Two claims, both true. Paid search remains the best instrument ever built for harvesting declared intent. And the measurement stack wrapped around it, last-click ROAS, in-platform attribution, cost-per-MQL arithmetic, now misprices it badly enough that teams are making real budget decisions on fiction.
When the numbers wobble, the instinct is to blame the channel. Wrong suspect. Blame the ruler.
What paid search still does better than anything
Someone types "corporate card and expense software" into a search box. That is not inferred intent or modeled intent. That is a person telling you what they want, in their own words, at the moment they want it, and three jobs built on that moment remain genuinely excellent.
Category capture: being present when a declared need matches what you sell. Competitor defense: your name adjacent to theirs at the exact moment of comparison, which is cheap insurance against a shortlist of one. And offer testing at speed: paid search is the fastest message lab you own. A new offer, a new angle, a new price framing can meet real buyers on Monday and return a verdict by Friday. Nothing else in the stack tests copy against declared intent in days rather than quarters.
None of that has degraded. What has degraded is your ability to see it clearly.
Where the numbers lie
Start with brand terms. Brand campaigns post heroic ROAS because they intercept people who had already decided, and last-click attribution hands them full credit for the interception. Last click does not measure influence. It measures proximity to the checkout. A meaningful slice of what your dashboard calls paid-search revenue is demand created somewhere else entirely and merely collected at the toll booth.
Now the newer distortion: the invisible assist. A growing share of B2B evaluation happens inside AI answers before any click fires. A buyer asks an engine for a shortlist, reads what it says about you, asks the follow-up questions, and only then goes to a search box. If the answer layer recommends you, your campaigns harvest pre-warmed demand and look brilliant. If it recommends competitors, your non-brand clicks arrive cold, conversion rates sag, CAC climbs, and the platform's recommendation engine will suggest you raise bids. The platform grades its own homework, and it is a generous grader. In neither case did your paid execution change. The upstream answer changed, in a surface your dashboards do not watch.
Cost-per-MQL math has a related defect: it prices the form fill, not the buyer. When evaluation moves into the answer layer, the person who eventually clicks and converts is further along, better informed, and rarer. Cost per lead drifts up while cost per real opportunity holds or falls, and a team reading the lead line alone will conclude the channel is decaying when what changed is where the thinking happens.
Then there is incrementality, the question almost nobody asks of their own spend: how much of this would have happened anyway? Brand-term pause tests and geo holdouts are cheap, well understood, and almost never run, because the number that comes back might be smaller than the number in the deck. Draw your own conclusion about whose interests the current numbers serve.
The fix: run paid like an experiment, not a faith
First, pre-register. Before a dollar moves, write down the hypothesis, the success criteria, the observation window, and the kill line: the point at which spend stops regardless of who is fond of the campaign. We ran demand budget this way for a global travel and expense platform, success criteria and kill lines registered before launch, and the unglamorous outcome was exactly the point: budget stopped defending itself and moved to what worked. Pre-registration does not make anyone smarter. It makes it impossible to grade on a curve after the fact.
Second, test incrementality with real holdouts. Geo splits, matched markets, scheduled brand-pause windows. The number that comes back may indeed be smaller than the deck number, and that is the reason to run the test rather than the reason to avoid it. A smaller true number beats a larger fictional one everywhere except the deck.
Third, stop reading paid in a vacuum. Track what engines say about you for your money queries with the same discipline you apply to impression share; that is the core of AI search optimization, and it belongs in the same weekly review as your paid numbers. When answer-layer presence shifts, expect paid metrics to move for reasons that have nothing to do with bids or budgets, and stop interrogating campaign settings for a confession they do not contain. An attribution model with no line for the answer layer is not conservative, it is blind in one eye, and a measurement stack built for this decade has to see out of both.
What this looks like on a Monday
A weekly read where paid performance, answer-layer visibility, and holdout results sit on the same page. Campaigns with expiry dates and kill lines instead of tenure. Brand spend justified by pause tests instead of by ROAS screenshots. A budget line that can name, for every dollar, the experiment it belongs to and the evidence that would take it away. Fewer arguments about attribution models, more arguments about experiments worth running next, which is the better argument to be having.
It is less flattering than the old dashboard. It is also true, which the old dashboard was not.
Keep buying the clicks; just stop letting the click be the judge of what the clicks are worth.