Working out what a customer costs is cheaper than the advertising most local businesses buy on instinct

The one number that decides all the others

Ask the owner of a local business what they spend on marketing and the answer arrives in seconds. Ask what a new customer costs them and the answer turns into a story: a good month, a bad month, a leaflet that worked once. The distance between those two answers is where most local marketing budgets are set.

It is a large amount of money set that way. There were 5,690,265 private sector businesses in the UK at the start of 2025, and 4.3 million of them, roughly 75 per cent, employed nobody besides the owners ^[2]. UK advertising spend reached an estimated £46.9bn in 2025, with search and online display formats accounting for 83 per cent of spend in the third quarter ^[21]. Little of it is advised upon: the Longitudinal Small Business Survey found that 27 per cent of SME employers had sought external information or advice in the preceding 12 months ^[16].

The argument here is narrow. Cost per enquiry and cost per customer are arithmetic, not analytics, and almost every local business already holds the inputs. The reason they so rarely get worked out is that the dashboards nearest to hand answer a neighbouring question confidently enough to feel like an answer, and the neighbouring question is the wrong one.

Four numbers, and only one of them is genuinely hard

The vocabulary is worth fixing first, because most of the confusion in this subject is vocabulary wearing a suit.

• Cost per enquiry is everything spent on winning work in a period, divided by first contacts from people who were not already customers.
• Cost per customer, or customer acquisition cost, is the same spend divided by the number of those contacts that became a paying job or appointment.
• Gross margin per customer is what the customer paid, less the direct cost of delivering the work, before overheads.
• Lifetime value is the total gross margin a customer is expected to produce across every job they will ever bring.

The first three are division. They need a bank statement, a diary and a decision about what counts. The fourth is a forecast, and forecasts about human behaviour are a research field rather than a spreadsheet cell. Fader, Hardie and Lee set out a repeat buying model whose output feeds lifetime value calculations, and state plainly that it must be augmented by a model of purchase amount first ^[7]. Even the published methods treat lifetime value as two forecasts stacked on each other: how often, and how much.

A crude lifetime value is still better than none, provided the crudeness is declared. Average gross margin multiplied by the number of visits customers have historically made, over a period short enough to have been observed, is defensible. A ten year horizon invented to make a budget look sensible is not.

What the published benchmarks say, and what they quietly do not

Published cost per lead benchmarks do exist. The most widely circulated set comes from LocaliQ, working with WordStream, drawn from what it describes as data from thousands of its customers' campaigns across Google Ads and Microsoft Ads ^[9].

Its 2026 table puts the average cost per lead for search advertising at 66.69 US dollars. Dentists and dental services appear at 72.97 dollars on a conversion rate of 10.67 per cent, physicians and surgeons at 40.04 dollars, attorneys and legal services at 131.63 dollars on 5.55 per cent, and beauty and personal care at 39.25 dollars ^[9].

Three limitations matter. The currency is United States dollars and the publisher does not state which countries are covered ^[9]. The sample is self selected, consisting of businesses that already advertise through one provider. And a lead, in that table, is whatever each advertiser configured as a conversion action, which is the next and larger problem.

Why the lead count on the dashboard is not the lead count in the diary

This is the part that does the most damage, because it is invisible. The free reporting attached to search advertising and to a Business Profile is accurate about what it measures, and what it measures is not what an owner means by an enquiry.

Counting rules decide the answer before the data arrives

Google Ads allows an advertiser to count every conversion from a click, or only one. Its documentation frames counting every conversion as the choice for sales, because every sale likely adds value for your business, and counting one as the choice for leads, because usually only one, unique lead per ad click probably brings value ^[12]. An account left on the wrong setting reports a lead count inflated by people who filled in a form and then rang as well, and the same documentation warns that a double tagged site can cause duplicate conversion reporting ^[5].

A tap is not a conversation

Calls are where the gap opens widest. A Business Profile reports calls as the number of times a customer clicked on the call button on the profile ^[4]. In an advertising account, call clicks are the number of times that potential customers tapped the Call button in an ad from mobile devices, and the documentation describes the figure as the approximate number of calls coming from the ad ^[20]. Neither metric claims that anybody spoke to anybody.

Call reporting through a Google forwarding number records more, including call duration and the caller's number for calls over 15 seconds, and allows calls of a specified duration to count as conversions ^[11]. Google's documentation lists the caveats: the forwarding numbers belong to Google and can be reassigned, caller ID sometimes shows a Google number rather than the caller's, and a local number may not always be available ^[11].

Consent, and the conversions nobody observes

Behind all of it sits the law. The Information Commissioner's Office requires prior consent to the UK GDPR standard for storage and access technologies unless an exemption applies, and is unambiguous that analytics is not exempt.

