Answering the night shift

This is the part of the engagement most agencies would not have initiated. Search was not broken in any way that shows up in a status meeting. It was live, it was spending, it was reporting conversions, and the client had not complained. The honest read of the account was that it had stopped being the right place for this budget, and nobody was going to tell us that except the numbers.

Across the first month of this window, search absorbed 67% of total ad spend and returned 17% of the phone calls. A phone-call lead cost $225.62 there. The same lead cost $23.07 on paid social. By the following month search had deteriorated to $561.48 per call while taking 28% of the budget for 2% of the calls. Continuing would have been defensible — the campaigns were optimised, the keywords were tight, the reports looked busy. It would also have been wrong.

So we recommended reallocating almost the entire budget to paid social call-only campaigns, and we brought the recommendation with the replacement already built rather than as a problem for the client to weigh up. Search was wound down and eventually paused. It is a strange thing to argue yourself out of a channel you are being paid to run, and it was plainly the right call.

Five studios, five separate campaigns, each pointed at its own phone line. A franchise operator earns per location, not per group average, so each studio is funded and judged on its own cost per call. Pooled into one campaign, the algorithm quietly funnels budget to whichever location converts most cheaply and starves the rest — which reads as a healthy blended number in the report while a real studio sits with a silent phone.

Delivery was kept continuous rather than scheduled to opening hours. Restricting delivery windows repeatedly pushes campaigns back into learning, so tidier call timing is bought with worse efficiency across the whole account. We took the efficiency and solved the timing problem downstream, on the business side — which is the thread that runs through the rest of this case study.

From month two we layered in customer-level optimisation, feeding the platform the operator’s own customer profiles so the algorithm targets people who resemble the studios’ best existing members rather than a demographic guess. Creative is cycled deliberately and often: the delivery algorithm matches assets to users, and stale assets raise cost per call long before anything else in the account does.

The clearest way to read this engagement is not the paid-social number in isolation — it is what happened to the operator’s total marketing bill and total lead volume at the same time. Across the three months, total ad spend across every channel fell by 20% while total phone-call leads rose by 37%.

Blended cost per phone-call lead went from $58.00 to $33.99, a 41% reduction, and that figure includes the wasted search spend still sitting in the first month. The business is now buying substantially more of the only outcome it cares about, for materially less money, from a single channel it was not using at the start of the year.

Cost per call did rise within paid social as we scaled — $23.07 to $32.00 to $33.99. That is the expected cost of moving past the cheapest, highest-intent slice of an audience, and it is the correct trade at this volume: month three bought 199 calls where month one bought 120. We report the rise rather than quoting the opening month and calling it the result.

Volume alone is the easiest number to flatter. A campaign can generate a great many dials that end two seconds after they connect, and the report will still read well. So we measure how long each call stays alive: a dial is intent, a call still connected at twenty seconds is a real person, and a call past sixty seconds is a conversation long enough to have covered a service, a price and a date.

Every one of those measures improved across the three months, in both absolute and percentage terms. Sixty-second conversations nearly tripled, from 14 to 41, and rose from 11.7% of calls to 20.6%. Callback requests — people who asked to be rung back rather than waiting — went from 5 a month to 27.

That combination is the one worth having. Cost per call rose modestly as we scaled, and the proportion of those calls turning into real conversations rose faster. The customer-level optimisation introduced in month two is the most likely driver, and we state that as the reading of the data rather than as a proven cause.

Running delivery continuously bought efficiency, and it also meant a large share of calls landed outside opening hours. Across the ninety days, 206 of the 496 calls — 41.5% — arrived when no studio was staffed. In the first two months that was the majority of all call volume: 53.3% and 53.1% respectively.

At the blended rate those after-hours calls represent roughly $2,100 of media a month ringing into an empty room. The platform stores no phone numbers for call-only ads, so not one of them could be retrieved and called back the next morning. The demand was already bought and paid for; it was being lost between the ad and the desk.

By month three the after-hours share had dropped sharply to 24.1%, as the customer-level optimisation pulled delivery toward the hours the studios are open. That is a genuine improvement, and it is not a fix. Forty-eight calls a month still arrive after closing, and a smaller leak with a known cause is an argument for plugging it properly rather than for leaving it alone.

The obvious commercial response to a leak like this is to sell more media. It is also the wrong one: raising spend to generate more calls into an unanswered line multiplies the loss rather than fixing it. The cheaper, larger win was downstream of the ad account entirely — which put it outside any media scope of work, and therefore squarely inside our job.

We designed, built and shipped an AI voice receptionist that integrates directly with the operator’s existing VoIP phone system. When a studio is closed, it answers the call live. It fields questions using the studio’s actual services and pricing, establishes whether the caller has had extensions before, captures their name and preferred appointment times, and texts them the studio’s direct booking link while they are still interested. Every morning the studio manager receives an email breakdown of the night’s calls — name, number, service requested, preferred times — so the team opens the day with a list instead of a void.

It was deliberately built to be adopted rather than admired. Setup takes under ninety minutes in total, requires no new contract with the phone provider, and costs roughly $60 to $70 a month in usage across all five studios. That is about three percent of the after-hours media it protects each month. An elegant system the operator will not switch on is worth nothing; this one went live.