You get what you optimize for
Ad platforms are extremely good at finding more of whatever you tell them a conversion is. If the conversion event is a form submission, they will find you people who fill in forms. That is a real population, and it is not the same population as people who buy.
This is why cost per lead can fall for three months while the pipeline stays flat. Nothing is broken; the system is working exactly as instructed, on the wrong instruction.
The definition has to come from sales
A good lead is not a marketing concept. It is a description of the accounts your sales team can actually convert, and only they can write it.
The conversation is short and specific. Which inquiries from last quarter did you want more of? What did they have in common — size, sector, urgency, budget, whether they had already tried to solve it themselves? Which ones did you know were wrong within thirty seconds, and what told you?
Write the answer down as a rule, not a feeling. It should be specific enough that two people looking at the same inquiry agree on whether it qualifies.
Then move the optimization event
Once the definition exists, the qualified stage has to be sent back to the platforms as the conversion. That usually means a CRM integration — the lead is created, the sales team marks it qualified or not, and that outcome flows back to Meta or Google as the event worth optimizing.
Expect volume to fall and cost per lead to rise. Both are supposed to happen. You are now paying for a rarer, more valuable thing, and the number to watch is cost per qualified conversation, not cost per lead.
The uncomfortable part of this is that it makes the marketing numbers look worse before they look better. Agree that with whoever reads the report, before you change anything.
Qualification belongs before the calendar
The last piece is filtering. Every question you ask on the form is friction, and every unqualified call is thirty minutes of a salesperson's day, so the trade is real and worth thinking about rather than defaulting.
As a rule: filter on the one or two attributes that reliably predict fit, and leave everything else to the conversation. Two well-chosen questions remove most wrong-fit inquiries without materially reducing the right ones.