The last click did not do all the work — stop paying it as if it did

Why last-click quietly defunds the work that is working

Last-click attribution has one enormous virtue: it is unambiguous. Somebody clicked, then they converted, and the click gets the credit. Nobody argues with it because there is nothing to argue about. That clarity is exactly why it survives, and it is also why it systematically misallocates budget in almost every local services account.

Consider the shape of a real decision. A homeowner notices a damp patch. Over the next three weeks they see a remarketing ad twice, read a blog post about what causes it, look at two competitor sites, ask a neighbour, and then — when they finally decide to spend the money — type the company name into Google, click the top result, and ring. Last-click hands every penny of that job to a branded search click that cost forty pence and did nothing except be present at the end.

Now watch what happens next, because this is the part that costs agencies accounts. Branded search looks phenomenal, so it gets more budget. The blog post produced no conversions, so content gets cut. Remarketing shows a poor direct return, so it gets cut. Six weeks later branded search volume falls, because nobody is being introduced to the brand any more, and the only channel that ever looked good in the report is now producing less. The client sees a decline and asks what changed. The honest answer is "we defunded the top of the funnel because our measurement could not see it", and almost nobody says that out loud.

The mirror image happens to first-touch, which is why swapping one for the other is not a fix. First-touch loves whatever channel introduces the most people and takes no view on whether any of them ever bought. Run first-touch on a local account and display advertising will look like the engine of the business, because display is very good at being the first thing somebody saw and historically poor at being the reason they purchased.

The useful conclusion is not that one model is right. It is that a single model is a position, and a position you cannot compare against alternatives is indistinguishable from a guess. What changes the conversation with a client is being able to show the same set of real leads scored five different ways, and to say honestly: here is the range, here is where the models agree, and here is the one channel where they violently disagree — that disagreement is the thing worth talking about.

How a journey gets assembled

Attribution is only as good as the touch data underneath it. Most of the difficulty is not in the maths — the models are arithmetic — it is in stitching together a journey that crosses sessions, devices and channels, and ends in a phone call nobody clicked.

1. 1

   Every arrival is recorded as a touch

   Source, medium, campaign, term, content, the click identifier the platform passed, the landing page, the referrer and the timestamp. Not a session count — an individual, dated, attributed arrival that can be listed.

2. 2

   Returning visitors are recognised

   A first-party identifier survives the visitor leaving and coming back, so the March click and the May click are known to be the same person rather than two unrelated strangers who happen to have the same interest.

3. 3

   The conversion is joined to the journey

   A form submission carries its session. A phone call is joined by the tracked number the visitor was shown. Either way the lead lands attached to the whole chain of touches, not just the last one.

4. 4

   Credit is allocated by whichever model you asked for

   The same journey, scored five ways. Nothing is recomputed destructively — switching models is a view, so you can hold the same period and the same leads still and change only the question being asked.

5. 5

   The channel report is the sum of those allocations

   Which means a channel can hold 4.3 conversions rather than 4, and that fraction is the point: it is the share of journeys it contributed to without closing.

The five models, and what each one is honestly for

None of these is the truth. Each is a defensible way of dividing credit that cannot be observed directly, and each is wrong in a specific, predictable direction. Knowing the direction is what makes them useful.

The part everybody else skips: attributing the phone call

Almost every attribution product on the market models the journey right up to a click on a submit button and then stops. For a SaaS company that is fine, because the conversion genuinely is a form. For a roofer, a solicitor or a dental practice it is close to useless, because most of the money arrives through the telephone and the telephone has no referrer.

This is where a lot of agency reporting quietly becomes fiction. The form leads have beautiful multi-touch journeys attached. The calls — which might be sixty or seventy per cent of the actual business — arrive as an undifferentiated block labelled "phone", sitting outside the attribution model entirely. Then a channel comparison is run over the third of leads that happened to be forms, and a budget decision is made on it.

Because dynamic number insertion assigns a number to the session, a tracked call arrives already joined to the same journey a form submission would have been joined to. It has a first touch, a last touch and everything in between. It scores under every model exactly as a form does. And so a channel comparison is finally being run over all of the leads rather than the convenient subset.

The practical effect on a local account is usually dramatic and usually in the same direction: organic search and Google Business Profile look far better than they did, because both produce disproportionately many calls and disproportionately few forms. Agencies who have been quietly under-selling their own SEO work for years find the evidence for it was there the whole time, sitting in a phone log that was never joined to anything.

