank the speaker has signed off. Agents may propose a quote for you to approve; they cannot publish one as said. Coverage reports list only pages we actually fetched that actually mention you. If something we published turns out to be wrong, a correction is drafted within 24 hours and queued for approval, because that is what the PRSA code asks for1 and what a decent person does.

A third of journalists bin a pitch for missing or unverifiable sources.2 The ledger is a feature for them as much as for you.

Every action an agent can take has a setting. Four levels, per action, per client.

You can tighten any setting. You cannot loosen one past our ceiling, and neither can we without two admins agreeing. Every change is logged with who made it and when. Clients can be given read-only access to the desk to watch it work.

The defaults and the ceilings, read from the code →

Your chapter arrives on Friday and is numbered, so a year from now you can read the arc of your story from chapter 1. It reports the way the industry's own measurement standard says to:3 what we did, what ran, what it changed and what it led to.

Things the chapter never contains: advertising-value equivalents, "impressions" from a media database, and "pitches sent" presented as a result. The fact-checker reconciles every number to the database before your lead reads it, and your lead reads it before you do.

A crisis starts one of two ways: the monitor sees volume and sentiment move together and pages a human, or you press "Declare incident". Within five minutes the crisis lead has a brief on the desk: what is confirmed, what is unknown, when each fact was last checked, how far it has spread, who is driving it and what you are exposed to. If the criteria are met it sets a legal hold, which pauses outreach and review replies for your brand until a person lifts it.

The holding statement is drafted from the templates you approved on a calm day: aware, concerned, acting, next update, press contact. The fact-checker verifies every sentence against a receipt. Two named people approve it. Then, and only then, it goes to your newsroom page and to any journalist who asks. Inbound press questions are answered with the approved statement and nothing else; anything beyond it goes to a human.

Monitoring runs every 30 to 60 minutes with a delta brief until the severity rule says it is over. Within 72 hours you get a post-mortem: how fast we were, how consistent, what ran, what we would change in the playbook.

Headline: a human acknowledges within 30 minutes in business hours. Situation Room: a human acknowledges within 30 minutes, any hour, and a holding statement is on the desk within two.

When someone asks ChatGPT, Claude, Gemini or Perplexity about your category, are you in the answer, where, in what tone and citing which page? Every week (every day on Headline and above) the auditor runs your prompt panel: fifty to two hundred questions, branded and unbranded, three samples per engine, pinned to your region, raw answers stored so we can show our working.

You get a visibility trend, the list of outlets and creators the engines actually cite (which overlaps with the journalists most agencies pitch about two per cent of the time),4 and an inaccuracy log: what each engine got wrong and the exact page it learned it from.

The fix plan is ordinary, honest publicity: correct your own pages, put a proper newsroom on your site, pitch the outlets the engines cite, mark up the structured data. We do not try to trick an engine, build fake pages or inject prompts. It does not work for long, and when it stops working it takes your name with it.

See what they say about you now→