Field notes: the week of October 9, 2026

We spent most of this week staring at a dashboard that looked great and meant nothing. A client's lead volume was up 40 percent month over month, and the sales team was quietly furious. The leads were worse. Not fewer qualified people, just more unqualified ones diluting the pool. The number went up and the business got harder to run. That gap between a metric and a result is the thing we keep running into, so that is where these notes start.

The vanity of leading indicators

A leading indicator is supposed to predict a result. Lead volume predicts revenue, the theory goes, so you optimize for lead volume. The problem is that leading indicators are easy to move in ways that break their connection to the thing they were supposed to predict.

Drop the form fields from seven to three and submissions jump. Add a discount offer to the top of the funnel and clicks jump. Loosen the targeting and impressions jump. Every one of those moves improves a number that a dashboard rewards, and every one can make the downstream result worse by sending more unserious people into the pipeline.

This week we rebuilt the client's reporting around a single question: of the leads we generated, how many closed, and at what size? That meant waiting. You cannot judge October's leads until November's deals close, which is uncomfortable when everyone wants a weekly number to look at. We added two intermediate metrics to fill the gap:

• Lead-to-meeting rate, pulled from the CRM, segmented by campaign
• Meeting-to-opportunity rate, which tells us whether the people showing up actually have a problem we can solve

Those two together caught the dilution within days. The campaign driving the volume spike had a lead-to-meeting rate of roughly half the account average. More leads, fewer meetings per lead, no change in closed revenue. The dashboard was celebrating work that produced nothing.

A metric you can move without moving the business is not a measurement, it is a distraction with a chart attached.

Attribution arguments we are tired of having

Three times this week we got pulled into a version of the same debate: which channel gets credit for a deal. Paid search says it was the branded search click. Content says the prospect read four blog posts first. Paid social says the first touch was an ad six weeks earlier. Everyone has a screenshot.

These arguments are mostly a waste of time, and we say that as the people who build the tracking. Here is the honest position we have landed on.

Attribution models are not truth. They are accounting conventions. Last-touch overcredits the bottom of the funnel. First-touch overcredits the top. Multi-touch spreads credit across everything and makes no channel look responsible for anything. Each model is a different opinion about which lie you prefer.

What we actually do is simpler and less satisfying:

• Pick one attribution model per account and do not change it mid-quarter, so trends stay comparable
• Use holdout tests and geo splits when a channel's contribution is genuinely in question, because an experiment beats a model
• Keep a separate "self-reported" field on the lead form asking how people heard about us, and treat the mismatch between that and the tracked source as useful information rather than an error to reconcile

The self-reported field is the one most teams skip. It is noisy, people misremember, and it does not sum to 100 percent of anything. It is also the only signal that captures the podcast mention and the word-of-mouth referral that your pixels will never see. When tracked attribution says paid social drove a deal and the buyer says a friend recommended you, the friend probably mattered more.

A note on AI-written pages

We graded a few dozen sites this week, and a pattern is getting hard to ignore. More landing pages read like they were generated in one pass and shipped without a human deciding what the page was for. The giveaways:

• Three different value propositions on one page, none of them committed to
• Headlines that describe a category instead of making a claim ("Marketing solutions for growing businesses")
• FAQ sections answering questions no real customer asks, clearly written to hit keywords
• Copy that is grammatically perfect and says nothing a competitor could not also say

The tooling is not the problem. We use language models every day, and our own MCP servers let Claude work directly against client data. The problem is skipping the judgment step. A model will happily produce a plausible page for any input, including an input that has not decided who it is talking to. When the page has no point of view, the model fills the vacuum with the average of everything it has read, and the average of everything is forgettable.

The fix is not to write by hand. The fix is to do the thinking a model cannot do for you: pick the one customer, pick the one claim, name the specific objection you are overcoming. Then let the tool draft against those constraints. We saw the difference clearly this week on an account where we fed the model a sharp brief versus a vague one. Same tool, same hour, completely different page.

Pricing when you are paid on outcomes

Because our compensation is tied to results, we spend more time than most agencies arguing about what counts as a result before the work starts. This week that meant turning down a scope.

A prospect wanted us to improve "brand awareness" and tie our pay to it. We said no, not because awareness does not matter, but because we could not agree on a measurement that we would both trust in six months. Awareness proxies (branded search volume, direct traffic, survey recall) are all gameable or slow or expensive to measure cleanly. If we cannot define the number honestly up front, we should not get paid on it. Taking the deal would have meant a quarter of fights about whether the needle moved.

The version we would say yes to: tie pay to pipeline or revenue, and run awareness work as an input we believe will improve those, measured by holdout where possible. Same activity, honest scorecard. The discipline of outcome-based pay is that it forces this conversation early, when it is cheap, instead of at renewal, when it is a relationship-ending argument.

That is the week. If you want a blunt read on whether your pages are deciding what they are for or letting a model decide for them, run the free grader. If you would rather have a partner whose pay depends on the result and not the activity, apply for partnership.