What you can actually answer once journeys exist

These are the questions that come up in real client meetings, and the ones a single-model report cannot answer without hedging.

Lookback windows, and why the default matters more than you think

A lookback window is the period before a conversion in which a touch is still considered part of the journey. Set it at 30 days and a customer who first heard of the business in February and bought in April has their entire introduction erased; the journey starts at whatever they happened to do in the final month. Set it at 365 days and every conversion drags along a year of incidental brushes with the brand, most of which had no causal role at all.

The default here is 90 days, which is long enough to hold the considered purchases that dominate home improvement, legal and healthcare, and short enough that credit does not diffuse into meaninglessness. It is adjustable per client, because a 90-day window is obviously silly for an emergency locksmith and obviously too short for a firm selling twenty-thousand-pound extensions.

The important discipline is to pick the window before you look at the result. Attribution windows are the easiest knob in marketing to turn until the answer flatters the channel you were already planning to recommend, and everyone who has ever built a client deck knows the temptation. Setting it per client, once, on the basis of how long that business's customers actually take to decide, is the difference between measurement and advocacy.

The same applies to model choice. Choosing position-based at the start of an engagement and holding it is defensible. Choosing whichever of five models makes this quarter look best is not, and clients who have been through it once can smell it. The comparison view exists so that model disagreement is something you show the client, not something you shop through in private.

What is honestly measurable, and what is not

Attribution is a field with a lot of confident nonsense in it. Here is the line, drawn where it actually falls.

• Tracked digital touches: yes. Any arrival on the site with a source attached, across sessions, joined to the lead it eventually produced. This is a record, not an estimate.
• Cross-device journeys: partly. Phone-then-desktop is recovered when the visitor identifies themselves — a form fill, a login, a call from a number already seen. Otherwise it is two journeys, and the report says so rather than inventing a join.
• Offline influence: no. A van, a neighbour's recommendation, a leaflet. Where those matter, and for trades they matter enormously, the tracked picture is a subset and should be presented as one.
• Blocked and consent-declined visitors: no. A visitor who declines tracking has no journey. That is a real hole, it is reported as unattributed rather than redistributed, and a report with an honest unattributed bucket is worth more than one without.
• Walk-ins and inbound email: no. Nothing digital happened. Some businesses get a third of their work this way; if yours does, the tracked total is not the business total and the client should hear that from you first.
• Which touch caused the sale: nobody can. Attribution allocates credit under a stated rule. It does not establish causation, and any vendor telling you otherwise is selling a model as if it were a measurement.

What this is worth to an agency, in plain terms

The commercial case is not that better attribution improves the campaign. Sometimes it does. The reliable case is that it changes what you can defend — and agency revenue is lost far more often to an undefended channel than to a badly run one.

The typical pattern: an agency runs SEO and paid search for a home services client. Paid search shows a clean, attributable return. SEO shows traffic and a modest number of last-click conversions. At the eighteen-month mark the client, entirely reasonably, asks why they are paying a monthly retainer for the channel with the worse numbers. The agency knows the answer — organic is introducing most of the people who later convert through brand and through the phone — but knowing it and evidencing it are different things, and without the evidence the retainer gets cut.

With journeys assembled and calls included, that conversation has a different shape. Organic holds a large share of first touches. A material fraction of the calls that closed have an organic touch somewhere in the chain. The assisted number is visible. The client can see the same data you can, and the decision they make is an informed one — which, notably, is sometimes still to cut, and that is fine. An informed client who stays is worth more than an uninformed one who was going to leave at the next review.

There is a second, less obvious effect. Once an agency can see assisted conversions honestly, it starts making better internal decisions about where to put its own effort — which content to commission, which campaigns to build, which channels to stop pitching. A surprising number of agencies discover they have been quietly subsidising a channel that never worked, protected from the truth by exactly the same measurement gap that was hiding their best work.

How it fits with the rest of the system

Attribution is downstream of capture and upstream of economics. It needs call tracking and form capture underneath it, because a model over an incomplete set of conversions is a model of the wrong thing. And it feeds cost per lead above it, because once credit is allocated across a journey, spend can be divided against the credit rather than against the last click.

That chain is why the pieces are one product rather than four integrations. The moment attribution lives in a different system from the call log, somebody has to reconcile two lead counts, and the reconciliation is always the first thing to be abandoned when the month gets busy.

Downstream, the client-facing report shows whichever model you have chosen, labelled, with the window stated. Not because clients want to read about attribution theory, but because a number with its method attached survives scrutiny and a number without one does not.

Common questions