Display advertising attribution: how to tell if programmatic ads bring in sales

Prospecting display reaches so many people that plenty of your customers will have been served an ad before buying, whether or not it did anything. View-through counts measure that overlap, not cause. Judge display on matched sales from leads it recorded first, keep junk placements out, and when spend is large, settle it with a geo holdout.

A roofing company in a metro area of about 800,000 households pays an agency $9,000 a month for prospecting display: a line on The Trade Desk, a line on StackAdapt and a Google Display Network campaign. The monthly report shows 3.1 million impressions, 2,600 clicks, and 61 conversions at $148 each. Forty-nine of those conversions are view-throughs. The agency calls the campaign efficient, and on its own numbers it is.

The owner's question is simpler: how many roofs did it sell? Nobody in the meeting can answer, because the conversions are web events inside the ad platforms and the roofs are invoices in the job software. This post is about connecting the two, and about the uncomfortable fact that a broad display campaign generates most of its own conversion count by coincidence.

What prospecting programmatic display is

Prospecting programmatic display is banner and native advertising bought by automated auction to reach people who have not visited you before, chosen by audience, location, context or lookalike modeling. It is bought through demand-side platforms or Google's Display Network. Its job is to create new demand, which is exactly what its standard reports cannot show.

• Google Display & Video 360, The Trade Desk, StackAdapt. Demand-side platforms (DSPs) that buy impressions across ad exchanges, usually through an agency, with audience data layered on top.
• The Google Display Network. Bought inside Google Ads, across sites and apps that sell space through Google. Not called a DSP, but the auction, the reporting and the problems are the same.
• Native and contextual lines. Ads styled as content, or placed next to articles about a topic, often sold as part of the same DSP buy.

Three neighbors have their own posts and are not repeated here. Showing ads to people who already visited is retargeting, which has a different and worse inflation problem. Connected TV and online video, which DSPs increasingly bundle into the same buy, are covered in video attribution. Placements inside mobile games and utilities, where accidental taps dominate, are in in-app ads. Ask the agency to split the report by these lines before reading any total.

Impressions, clicks, and what each one can prove

An impression proves an ad was served to a device. A click proves someone touched it, which may not mean interest. Neither proves a sale. For a business that sells by phone or in person, the only display metric that can be tied to revenue is a lead that carries a display source and later matches a closed sale.

Read the table top to bottom and the metrics get scarcer and harder to fake. Millions of impressions become a few thousand clicks, a few dozen leads and a handful of sales. A report that stops at the top three rows is reporting delivery. The roofer's 2,600 clicks on 3.1 million impressions is a click-through rate of 0.08%, so even the clicks are a small fraction of what the budget bought.

View-through conversions and their windows

A view-through conversion is a site event, such as a form or a call-button tap, recorded after someone was served an impression and did not click. Its window is how long after the impression the event can still count. The longer the window and the broader the audience, the more view-throughs a campaign collects whether or not it works.

DSPs let the advertiser or the agency set their own windows, and reports rarely print them. Seven, 14 and 30 days are all common. The setting matters more for prospecting than for anything else, because a prospecting campaign is designed to reach as much of a market as possible, and every person it reaches becomes eligible to convert by coincidence.

Here is that coincidence worked through for the roofer, assuming the ads did nothing at all. The campaign reaches about 400,000 of the metro's 800,000 households in a month, so its 3.1 million impressions average 7.75 per reached household. The website takes 120 form submissions a month from all sources, and half of them, 60, come from reached households. Spread evenly across 30 days, the chance that a reached household was served an impression in the last d days is 1 − e^(−7.75 × d ÷ 30).

The roofer's real report showed 49 view-throughs on a 7-day window. The table produces 50 from a campaign that changed nobody's mind. That does not prove the campaign is worthless; it proves the view-through count cannot tell a working campaign from a useless one, because both land on the same number. Lengthening the window from one day to 30 multiplies the count by four without a single extra customer. The general case is in attribution windows.

Why display conversions are often people who would have bought anyway

Display reaches people who are online a lot, in the market and nearby, which also describes people about to buy. Targeting picks them on purpose; conversion-optimized bidding finds them faster. So viewers and buyers overlap heavily even when the ads change nothing. Comparing exposed and unexposed people measures that selection, not the ad.

