Half of Slack Messages Now Come From Agents. Your Messaging Needs an Eval Suite.

Half of Slack messages come from apps and agents. The first reader of your messaging is a model now. Give it a context file and a message eval suite.

Falk Gottlob · · 8 min read

Brick-red Falkster cover: a single index card pinned at the center, with taut threads running out to a ring of loudspeakers, radios, and robot heads that lean in to listen, and one blue thread snapped and curling loose.
4/4
SeriesThe Product Marketing Transition

Salesforce published Slack's numbers on November 3. Half of Slack messages now come from apps and agents. 127 million MCP tool calls a week, doubling every quarter, by the company's own count.

Same day, Louise Linehan at Ahrefs published an analysis of 260 million AI prompts across 10,000 brands, with data work by Xibeijia Guan. What correlated most with how often an assistant mentions a brand? Mentions in Reddit threads, at 0.75. General webpage mentions came in at 0.70. Correlation, not cause, and they say so plainly.

Read those two side by side and one thing follows. The first reader of your messaging isn't a person anymore. It's a model. Sometimes it's your agent, sometimes it's the buyer's, and only after that a human reads whatever the model said.

So messaging needs what every other model input has. A context file. And an eval suite.

This is part 4 of The Product Marketing Transition. Part 2 put one positioning file in the repo with two owners, the PM and the PMM. Part 3 sorted every change by what it does to the claim. This part is about how you know the claim survived the trip through a model.

The short version

  • Salesforce reported on November 3, 2026 that half of Slack messages come from apps and agents, and Ahrefs found the same day, across 260 million prompts, that off-site mentions correlate most with how often AI assistants name a brand.
  • Your messaging now has three readers: your own agents, where you control the context; the buyer's assistant, where you don't, and what others say carries the weight; and people, who mostly read what the first two produced. A messaging doc nobody tests is a guess.
  • So write the claim as one versioned context file every customer-facing agent loads, build a message eval suite from the 25 or so questions buyers ask before a shortlist, run it on every change and weekly against the public assistants, and track message drift.
  • The agents that read the file are product, their prompts and evals live in the repo, and that's why this work sits in the product pipeline under the CPO.

Three readers, and you only write for one

Your own agents are the reader you control. Sales agents, support agents, the onboarding agent inside the product. HubSpot rebuilt its platform in September around what it calls Growth Context, a running picture of the company, its teams, and its customers that grounds its agents, with a Context Home that scores how complete that picture is (MarTech's Constantine von Hoffman covered it). That's a CRM vendor telling every one of its customers that the positioning has to live somewhere an agent can load it.

Melissa Kwan showed what that looks like when one person does it on purpose. In Kyle Poyar's Growth Unhinged on October 21, she described a positioning brain at eWebinar: customer quotes, objections, alternatives, a language guide, and a set of instructions the model maintains. Every finished page goes back in. Her team now writes emails and posts from it without a briefing. Her line is the one I'd frame: "positioning is retraining, and retraining only works if every touchpoint says the same thing."

Agents are touchpoints now. Most of them, if the Slack number is any guide.

The buyer's assistant is the reader you don't control. Ahrefs' numbers say its picture of you is assembled mostly from what other people wrote. 10x more webpage mentions lined up with about 3.7x more AI mentions. 10x more Reddit mentions, about 5.5x. And the median brand had eight Reddit threads mentioning it, with 29% having none. So for this reader, the context file is your claim plus the proof that lets someone else repeat it in their own words.

People are the third reader. They still decide. But more and more, they read the summary a model gave them.

A messaging doc nobody tests is a guess

April Dunford put it well at Mind the Product in July: "The fundamentals hold. The execution doesn't." Who it's for, what they'd do instead, why you win. None of that moved. What moved is that the claim now passes through a model before it reaches the buyer, and models paraphrase, compress, and sometimes invent.

So the execution is a test. I'd call it a message eval suite.

Start with the questions buyers ask before a shortlist. Twenty to thirty of them. Which tools do this for a company like mine? How is X different from Y? What does it cost at our size? Is it safe for regulated data?

Then write a rubric for each answer from the claim itself. Does it name the alternative you beat? The differentiator? The proof? Does it avoid the one line you never want said, the stale pricing, the retired product name, the capability you sunset?

Run it in two places. Inside, against your own agents, on every change to the context file, the way a prompt change runs against product evals. Outside, weekly, against ChatGPT, Claude, Gemini, and Perplexity. Two numbers come out. A pass rate for your agents. And message drift, the share of answers anywhere that contradict or miss the current claim.

The inside number you can fix in an afternoon. The outside number tells you where the proof is missing.

The file sits next to the agents, so it sits under the CPO

Here's this seat's piece of the one-function case. The agents that read the messaging file are product. The support agent, the sales agent on the website, the onboarding agent in the app. Their system prompts live in the repo. Their evals run in the pipeline. The messaging file they load belongs next to those prompts, and its evals belong in the same run.

A messaging doc owned by an org outside the product pipeline gets copied into the agents' prompts once and then drifts. Nobody notices until a buyer forwards an answer with last year's pricing in it.

