At Local Media Fest this summer, Phil Pracht, director of core products at BLOX Digital, asked a room of 100 to 150 publishers who knew what MCP was. About five people raised their hands.
In a recent Local Media Association webinar, Pracht came back to the topic. This time he skipped most of the theory and showed what it looks like when a publisher uses it.
MCP stands for Model Context Protocol. It’s a standard way for an AI tool like Claude or ChatGPT to connect to another system and ask it questions directly. Pracht described it as asking your CMS’s “brain” what’s going on. Your CMS already knows which stories ran, who read them, who subscribed and who left. Usually getting those answers means clicking through several dashboards, building a report or waiting on the one person in the newsroom who knows where everything is.
“Your dashboards probably only tell you one layer of those at a time,” Pracht said. The value of MCP is in questions that cross layers, like which reporter brings in the most subscribers, or whether a big story paid off.
A brief, not a dashboard
Pracht’s main demo started with one prompt: pull last month’s performance and build a one-page publisher brief. It asked for four things: how the month went, one thing the publisher wouldn’t have guessed, what drove subscription starts and stops, and three things to do on Monday.
The prompt doesn’t name a report, a date filter or an export. It says what the publisher wants to know and who will read the answer. Behind the scenes, the AI picked which CMS tools to call and in what order: page views compared with the prior month, stories by section and byline, what each new subscriber read just before paying, and the reasons logged for cancellations.
The result for an anonymized mid-size daily, called the “Prairie Ledger” in the handout, opened with a line no dashboard would write: the paper signed up 172 more subscribers than in August and brought in less money. Most of the gain came from free classroom accounts and print subscribers turning on digital access they already had. Paying subscribers actually went down.
The surprise was in the bylines. The paper’s most-read writers sold almost nothing, while its court reporter outsold every other byline combined, and it happened two months in a row. One murder trial, covered in three stories and a photo gallery, brought in 22 signups.
The Monday recommendations were specific. Cover the next trials starting at jury selection instead of waiting for the verdict. Put an annual offer on high-traffic stories that don’t ask for anything yet. Offer digital to every print subscriber who cancels. The brief named reporters and stories (the names in the handout are changed), which Pracht said is the point. “It’s important to you that crime content performs well” doesn’t mean much to anyone. “This trial, this reporter, these 22 subscriptions” does.
The rules do a lot of the work
Pracht said the rules attached to his prompt were the most important part. Every number has to come with a comparison or it gets left out. No analytics jargon, because the reader runs a newspaper, not an analytics team. Use real names.
His biggest lesson: “AI is going to happily, it wants to, it’s begging to, create a dashboard.” If you want something else, you have to say so.
Follow-up questions
Because the AI keeps the context of the conversation, follow-ups don’t require a new report. Pracht asked who was leaving and whether they had been paying. Fewer people canceled overall, but more of the ones who left were paying customers: 132, up from 99. Most were 99-cent trial buyers who turned off auto-renew about three weeks in, just before their first full charge.
The suggested fix was to send those buyers a short note in their first week pointing them to the court coverage. Pracht never mentioned courts in that follow-up. The AI remembered it from the brief.
From there he asked it to draft an “in case you missed it” roundup of the stories that sold best. In BLOX, anything the AI creates is saved as a draft. An editor still reads it, fixes the tone and decides whether to publish.
Same prompt, different papers
Pracht ran the identical prompt on other anonymized papers to show it wasn’t returning the same canned answer.
At Prairie Valley, a smaller site, an athlete-of-the-week poll drove a big spike in traffic, but only five of the month’s 32 new subscribers came from the winner write-ups. The Monday suggestion was to link last week’s winner at the top of every ballot. It’s about five minutes of work a week.
At the Riverbend Courier, Friday football sold as many subscriptions as crime news with about one-sixth of the readers.
Some patterns showed up across all three papers, which have different owners. Subscriber counts and revenue moved in opposite directions, because free and bundled accounts made the totals look better than the business did. The most-read stories weren’t the ones that sold. Long-running coverage like trials and high school football was. All three were also losing paying readers through problems that are easy to miss, like print readers losing digital access or failed cards that were never retried. Pracht didn’t ask about any of those specifically. He told it to keep digging.
Your archive is searchable too
MCP also works on content. Pracht asked what the paper had published about property taxes over five years, with a timeline and a list of questions it raised but never followed up on. The answer surfaced open threads like whether landlords ever signed up for a program and who didn’t get a break. It also noted that the paper had done almost no reporting on what any of it meant for a local homeowner’s bill.
He also suggested using it for reader questions like “why is the Fifth Street bridge still closed?” Ask the CMS what you’ve reported, and you get a short answer with links to send back.
Guardrails
MCP doesn’t give the AI your admin password. It gets a specific list of permissions, and whoever manages your BLOX instance can turn them on or off. In the demo, Pracht asked the AI to save a roundup draft to BLOX NXT, and it said it couldn’t because article creation was turned off for that connection.
On training: business and enterprise accounts with Claude, ChatGPT and similar tools generally have training on your data turned off by default. On a personal account, you may need to turn it off yourself. BLOX has published a guide to how each provider handles this.
Pracht also said to check the numbers before acting on them. BLOX NXT has per-article stats you can use to confirm what the AI tells you. “A wrong number is easier to catch in a table than a polished sentence with someone’s name attached,” he said.
Making it repeatable
A common complaint about MCP is that it burns through a lot of usage while it searches. Pracht’s fix is to turn a prompt that works into a reusable “skill,” so the AI already knows where to look. His first run of each brief took about five minutes. Later runs took a minute or less.
What this means for publishers
MCP changes the job from pulling reports to asking better questions of the data and content publishers already have. What drove subscriptions? What did we miss? What should we do Monday? The answers still need a human to check the numbers, apply judgment, and decide what happens next. But for lean local media teams, getting from question to useful answer in a minute instead of an afternoon is a meaningful change.
If you’d like to watch the full webinar, you can do so at this link.
Passcode: YGZ80Or^