The same paper's third experiment is even more pointed. Users were much more likely to sign up at the advertiser's competitor on the day they saw the advertiser's ad, and a control group that saw an unrelated ad showed a nearly identical spike, so the causal effect was close to zero. Exposure was simply a marker for being online that day. A view-through report is built on exactly this comparison, and so are the "audience lift" charts many DSPs supply.

Targeting makes it worse in a predictable way. An in-market audience for roofing is people who have recently searched for or read about roofing. A lookalike audience is people who resemble past customers. A conversion-optimized bid strategy learns which impressions precede conversions and buys more of them. Each of those does its job, and each concentrates impressions on people already likely to call, so each raises the coincidental view-through count in the table above.

But display can sell, mostly without a click

None of this means display does nothing. The best experiment on display and in-store sales found a real, profitable lift that came almost entirely from people who bought offline and never clicked. That is why click-based matching undercounts display while view-through counts overcount it, and why neither is the answer on its own.

Put the two experiments side by side and the measurement problem is clear. Observational comparisons can be wrong in either direction, by multiples. A business that sells in person or by phone should expect most of a working display campaign's effect to arrive without a click, so click-sourced leads matched to sales give a floor that may sit well below the truth. View-through counts give a ceiling that may sit well above it. Only an experiment says where in between the truth lies.

Accidental clicks and made-for-advertising sites

Cheap display inventory buys cheap clicks, and many are not interest: fat-finger taps on phones, ads placed beside buttons, and sites built only to resell traffic as impressions. These made-for-advertising sites inflate impressions and clicks while producing almost no leads. The fix is placement-level evidence: which sites the money went to, and which sites produced leads that bought.

Those are big-brand budgets, but a local campaign is more exposed, not less: a small budget on default settings drifts to the cheapest inventory the auction can find, and the cheapest inventory is where made-for-advertising pages and accidental mobile taps live. Accidental clicks also hurt the click-based floor in a second way. They inflate the click count, so any "cost per click" or "landing-page visits" metric improves while leads stay flat. The app-specific version of this, with Google's own data on accidental taps, is in in-app ads.

1. Get the placement report. Every DSP and Google Ads can list spend, impressions and clicks by domain and app. Ask for it monthly, in full, not as a top-20 summary.
2. Put the placement on the lead. Google Ads can write the site or app into the landing-page URL; DSPs generally offer a site or domain macro too; ask the agency which. Save it in a hidden form field so a web lead says not just "display" but where.
3. Judge placements on leads that bought, not on clicks. A domain with 900 clicks and no matched sales is a cost; a domain with 40 clicks and two roofs is the campaign.
4. Prefer an inclusion list. The ANA recommends buying from a list of approved sites rather than trying to exclude bad ones, because there are too many to exclude.

Geo holdouts and matched-market tests: the honest measure

A geo holdout turns display off in randomly chosen regions and leaves it on in comparable ones, then compares closed sales. It measures what display causes, including sales from people who never clicked, without needing any impression data. It is the only method here that answers the owner's question, and it needs more volume than most local businesses have.

The design, run on your own closed sales rather than a platform's conversions, is set out in incrementality tests. What matters for display is the arithmetic of size. With sales that vary at random month to month, detecting a lift of d at the usual 95% threshold, with even odds of success, needs roughly 7.68 ÷ d² sales in each group during the test. That ignores the extra noise between regions, so treat it as a floor.

So the roofer, closing about 70 jobs a month and splitting its market in two, can find out within a season whether display adds 20% or more, and can never find out whether it adds 5%. That is still worth knowing. A display line costing $9,000 a month on a business whose 70 jobs bring in $770,000 needs to add only about one job to cover its media cost, so a test that rules out a large effect is a strong argument for a smaller budget, and a test that finds one is the only evidence that should justify a bigger one.

Matched-market tests, which pick a few similar regions by hand rather than randomizing many, are the realistic version for regional and multi-location businesses. They are weaker because the regions can differ in ways nobody chose, so run them longer, compare against the same months last year, and state the confidence interval rather than the point estimate. A single-metro business without enough regions can turn display off for alternating months instead, which is weaker again and still better than reading the view-through column.

Tracking phone leads from display

Display leads reach a phone business three ways: a click that becomes a form, a click that becomes a call from the landing page, and a call from someone who saw a banner and never clicked. Give each its own trace: a source tag on clicks, a tracking number on the landing page, and a dedicated number in the creative.