You still must obtain consent for all analytics purposes ^[17]

The same guidance says online advertising purposes, including ad measurement, are not exempt from the consent requirements of the Privacy and Electronic Communications Regulations and never have been ^[17]. Every visitor who declines is a hole in the measurement, by design and by law.

Google's answer to those holes is modelling. Its documentation says that without cookies advertisers lose visibility into user paths on their site, and that conversion modelling uses Google's AI to analyse observable data and historical trends ^[6]. Modelled conversions then appear in the Conversions column alongside observed ones.

There is a threshold, and it is the detail most relevant to a single site advertiser. Google states that modelling requires a daily ad click threshold of 700 ad clicks over a 7 day period, per country and domain grouping ^[6]. A business spending a few hundred pounds a month will not reach it, so its reported conversions are the observed ones only, an undercount of unknown size.

Working the number out from what a business already has

The honest method does not start in a dashboard. It starts with a bank statement, a telephone log and a diary, and it takes an afternoon. Take one month and gather four quantities.

1. Acquisition spend. Everything paid out that month to make the telephone ring: advertising, agency or freelancer fees, leaflets, signage, sponsorship, the share of the website and booking system that exists to win work, and staff time spent on marketing rather than delivery.
2. Enquiries. Every first contact from someone who was not already a customer, counted once per person, from the telephone log, the inbox, the messaging apps and the booking system, with duplicates merged.
3. New customers. The enquiries that turned into a paid appointment or job.
4. Gross margin per customer. Price paid less the direct cost of delivery, before overheads.

Then divide. Spend over enquiries gives cost per enquiry, spend over new customers gives cost per customer, and set that beside gross margin to see whether the month paid for itself.

Suppose a business spends nine hundred pounds across a month, logs 60 first contacts, converts 18 of them, and earns an average gross margin of 140 pounds a customer. Its cost per enquiry is 15 pounds, its cost per customer is 50 pounds, and it is holding gross margin of a little over two and a half thousand pounds against nine hundred pounds of spend.

Change one input and the conclusion inverts. Had only 6 of those 60 contacts converted, the cost per customer would be 150 pounds against a margin of 140, and the month would have lost money while the advertising account reported a respectable cost per lead. That is why the conversion step, which happens on the telephone rather than online, is usually the cheapest thing to fix.

Anyone running automated bidding already declares a version of this number, calculated or not. With Target CPA bidding the advertiser sets the desired average cost per conversion, and Google states that some conversions may cost more than the target and some less ^[19]. Setting it without knowing gross margin per customer is a guess wearing the clothes of a setting.

Word of mouth, and the arithmetic of the unmeasured channel

“Most of our customers come by word of mouth” is the commonest answer to the question this article is about. It is often true. It is almost never measured, and the tools nearest to hand make it look truer than it is.

In Google Analytics, Direct is defined as the channel by which users arrive at your site or app via a saved link or by entering your URL, and the rule assigning it fires when the source matches (direct) and the medium is (not set) or (none) ^[3]. Anything the rules cannot place is labelled Unassigned, which Analytics uses when there are no other channel rules that match the event data ^[3].

Direct is therefore a residue as much as a channel. A link pasted into a family group chat, a leaflet QR code without a tracking parameter, a tap from an app that strips the referring page: all land in the same bucket as genuine recommendation. A business reading a large Direct number as proof of reputation may be reading proof that its leaflets are untagged.

What credit remains is assigned by a model the advertiser no longer freely chooses. Google states that the first click, linear, time decay and position based attribution models are no longer supported, that conversion actions which used them have been upgraded to data driven attribution, and that last click is still supported ^[1]. Data driven attribution is now the default for most conversion actions ^[1].

The path being modelled is also getting longer. Ofcom reports that Google Search is used by 82 per cent of UK adults, that about 30 per cent of searches now show AI overviews, and that ChatGPT recorded 1.8 billion UK visits in the first eight months of 2025 ^[8]. More steps in a journey, and more of them inside systems that report nothing back to the business being chosen.

The practical response is not better software. It is one question, asked at booking and recorded in the same field every time: how did you hear about us. It is crude, and it remains the most reliable attribution instrument available at this size.

What the experiments say about crediting advertising at all

Even a clean cost per customer is an accounting figure, not a causal one. It says what was paid and what arrived. Whether the payment caused the arrival is answerable only by experiment.

Large advertisers have run them, and the results are sobering. A field experiment at eBay by Thomas Blake, Chris Nosko and Steven Tadelis switched off paid search advertising and watched what happened.

natural search is close to a perfect substitute for paid search ^[14]

The authors report that brand keyword ads have no measurable short term benefits, that most, if not all, of the lost traffic finds its way back through natural search and direct navigation, and that new and infrequent users are positively influenced by ads while frequent buyers were not significantly affected ^[14].