The Design Transition reached the same place from the other side. Its part on the design system as context argued that design owns the system agents generate interfaces from, and someone on design signs the weekly diff. Product marketing owns the claim those agents speak. Two context files, one pipeline, one owner of the outcome. That's The Product OS, and it's why the CPO Transition keeps landing on one function under the CPO.

The decision for a CPO this quarter: name the messaging context file, put it in the repo, make every customer-facing agent load it, and make the PMM a required reviewer on changes with a message eval that has to pass. Then put drift on the same dashboard as adoption.

What I expect

These are my predictions, not facts.

By the end of 2027, every customer-facing agent at AI-native software companies loads one versioned messaging file, and changes to it run against a message eval suite before they merge, the same way prompt changes do.

By 2028, message drift is a number the head of product marketing reports monthly, next to story lag from part 3.

Within two years, by late 2028, PMM teams at AI-native companies put more effort into earning third-party proof, the reviews, community threads, customer write-ups, and analyst notes the assistants read, than into owned web copy. Ahrefs' correlations point that way, and the teams that test the outside number will see it first.

One thing to try this week

Write down ten questions your buyers ask before they build a shortlist. Ask your own support or sales agent all ten. Then ask ChatGPT, Claude, Gemini, and Perplexity the same ten.

Grade every answer on two things: did it name the alternative you beat, and did it give the proof? Count the passes. If your own agent scores lower than the public assistants, the context file is the problem, and you can fix it this week.

Sources: Salesforce, "Half of Slack Messages Are No Longer From People. Here's Slack by the Numbers.," Salesforce Newsroom, November 3, 2026, company figures. Louise Linehan, with data analysis by Xibeijia Guan, "Which Types of Mentions Correlate the Strongest With AI Visibility? (260M AI Prompts Analyzed)," Ahrefs, November 3, 2026. Melissa Kwan, "How to build an AI brain for your positioning," Growth Unhinged (Kyle Poyar), October 21, 2026. Constantine von Hoffman, "HubSpot rebuilds its platform around AI agents," MarTech, September 16, 2026, context only. Louron Pratt, "Everything has changed except for the fundamentals of positioning: April Dunford at #mtpcon London 2026," Mind the Product, July 13, 2026, context only.

260 million

prompts in the Ahrefs study. Off-site mentions correlate most with how often AI assistants name a brand.

Part of the running argument on Product Leadership: the product org chart was calibrated to a build cost that collapsed, and when the first reader of the messaging is a model, the claim becomes a context file with evals, maintained in the product pipeline under the CPO.

Related answer: Half of Slack messages come from agents. Who writes the messaging they read?

Share this post

Frequently asked

What does it mean that messaging is model context?+

The first reader of a company's messaging is now usually a model: its own sales, support, and onboarding agents, and the AI assistants buyers ask. Salesforce reported on November 3, 2026 that half of Slack messages come from apps and agents. Messaging that agents load has to be written, versioned, and tested like any other model input, in one context file with an eval suite.

What is a message eval suite?+

A fixed set of the questions buyers ask before a shortlist, usually 20 to 30, each with a rubric derived from the claim: does the answer name the alternative you beat, the differentiator, and the proof, and does it avoid the line you never want said. It runs against your own agents on every change to the messaging file and weekly against the public assistants.

Can you control what ChatGPT or Gemini says about your product?+

Not directly. Ahrefs' November 2026 analysis of 260 million prompts across 10,000 brands found off-site mentions correlate strongly with AI visibility, with Reddit thread mentions at 0.75 and webpage mentions at 0.70. Those are correlations, not proof of cause. For the buyer's assistant, the context is mostly what others say, so the PMM's job is a claim others can repeat with proof, and a weekly test of what the assistants actually say.

Who should own the messaging context file?+

The PMM owns the claim and the evals, and the PM co-owns the file, as part 2 of the series argued. The file sits in the repo next to the agents' prompts and evals, so changes run in the same pipeline. That's why the seat belongs in the product org under the CPO.

What is message drift?+

The share of answers, from your own agents or from public assistants, that contradict or miss the current claim. Falk's prediction is that by 2028 the head of product marketing reports it monthly next to story lag, the days until assistants describe a change correctly.

What are Falk's predictions on messaging as model context?+

Three, labeled as predictions. By the end of 2027, every customer-facing agent at AI-native software companies loads one versioned messaging file, and changes run against a message eval suite before they merge. By 2028, message drift is reported monthly next to story lag. And within two years, PMM teams at AI-native companies put more effort into earning third-party proof than into owned web copy.

product marketingPMMgo-to-marketCPOProduct Marketing TransitionProduct OSmessagingpositioning
Helpful?

Keep reading

The short answer

The Claim FilePositioning is one versioned file in the repo that every agent, page, and assistant profile loads. Sort every change by the claim, and track story lag.Read the chapter

Part of

Product Leadership

About the author

Falk Gottlob

Falk Gottlob

Product Executive · Founder, Falkster.AI

Thirty years shipping product, from Microsoft Research and Adobe to Salesforce, where he grew Quip into what became Slack Canvas. Four startup exits, including a $6.5B healthcare merger and a company Microsoft bought. Four-time Chief Product Officer. Now founder of Falkster.AI, building Heidi, the company brain for AI-native businesses. This notebook is written from inside the build, not above it.

Sign in with LinkedIn to leave a comment.

Sign in with LinkedIn