1. Tag every click with its own source. A UTM source and campaign that separate prospecting display from retargeting and from each DSP line, carried into a hidden field on forms. The rules for writing them so they survive are in UTM parameters.
2. Swap the landing-page number by source. Dynamic number insertion from a call tracker shows display visitors a number that belongs to display, so their calls arrive tagged. See call tracking for getting the caller's number into the lead file.
3. Print a dedicated number in the creative where the format allows. Someone who sees a banner and calls an hour later leaves no click. A number used nowhere else catches some of those people, and it is the only non-click display signal that ends in a lead with a contact.
4. Keep one lead file with dates. Calls and forms from every channel in one export, each row with phone or email, source and date, so a display lead can be compared with whatever came before it.
5. Normalize before matching. Phone numbers in one consistent international format, emails lowercased and trimmed; the reasons are in phone normalization.

Then match closed sales to the leads and credit each sale to the buyer's earliest lead. A display-tagged lead whose phone number was already in the file from a search call two weeks earlier is search's customer, and display touched them on the way. A display lead that is the buyer's first record is a sale display started. Those first-recorded matches are the floor described above, and they are the figure to put in the channel table.

CloseRev does that match from two CSV exports, your closed sales and your leads. It matches on normalized phone and email and, by default, credits each sale to the earliest matched lead, with last touch as an option. Only high-confidence matches count automatically; medium- and low-confidence ones stay reviewable, and sales with no matching lead stay in Direct / Unknown. It reads no impressions, view-throughs or DSP reports, and it does not run geo tests.

The roofer's month, counted four ways

Put the same month of prospecting display through each measure and the answers range from impressive to inconclusive. The differences are not rounding: they come from what each measure is allowed to count. The figures below are illustrative assumptions built on the roofer's report.

The lead file holds 31 leads tagged as prospecting display: 18 forms carrying the UTM source and 13 calls to the landing-page or banner numbers. Matched against the sales file over the next 90 days, 5 of those people bought. Two had called from a search ad before the display lead appeared, so 3 sales have display as the earliest lead. The average roof is $11,000. Some agencies also report "modeled revenue": conversions multiplied by an assumed close rate and deal size.

The arithmetic: 61 × 25% × $11,000 = $167,750, and $167,750 ÷ $9,000 = 18.6. Then 5 × $11,000 = $55,000, or 6.1 per dollar, and 3 × $11,000 = $33,000, or 3.7. Revenue per dollar is not profit; a roof carries materials, crew and overhead. And at 35 sales per group a month, two months of testing is about 70 per group, below even the 30% row of the size table, so the holdout's honest result is a range.

The modeled figure is the one to refuse: 49 of its 61 conversions are the view-throughs the coincidence table generates from nothing, and multiplying them by a close rate turns an exposure count into invented revenue. The earliest-lead figure is the one to manage by, with the caveat written next to it that display's no-click effect is missing. The holdout is the one that settles the question once volume allows.

A display report that multiplies conversions by a close rate is not measuring display. It is measuring how many of your customers were in a large audience, which, for a campaign built to reach half your market, is about half of them.

What to report for programmatic display

Report prospecting display on matched sales from leads it recorded first, beside its cost, in the same table as every other channel. Keep platform conversions in a separate, labeled line with click and view-through split and the window stated. Add placement quality, and a holdout result whenever one exists, as its own line rather than a multiplier.

• Matched sales and revenue from first-recorded display leads, with cost per sale. The floor, and labeled as one.
• Display leads that were already known, and which channel found them first. Credit for those belongs, at least in part, to that channel.
• Platform conversions, labeled as such: click-through and view-through separately, window stated, never added to revenue.
• Placement quality: share of spend on the top domains, spend on sites and apps with no leads, and the list of placements that produced a sale.
• Lift, if tested, with dates, regions and the confidence interval. See unattributed revenue for why sales display may have influenced but cannot be traced stay in their own bucket rather than being spread across channels.

If that report shows prospecting display producing a few first-recorded sales and a lot of coincidental conversions, the decision is not necessarily to stop. It is to cut the long tail of placements, shorten the view window in the report, track the calls properly for a quarter, and size a test before any increase. If a holdout later finds a lift that covers the cost, that is the number to defend the budget with, and it will be smaller than the dashboard's.

Programmatic display can sell, mostly to people who never click, and it also takes credit for everyone it happened to reach. Count the sales its leads closed, buy placements that produce them, and let a geo test, not a view-through column, decide whether it earns more budget.