Measuring such effects reliably is harder than it looks. Randall Lewis and Justin Rao analysed 25 large digital advertising field experiments covering about 2.8 million US dollars of spending ^[15].

The median confidence interval on return on investment is over 100 percentage points wide. ^[15]

They note that individual level sales are so volatile that a coefficient of variation of 10 is common, and that informative advertising experiments can easily require more than 10 million person-weeks of observation ^[15]. A clinic with two hundred enquiries a month is several orders of magnitude short of that.

Two conclusions follow, pulling in opposite directions, which is why both need stating. A small business cannot prove the causal return on its advertising, so confident claims of return deserve suspicion, including from whoever is selling it. Equally it cannot prove waste, so switching everything off after a quiet month is not evidence based either. The affordable middle is a holdout: turn one channel off for a set period and watch total enquiries rather than that channel's reported ones.

How long a customer lasts, and how few of them are needed

Lifetime value calculations assume a long life for two parties, the customer and the business. The second gets far less scrutiny than it deserves.

The Office for National Statistics puts the five year survival rate for UK businesses born in 2019 at 38.4 per cent, and recorded 317,000 business births against 280,000 business deaths in 2024, the death counts provisional ^[13]. A lifetime value built on a decade of repeat custom is also a forecast that the business will be there to receive it.

The counterweight is how few customers a local business needs, which is what makes the arithmetic worth doing. VAT registration becomes compulsory when total taxable turnover for the last 12 months goes over the threshold of 90,000 pounds, or when turnover is expected to pass it within the next 30 days ^[18]. A business trading near that level, earning a few hundred pounds of margin a job, runs on a few hundred customers a year rather than a few hundred thousand.

At that scale a twenty pound error in cost per customer is not a rounding difference, it is a misallocation repeated several hundred times. It is also why an afternoon spent on the sums returns more than most of the tactics it will be used to judge.

What it means for a local business

Five things follow, and none requires new software or a new supplier.

1. Count first contacts by hand for one month, once per person, across every channel, with a note of where each said they came from. Decide what counts as an enquiry before counting, not after.
2. Match the advertising account to that definition. Where the goal is a lead rather than a sale, count one conversion per click rather than every conversion, and set conversion actions as primary or secondary deliberately ^[12][5].
3. Work out gross margin per customer before setting any acquisition target, because a Target CPA figure is only meaningful as a fraction of the margin it buys ^[19].
4. Read reported conversions as a floor rather than a truth. Taps on a call button are not conversations ^[4][20], and modelled conversions will be absent entirely from accounts below the stated click threshold ^[6].
5. Test by switching something off for a defined period and watching total enquiries, rather than by reading the same dashboard more closely.

This produces a number with known error bars rather than a precise one, from records the business already keeps, which is the difference between a decision and a hunch. It also reframes the usual question. The question is not whether to spend more on advertising. It is whether the business knows what it currently pays for a customer, and whether that figure sits comfortably inside the margin the customer brings.

What remains uncertain

Several things here are unsettled, and the method above should not be presented as more solid than it is.

The benchmarks are the weakest evidence. The published cost per lead figures come from one provider's own advertisers, the publisher does not state which countries the data covers, and the figures are in United States dollars ^[9]. No comparable open UK dataset for local service sectors appears to be published, so UK businesses are benchmarking against a market that may not be theirs.

The incrementality evidence is strong but narrow. The eBay experiment was run by a very large retailer with an exceptionally well known brand, and its finding on brand terms does not transfer automatically to a business nobody is yet searching for by name ^[14]. The warning from Lewis and Rao cuts both ways: if reliable measurement needs samples far beyond a small advertiser's reach, that argues against confident claims of effectiveness and of waste alike ^[15].

Lifetime value for local services has no good public benchmark. The published models require a separate model of purchase amount before they can produce a lifetime value at all ^[7], which is to say the figure is an estimate built on an estimate. For a clinic or a firm it will be an internal number for some time yet, and should be labelled as one.

Finally, the measurement ground keeps moving. Google has withdrawn four attribution models and upgraded the accounts that used them ^[1], renamed what used to be conversions as key events in Analytics in order to align the definitions between Analytics and Ads ^[10], and made estimated conversions part of the reported total for accounts large enough to qualify ^[6], while the ICO's position is that consent is required for all analytics purposes ^[17]. Any method that depends on a particular dashboard behaving in a particular way will need revisiting when it changes. A method built on a bank statement, a diary and one question asked at booking will not.