Would those leads have arrived anyway, and the three comparisons that answer it

Counting an agency's leads accurately and proving the agency caused them are two separate jobs, and this page is about the second one. An agency claiming 120 leads is reporting attribution, which records which touchpoint was last seen before an enquiry. Causation is a different measure: how many of those enquiries would not have existed without the work. The gap between the two is almost always made of demand you already had, and it is measurable with three tests that a mid sized business can run without a data team. Hold one channel back, pause one campaign in a quiet period, and record the source at the point of enquiry from the human who sent it. Those three, run over a quarter, will tell you more than any dashboard.

If the argument you are having is about whether 120 is the right count in the first place, that is the reconciliation problem rather than the causation one, and it is worked through separately in our note on reconciling four different lead counts. Everything below assumes the count is agreed and asks the harder question.

Why the agency's number is larger than yours

The arithmetic is not usually dishonest. It is four separate counting choices, each defensible on its own, which stack up.

• Every platform claims the same conversion. Run paid search, paid social and email in the same month and each platform will count the order it touched. Add the three reports and you can exceed your own sales ledger.
• Branded search is counted as acquisition. Somebody typing your company name had already chosen you. Those clicks belong in a separate row from the non branded ones, always.
• View through windows are generous. A conversion credited to an advert that was displayed and never clicked is a modelling decision, not an observation.
• Existing demand arrives through new routes. Improve a page and some customers who would have phoned you anyway will arrive through it instead. The channel grows and the business does not.

Test one, the holdout, which is the strongest of the three

Pick two regions, two city areas or two customer segments that behave alike. Run the activity in one and not in the other, for at least six weeks, and compare the total enquiry count in each. Not the channel count, the total.

This works because the held back group tells you what would have happened anyway. It is the only one of the three tests that gives you a counterfactual rather than an inference, and for a national business with regional sales data it is the most convincing evidence available.

Two rules make it reliable. Choose the pair before you start, and write down which is which. Pick a stretch with no promotion, no seasonal peak and no price change in either group.

Test two, the pause, for paid media only

Turn one campaign off for two to three weeks in a flat trading period and watch total orders, total enquiries and total revenue rather than the paid channel's own numbers. If the business keeps selling the same volume, that spend was buying demand you already had.

Choose a quiet window deliberately, tell the sales team nothing in advance, and run it long enough to cover your normal sales cycle. A one week pause on a six week cycle measures nothing.

Paid media can be paused. Search and content work compounds and cannot be switched off and back on in a month, so it needs the third test.

Test three, ask the human

Put one mandatory question at the point of enquiry, asked by a person rather than by a form. "How did you come across us?" Record the answer verbatim in the CRM, as a free text field, and read the answers rather than counting them.

It is less precise than the other two and it catches things neither of them sees: the recommendation, the trade body, the article, the assistant that produced a shortlist with your name on it. When half your new enquiries say a friend recommended you, no amount of channel reporting was ever going to tell you that.

How to test search and content, where a pause is not available

Use the publishing schedule as the experiment, which it already is.

• Cohort the pages by publication month. Pages published in month one, two and three, each tracked as a group against pages that existed before the engagement.
• Separate branded from non branded in Search Console. Filter out every query containing your company name, then look at clicks for the remaining set only.
• Count enquiries by landing page. An enquiry that started on a page published eight weeks ago is attributable work. An enquiry that started on your homepage is not evidence about the new pages.
• Hold back a service. Where you sell three or four service lines, publish for two and not for the others. Same site, same brand, same authority, different treatment.

Settle the definition before the first invoice

Most disputes about lead quality are disputes about an undefined word. Write the definition down in month one, in the contract, as a test somebody can apply without judgement.

A quarter that answers the question

• Weeks 1 and 2. Agree the definitions above. Turn on the how did you hear about us question. Export a pre engagement baseline for total enquiries, by month, for the last year.
• Weeks 3 to 8. Run the holdout. Two matched groups, one treated, nothing else changed in either.
• Weeks 6 to 9. Run the pause on one paid campaign, inside the holdout window if trading is flat, after it if not.
• Week 12. Compare total enquiries between the two groups, total orders across the pause, and read the free text answers. Three readings, one decision.

The agency should be running this with you rather than against you, and a good one will suggest the holdout before you do. A supplier that resists a held back group is telling you something about its own confidence in the work.

Questions about agency lead attribution

Our agency reports more conversions than we have orders. Is that fraud?

Almost never. It is three platforms each counting the same order plus a modelled view through window. Ask for the figure from your own sales records, matched by order date, and use that as the only number in the results table.

Is last click attribution wrong?

It is one lens and it is useful for the final step. It gives no information about whether the enquiry would have happened anyway, which is the question you are actually asking, and that needs a comparison group.

How long does a holdout need to run?

At least six weeks, and longer than one full sales cycle. Shorter than that and the two groups differ by noise.

Can we hold back a region if we only sell in one city?

Use service lines or customer segments instead. The principle is the same: two similar groups, one treated.

What if the agency refuses to pause anything?

Pause it yourself for two weeks in a flat period and watch total revenue. The result belongs to you either way, and the refusal is worth noting in the same file.

Does branded search really need separating?

Yes, and it is the single highest value filter in the whole exercise. Brand demand is the output of everything you have ever done, so counting it as this month's acquisition inflates every figure in the report.

The enquiry nobody can attribute

Run the free text question for a quarter and a pattern appears in the answers that no platform reported: somebody asked an assistant for a supplier, got three names, and called one of them. There is no click, no referrer and no dashboard behind that enquiry, which is why it turns up in a sentence a human typed and nowhere else. AI Knows Us puts a number against that row, by asking those buying questions of ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews on a named business and recording which companies the answers put in front of a buyer. The counting rule behind figures of that kind is published in the 597 reading benchmark, and the British agency reporting and rules context is in the note on UK agency rules.