Who should pay for the insurance quote that never turns into a policy?
At EverQuote, insurance providers paid for the referral, and the carrier or agent owned the close.
Every marketplace makes this choice. Charge for the click and the supplier carries the sale that does not happen. Charge a commission and you carry it. Charge for the lead and the risk splits at the handoff.
Charge at the last step you control.
Running product at a fast-scaling SaaS company in Boston, we burned six months searching for a VP.
By the time the new hire onboarded, half the team had changed and the problems we hired for had tripled. The job description we spent six months filling described a company that was gone.
A search feels like progress because candidates keep moving through it. Write the job description for the company you will be when the hire starts, and staff the gap while you wait.
Gainsight said this week it is retiring per-seat pricing. From next year customers pay on the ARR and customers they manage.
The bigger move is Atlas, its services business: agents plus human renewal managers, on contracts tied to results, already managing more than $75M in customer ARR.
That is close to my own model, which prices on the decision and runs on agents. Seat pricing kept vendors out of whether the work got done. Results pricing puts them in the room.
A new member on a call of fractional executives asked how many hours we work. One person aims for 30 paid hours and lands at 60. I said 80, and that I love it, then told her to take my advice with a grain of salt.
Last year I helped a founder build her growth plan around a nine-to-three workday, the window she had before school pickup. I still believe that advice. I do not follow it.
Eighty hours is my price. A plan built for the week you actually have beats one built for mine.
Delete the column headers from your product OKR sheet. If what is left reads like a roadmap, it is one.
At BetterLesson the company had one number for the year, $5M to $9M. Eleven teams set six-month KPIs against it. The quarterly review gave each team seven minutes to walk through its number and thirteen for questions.
A key result is a number you have promised to explain in 90 days.
https://www.matthewmamet.com/blog/okrs-in-product-management-fail-when-key-results-are-features/
Casey Winters argues the eval is becoming the core product artifact, because it defines whether you did the thing or not.
My agents' release gate is my eval. It can tell whether a post is allowed. It cannot tell whether it is the fourth version of an idea my feed already ran. That takes a history read before anything is ranked.
Judgment needs memory before it needs rules.
On October 15 HubSpot started letting customers trade unused seats for AI credits mid-contract.
A seat was always a proxy for work getting done in the product. Letting customers cash an empty one in for credits concedes that the work is what is being bought, and that customers already know which seats are empty.
My practice runs on one human login and ten agents. Seat pricing undercharges me.
Seats were how software counted people. Credits are how it will count work.
Everything my ten agents know lives in a folder of plain files. If a better model shipped tomorrow, I could take the folder and go.
Andrew Chen argues agents are tools, not networks, because memory and context are portable. My setup backs him up. His harder question is where network effects will live once they show up. Every marketplace I have worked in kept its moat in reputation, the part hardest to carry out the door.
Whoever holds reputation for agents will hold the network.
For one season I paid for an AI cold-email platform and never once turned on the AI. Outbound won me no clients, and I shut the stack off.
The channel was the problem. Fractional work gets bought the week the pain becomes undeniable, and a stranger's email does not land that week.
Every AI tool I buy now gets one month to replace something. Almost everything from the last fifteen months is gone.
A tool that helps you do the wrong thing faster is the most expensive one you own.
A former TripAdvisor colleague told me he is surrounded by TripAdvisor people again at his new company. Then I said the other thing: if TripAdvisor were started today, it would not need half the people who built it. Probably not him. Probably not me.
He has stopped writing specs. He builds a simple demo and lets it be the document. We used to carry printed specs into product review.
The next generation of product leaders will have built more with their own hands than any of us did.
In August one of my agents staged a blog draft that said I pay for four AI tools. I do not. It read cleanly and passed my release gate. A month later another job picked the same backlog item twice because my spec never said the result had to be written down. The model did nothing wrong either time. That is what an AI product manager manages now: what the system may claim, what done means, and what happens when the output is fluent and wrong. https://www.matthewmamet.com/blog/an-ai-product-manager-now-manages-the-system-that-does-the-work/
Airbnb's host-only fee reaches Europe on October 13. Guests stop seeing a separate service fee; most hosts pay 15.5 percent from their payout.
I read it as Airbnb moving who absorbs the cost. At EnergySage, installers burned through our leads while a wasted lead cost them nothing. When the cost moved to them, they got selective within a week.
Airbnb bought a cleaner checkout with host margin. Watch European nightly rates this winter.
Before I scope a consulting engagement, I ask for one sentence: the decision the buyer needs made by the time I leave.
At Arkadium, the first version was a request for more games. The data said players arrived from search, played one game, and left. The real decision was about retention.
If a buyer can write the sentence, the engagement is short. If they cannot, finding it is the engagement.
Marty Cagan argues that experts lead experts, and that strong product leadership matters more as AI makes building cheaper. Four days later I posted that the CPO role is dead. We are closer than that sounds. The pyramid of managers translating between the business and a big engineering team shrinks with the team. The expert at the top still has to know the work. Where we split is whether that seat is full-time. The expert stays in the room. The pyramid under the expert goes.
One morning in 2009 I told my kids I was heading in to work. My six-year-old asked, "Do you mean the basement, daddy?" I had three clients, about four days a month each, and I would not have called it a practice. What made it hold up was a rule I set without running the math: no more than four clients at a time, so each got someone who still knew their business. I have tested it plenty since. The number is roughly right.
I recently handed off the fractional community I ran in New England for a year, and most of what I gave my successor was a list of my mistakes. I felt like I was failing because we were nowhere near 3,000 members, while the DMs I got were mostly positive. I kept rushing introductions that, for a lot of members, were the reason they came. A community grows at the pace its members set, and the organizer's job is to stop pushing a faster one.
OpenAI put apps inside ChatGPT this week and announced no billing or revenue share. Developers send buyers to their own checkout and keep the money. At TripAdvisor I ran Instant Booking, our push to keep the booking on our own pages. Zero to $200M in 18 months, and we still walked it back. The traffic came. The trust to buy from us never did. Either OpenAI learned that lesson, or zero is an opening price. The take rate is the business model, and it has decided to name it later.
I mapped 170 people who held product roles at TripAdvisor from 2004 to 2024. 45 now lead product somewhere, 21 as chief product officers, across 148 companies, never more than two at one.
Football would call that shape a coaching tree. Junior PMs present for both the hotel metasearch win and the Instant Booking unwind reached senior roles at twice the rate of their peers. Tenure alone showed no effect.
Elena Verna wrote that it is a great time to be freemium, and I cannot decide whether my own experience agrees with her.
At EditMe we ran a self-serve ladder at five, nineteen, thirty-nine and seventy-nine dollars a month, first month free. In 2010 I built the cohort table by hand in SQL.
The free month told me almost nothing. The signal was whether an account had a second person editing by day thirty. A free tier is a measurement instrument first.
My honest split on the people talking about AI. Eighty percent have read about it. Fifteen percent have opened a chat window and uploaded a document. Five percent are building.
I spent years in the first group, sounding current in leadership meetings with no application for any of it.
It changed at EnergySage. Upload an electricity bill, take a satellite image of the roof, and the model works out how many panels bring that bill to zero. Arithmetic nobody wanted to do by hand.