Waterproofs, fleeces and walking gear sold online across the UK. 60 staff.
Verdict: 5 channels. Spend: £228,000 last year; £211,000 next. The platforms claim 14,200 customers and your records credit 9,340, but only about 3,711 were truly extra: £67 each, against £83 of gross profit. Ads netted £59,158 last year; next year’s budget should net £80,208, a change of £21,050 (next year’s net would be -£21,861 if every channel were a third less extra than assumed: that’s below last year, so test before moving budgets). Best: “Google Shopping and generic search” (£56,400); worst: “Influencer gifting” (-£11,380). 3 channels are judged without a test. At the margin, the cheapest extra customer is on “Google Shopping and generic search” (£65).
Light: what the platform claims. Mid: what Corrinbeck’s own order records credit. Dark: the customers who wouldn’t have bought anyway.
| Channel | Spend last year | Per extra customer | Net last year | Spend next year | Net next year | Change |
|---|---|---|---|---|---|---|
| Google Shopping and generic search Generic search · 60% truly extra (not tested) | £96,000 + £6,000 other | £53 last: £65 | £56,400 | £110,000 | £59,887 | £3,487 |
| Google brand search Brand search · 15% truly extra (before-and-after test) | £18,000 | £46 last: £138 | £14,175 | £6,000 | £16,309 | £2,134 |
| Meta new audiences Prospecting · 55% truly extra (held-back test) | £72,000 + £9,000 other | £78 last: £90 | £5,213 | £80,000 | £4,491 | −£722 |
| Meta retargeting Retargeting · 20% truly extra (not tested) | £30,000 | £100 last: £130 | −£5,250 | £15,000 | −£478 | £4,772 |
| Influencer gifting Other · 40% truly extra (not tested) | £12,000 + £4,000 other | £286 last: £279 | −£11,380 | £0 | £0 | £11,380 |
Only 15% of brand-search customers are extra (a two-week pause showed most came anyway), and spending more can’t create more searches for the name, so cutting it from £18,000 to £6,000 loses little. Retargeting, untested, reaches people already coming back: halving it is right, and a held-back test should decide the rest. Adding £8,000 to Meta new audiences buys customers at a dearer rate than they’re worth. Net includes other costs: agency fees, creative and gifted stock. Ads are credited with 44% of all new customers.
Spend and new customers come from the business’s own records (not the platforms’ claims). The share truly extra comes from a held-back or before-and-after test where there is one; otherwise 60% for new audiences, 60% generic search, 15% brand search, 20% retargeting, 40% other. Each extra customer is worth the first order × 50% margin + £40 of later orders (undiscounted; check it against your repeat figures). Net is after spend and other costs (fees, creative, stock), not overheads. Customers follow spend on a curve (added budget about 30% dearer at the margin; a cut removes the least useful spend first, and no spend brings none). Your records credit each customer to one source, which is itself a model: channels that help each other (an ad seen, then a search) are judged one by one, and brand search may take credit for others’ work.
Three channels rest on cautious shares, not tests. A held-back test (hide the ads from a random 10% for a month) on retargeting and generic search would turn those into measurements.
Corrinbeck chose the guided route and started the new budget split on 1 November. At the last session, on 23 October, the agent was set up in an AI assistant the business already uses, connected read-only to the Google and Meta invoices and to the shop's order records (its IT confirmed this before it committed). On the first Tuesday of each month it loads the month's spend and the new customers each channel brought, by the first order's source, into a new dated copy of this review, and sends the managing director a short note. It uses the rules on Inputs: a channel's share of the month's spend more than 3 points from the plan, or its cost per customer in the shop's records more than 30% above the plan's, with at least 20 customers to judge. Nothing left the business, and no decision changed, without one of Corrinbeck's own people approving it. These events continue the fictional example.
| Date | What the agent found | What people decided |
|---|---|---|
| 1 Dec 2026 Note only | November, the first month of the new split: every channel within 3 points of its share and within 30% of the plan's cost per customer. | Nothing needed. The managing director read the monthly note. |
| 5 Jan 2027 Flag | Meta retargeting was off the split: 11.5% of December's spend against 7.1% in the plan (£2,150 against £1,250). The agency had added a Christmas retargeting audience, the spend the review had halved. | The managing director asked the agency to take retargeting back to £1,250 a month from January, and said any change to the split now needs the managing director's yes. |
| 2 Feb 2027 Flag | Generic search was dearer than the plan: £40.87 per customer in the shop's records in January (230 customers) against the plan's £30.96. | No change. January is the quietest month for waterproofs and last January's orders showed the same dip. The managing director chose to watch February. |
| 2 Feb 2027 Note only | Meta retargeting back within the split at 7.4% of January's spend. | Nothing needed. |
| 2 Mar 2027 Note only | February: every channel within the split; generic search back to £32.97 per customer, 6% above the plan. | Nothing needed. |
At the check-in on 7 December, after the first monthly run (about day 45), Geoff and the managing director read that run's log together and left the rules as they were. After that the agent, its accounts and its log were Corrinbeck's.
What the agent added. The method is unchanged: the same spreadsheet, the same checks and the split the managing director decided. The agent made sure the split was checked every month, and caught the agency's Christmas retargeting in the first week of January, before a second month at that rate. It also put a dear January in front of a person, who decided it was the season and not the channel. Checking the split took the marketing manager reading a note instead of gathering the invoices and orders by hand (estimate: about 20 minutes a month). The held-back test on retargeting is still to run; the full review is rerun when it reads. The months after add nothing to the figures above.
Fictional and labelled: these events continue the worked example. The same events are the sample rows in the Agent log of the worked-example spreadsheet, which also shows February channel by channel against the plan. The agent is part of the guided and done-for-you routes; the self-implement kit does not include it.