The account belongs to a UK homewares brand with roughly £34m of direct-to-consumer revenue and £4.2m of annual Google Ads spend. The head of paid — who agreed to be reported on but asked that we call her Frida — has been in the role for eighteen months. She inherited an account that had, for four years before her arrival, been run on manual CPC bidding with campaign-level structure that most modern PPC leads would find quaintly old-fashioned: individual ad groups per major keyword theme, individual bids per keyword, tight geographic and dayparting controls, a rigorous weekly manual optimisation cadence.
The account, on inspection when she started, was producing sound but unexciting numbers. Blended reported ROAS of 3.6 on the paid-search line, blended MMM-attributed ROAS of 2.4, contribution margin per pound of paid media of £2.10. Nothing that would embarrass anyone in a board report. Nothing that would win an industry award, either.
Frida's brief, as agreed with her CMO in her first ninety days, was to modernise the account. Move it off manual bidding. Consolidate structure. Take advantage of the platform's automated bidding, which had, by all accounts, become substantially more capable in the four years since the previous PPC lead had made his structural decisions.
The migration
She began the migration in the summer of 2025. The strategy, as she explained it, was cautious. She would migrate one campaign per fortnight from manual CPC to target CPA. She would set the initial target CPA at the campaign's actual trailing-90-day CPA, monitor for two weeks, adjust as needed. She would leave the highest-spending campaign — the one that carried roughly 40% of the account's annual spend and was, on the manual configuration, the account's most reliably productive line — until last.
The first four migrations went as expected. Small volatility in the first week, stabilisation in the second, no material change in the reported CPA or conversion volume. By the middle of September she had migrated eight of the account's fifteen active campaigns. The head of paid, on her monthly review with the CMO, reported cautious progress.
In late September, roughly six weeks before the account's largest seasonal peak, she migrated the biggest campaign. Same methodology: set the initial target CPA at trailing 90-day actual, monitor for two weeks. The dashboard's first-week readout, when the two-week check arrived, looked slightly better than the pre-migration baseline: reported CPA down 4%, reported conversions up 3%. Frida approved the campaign for the peak. She did not, at the time, have any specific reason to look further.
What happened in October
Reported performance through October and early November remained positive. Reported CPA was flat to slightly down against the previous year's peak, reported conversion volume was up mid-single-digits year-on-year. Frida reported strong peak performance to the CMO at the end of November.
The problem showed up in the finance team's month-end reconciliation, which is where these problems always show up when they are going to be found at all. The finance team's model of paid-search contribution margin was substantially below what the paid function had been forecasting. The gap, on the head of finance's calculation, was approximately £310,000 over the peak window. It was too large to be dismissed as measurement noise.
Frida spent the first week of December unpicking what had happened. The story, when it came into focus, was uncomfortable enough that she has been willing to have it published here.
The mechanism
The target-CPA algorithm on the migrated campaign had, when set to a target that matched the campaign's trailing-90-day observed CPA, done exactly what target CPA is designed to do: it had held the reported CPA at that level. What it had also done, quietly, was shift the campaign's traffic mix substantially.
Before the migration, the campaign's manual bid structure had produced traffic that was heavily weighted toward premium product terms — the higher-priced items in the brand's catalogue, where the AOV was £180 and the contribution margin per order was strong. This weighting was, on inspection, an artefact of the previous PPC lead's specific bidding decisions three years earlier, when he had bid higher on the premium keywords because they produced better margin per click.
The target-CPA algorithm did not, and could not, understand this. When it took over, it optimised for the CPA target it had been given — which was a blended CPA across all products. It quickly discovered that clicks on lower-priced product terms were cheaper to acquire and produced the same conversion event (a purchase), and it shifted spend towards those terms. Reported CPA held. Reported conversion count held. Average order value on the campaign dropped from £147 to £106 over the peak window. Contribution margin dropped further, because the lower-priced products in the catalogue carried a lower gross margin percentage.
None of this appeared on the campaign dashboard. It appeared on the finance team's month-end reconciliation, six weeks after it had started.
"The automated bidder was optimising for the metric I had given it. The metric was the wrong metric. The manual structure I replaced had, by accident, been optimising for a different and better metric. My migration threw away the accident."
What Frida did
The immediate move, in the first week of December, was to convert the campaign back to manual CPC with the original bid structure. Reported performance on the campaign dropped by the margin the target-CPA algorithm had been artificially adding — roughly 8-10% reported CPA increase, roughly 4-6% reported conversion count decrease. The traffic mix returned, over two to three weeks, to its pre-migration composition. Contribution margin, on the finance team's model, returned to its pre-migration level within a further two weeks.
The medium-term move, which she has been executing through the first half of 2026, has been to migrate the campaign back to a smart bidding strategy, but with a substantially more sophisticated setup. She has enabled value-based bidding using the merchant feed, has set custom values per product tier (so the algorithm now understands that a £180 order is worth more than a £106 order), and has restricted the campaign to a product feed segment that matches the premium-tier composition of the pre-migration mix. Under this configuration, the campaign is producing broadly the same reported performance as the manual version, with the operational benefit of automated bidding, without the AOV mix problem that produced the original loss.
What she'd tell another head of paid
The lesson Frida drew, when I asked, was not a lesson about smart bidding as a category. Smart bidding is, in her considered view, a substantial improvement over manual bidding for most accounts most of the time. The lesson was about the specific care required when migrating a campaign whose manual structure carries hidden information — hidden margin bias, hidden geographic bias, hidden product-mix bias — that the automated bidder cannot see.
Her practical rule, since the incident, is: before migrating any campaign to smart bidding, do a written inventory of what the manual structure is doing that the smart bidder will not be able to replicate. If the answer is "nothing that matters", the migration is safe. If the answer is "there is a hidden bias that is producing real margin", the migration must include a value-based-bidding component that captures the bias explicitly. Otherwise, the algorithm will optimise it away.
The £310,000 loss, in her assessment, was the cost of skipping this step. It is a cost that shows up in perhaps 15% of the smart-bidding migrations she has since audited on other accounts as a consultant. The lesson generalises. Not universally, but often enough to be worth stating.