Measurement when the click disappears

Attribution needs a click, and a growing share of the moments that persuade a buyer produce none: an answer read on the results page, a recommendation from an AI assistant, a link pasted into a group chat. The buyer arrives days later by typing the brand, and the report credits brand search or "direct". A measurement budget built for a clicking world spends most of its hours on the one instrument that cannot see the change, and the cost shows up as a cut to whichever channel stopped producing clicks.

Three ways the click disappears

Zero-click search. The results page answers the question, and the reader never visits. The page did its work and the analytics show nothing.

AI answers. An assistant names the brand or cites a page, and the reader visits only if they follow the citation. Module 5's running site has 20,000 organic sessions a month, of which 120 come from AI surfaces, 0.6%; search after the answer engine works those numbers. The referrals understate the influence, because the reader who was told the name usually searches for it later rather than clicking.

Dark social. Links shared in messaging apps and email arrive with no referrer, so analytics files them as direct traffic. A colleague's recommendation is often the most persuasive touch in a business purchase and the least visible one.

All three end in the same place: a visit that attribution files under brand search or direct, the two buckets that reward whatever happened earlier without saying what it was.

Expand

An AI answer, a zero-click results page and a link shared in a chat all send the buyer back later through brand search or a typed address, with no referrer; only a paid ad leaves a click attribution can follow, and only the question asked at sign-up recovers the other three100%−+FitSVG (opens in a new tab)CloseEsc

Pinch to zoom · drag to panScroll to zoom · drag or arrow keys to pan · 0 fits · Esc closes

The report that misleads

The mistake first. Take a company larger than module 5's site, though with the same organic traffic: it spends $100,000 a month on marketing and gets about 600 sign-ups, of whom about 240 pay. Comparing the third quarter with the first:

The proposal on the table is to cut the $20,000-a-month content budget by a third, because "organic is declining". Read the whole table and the conclusion inverts: sign-ups are flat, the 3,000 sessions organic lost are mostly matched by 2,400 more in direct and brand search, and brand search rising 30% means more people arrive already knowing the name. The likeliest story is that readers stopped clicking through to the content, not that they stopped reading it. Last-click cannot tell those two stories apart; it can only report that the click moved.

Ask the buyer

The cheapest instrument that sees a clickless path is a free-text field on the sign-up form: "How did you hear about us?" Assume 70% answer, about 420 of 610 a month, and an hour or two to code the answers each month with a model's first pass and a person checking it. Suppose the coded answers come back as 25% search engine, 20% a friend or colleague, 12% an AI assistant, 10% a podcast, 8% our articles or newsletter, and 25% other.

12% of 420 is 50 sign-ups a month naming an AI assistant. Analytics, meanwhile, shows 120 AI-referral sessions; at an assumed 3% sign-up rate that is under 4 sign-ups. The two instruments disagree by more than ten times, and the disagreement is the finding. Self-report has its own biases (people remember the last thing, and the memorable thing), but they are different biases from last-click's, so where the two diverge something is happening that the clicks do not show.

Moving the weight

Four instruments, each answering a different question:

Attribution keeps a job: optimising inside a channel, where clicks still exist, such as which of two ad variants converts better. It loses the job of judging channels against each other. Experiments take the big decisions; a geo holdout measures sales in regions that did not see the spend, so it counts the customer who searched the brand after seeing an ad as readily as the one who clicked it, the logic behind when an A/B test is the wrong instrument. Mix modelling takes allocation, once there is enough history; marketing mix modelling covers what a regression on weekly spend can recover.

The measurement budget, re-planned

Assume the company spends 5% of its marketing on measurement, $60,000 a year, almost all of it analyst time at the course's $100 an hour, so 600 hours. Today 450 of them maintain attribution dashboards and a multi-touch model, and 150 go to one-off questions.

                         before      after      working
attribution              $45,000     $15,000    150 h: within-channel reports only
experiments              $0          $20,000    2 geo holdouts x 100 h
mix modelling            $0          $15,000    50 h to build + 4 quarterly refits x 25 h
brand tracking           $0          $10,000    24 h coding the form + 12 h search volume
                                                + 2 survey waves x 400 x $8 (assumed)
one-off questions        $15,000     $0         answered by the four above
total                    $60,000     $60,000

Two conditions sit on that plan. The mix model needs about two years of weekly data, 104 points, with enough variation in spend to separate channels; a company with 14 months of history moves the $15,000 into a third holdout and builds the model next year. And a holdout has a price beyond its hours. Holding paid social out of 20% of regions for four weeks, on $30,000 a month of social spend, withholds $6,000; if social acquires at the course's paid-media $333 a customer, about 18 customers never arrive, $14,400 of lifetime margin at the default $800 each, so the test costs roughly $8,400 net. Against a decision about $360,000 a year of social spend, that is cheap.

In practice

To the founder or the board: "Organic did not fall; clicks did. Sign-ups are flat, brand search is up 30%, and one answer in eight on the sign-up form now names an AI assistant." That reframes a content cut as a measurement question with a known answer path. To an agency: ask for incremental results from a holdout, not attributed conversions, before any budget moves between channels.

The mistake that costs money is cutting the channel whose clicks disappeared because attribution can no longer see it. Cutting a third of a $20,000-a-month content budget on a last-click report saves $80,000 a year and may remove the source of the brand searches that the same report credits to brand search.

Check yourself

1. A consultant offers to build a mix model for $25,000. You have 14 months of weekly data and four channels. Buy it?

No. 61 weekly points with four channels and seasonality cannot separate the channels' effects; the intervals will cross zero, as module 4's coefficient table shows. Spend the money on a geo holdout of the largest channel, whose answer is usable in six weeks, and revisit the model at 24 months of data.

2. The sign-up field shows "AI assistant" rising from 6% to 12% of answers over two quarters, while AI-referral sessions stay at about 120. Which do you plan with?

Neither as a count; the field as the trend. Referrals measure clicks, which this path does not produce, so a flat 120 is expected. The field doubling is the signal, and the decision it supports is cheap: keep the pages answer engines cite current and track brand search alongside. It does not justify a large budget on its own, because no instrument here measures what an AI mention causes.

3. A four-week holdout on paid social would cost about $8,400 net. Social is $30,000 a month. Run it, or keep using last-click?

Run it. Last-click credits social with few conversions because its buyers come back through brand search, so the report is biased against exactly this channel. $8,400 is under 3% of a year's social spend, and the result decides the other 97%.