Programmatic advertising measurement without vanity metrics

What to take away

Why does programmatic measurement go wrong so often?

Programmatic buying splits a campaign across a demand-side platform, one or more ad servers, a verification vendor and an analytics package. Each counts something slightly different. Impressions may be counted at bid, at render, or after a fraud filter has removed invalid traffic.

A media agency in Manchester can therefore report 4.2 million impressions while the client's own analytics shows 3.1 million sessions from the same activity. Neither number is wrong. They are answers to different questions, and nobody wrote down which question was being asked.

The fix is a measurement plan agreed before the campaign goes live. It names the primary metric, the secondary metrics, the counting rules and the source of truth for each. Everything downstream, from dashboards to invoices, refers back to that document.

That plan should also state what will not be measured. Trying to prove brand effects, sales effects and direct response from a single display flight is how teams end up with twelve metrics and no decision.

Which metrics actually matter for a programmatic campaign?

Group metrics by the job they do. Delivery metrics keep the campaign running: impressions served, budget spent, pacing against plan. Quality metrics describe the exposure: viewability rate, invalid traffic rate, brand safety incidents.

Response metrics describe behaviour: click-through rate, landing page engagement, video completion rate. Outcome metrics describe the business: conversions, cost per acquisition, return on ad spend.

For a UK advertiser, the useful starting set is small. Viewability, invalid traffic, click-through rate, cost per acquisition and one incrementality read. Adding frequency, completion rate and assisted conversions gives context without overwhelming the report.

This site's breakdown of programmatic advertising key metrics in England sets out the definitions and the benchmarks worth tracking for campaigns bought in the England market.

Benchmarks should come from your own account history where possible. Industry averages vary by format, sector and season, and a display benchmark from a US retail study will mislead a UK financial services advertiser.

How should you handle viewability and invalid traffic?

Viewability is a threshold, not a score. The Media Rating Council standard for display is 50 per cent of pixels in view for one continuous second; for video it is 50 per cent for two seconds. Buying on a higher threshold costs more and delivers fewer impressions.

A practical approach is to set a floor, report against it, and treat anything above the floor as a negotiation point with publishers rather than a target in itself. Chasing 100 per cent viewability usually means buying premium placements at a premium price.

Invalid traffic is the other half of quality. General invalid traffic covers known bots and crawlers; sophisticated invalid traffic covers behaviour that mimics humans. Verification vendors filter both, but their definitions differ.

Traffic quality checks are not only a vendor job. If you run your own site and see unexpected crawler activity in logs, Google publishes guidance on verifying requests from Google crawlers and fetchers so you can separate genuine Googlebot from spoofed traffic.

The measurement plan should record which verification vendor is authoritative, which filters are applied, and whether filtered impressions are still billed. That single decision changes reported cost per acquisition by a wide margin.

How do attribution and incrementality fit together?

Attribution distributes credit across touchpoints using a rule. Last click gives everything to the final click. Linear spreads it evenly. Data-driven models weight touchpoints by their observed contribution. Each is a model, not a fact.

Incrementality asks a different question: would this sale have happened without the advertising? The cleanest answer comes from a controlled test, such as holding back a matched group of audiences or geographies and comparing outcomes.

Most UK advertisers need attribution for day-to-day optimisation and incrementality for budget decisions. Attribution tells you which placements to shift spend towards this week. Incrementality tells you whether the channel deserves the budget next quarter.

Comparing the two is informative. When a data-driven model claims a large share of credit for a channel that a holdout test shows adds little, the model is probably capturing demand that already existed. This site's guide to programmatic advertising attribution methods explains the main models and the trade-offs between them.

One caveat applies throughout. Consent and cookie restrictions mean some conversions are never observed in analytics. Models have to estimate the unobserved share, and that estimate should be documented rather than buried in a dashboard.

What does good reporting look like week to week?

A useful weekly report answers three questions: what happened, why, and what changes next. It should fit on one page before appendices. Anything longer goes unread.

The top line is spend, delivered impressions, viewability, invalid traffic rate and conversions, each with the prior week and the plan for comparison. Below that, a short commentary naming the two or three factors that moved the numbers.

Monthly reporting adds cost per acquisition by placement, creative and audience, plus the incrementality read if a test concluded. Quarterly reporting adds the channel-level view and the budget recommendation.

Definitions belong in the report footer. If viewability is measured post-filter and conversions are deduplicated over 30 days, say so. A reader who cannot reproduce your number will not trust it.

For teams running multiple campaigns across several platforms, this site's operations and delivery guide covers the workflows that keep reporting consistent when activity scales.

How do you set up analytics so the numbers reconcile?

Start with a single analytics property that receives campaign tagging from every platform. Google's documentation on the next generation of Analytics explains how properties and data streams are structured, which matters when you are deciding where campaign traffic should land.

Use one naming convention for campaigns, sources and creatives, and enforce it in the ad platform's templates rather than in a spreadsheet afterwards. Most reconciliation failures are naming failures.

Decide how conversions are counted. A purchase, a lead and a newsletter sign-up are different events with different values. If a lead generation campaign runs on LinkedIn, the platform's own lead generation ads collect submissions inside the platform, so those leads must be exported and matched against your CRM rather than counted twice.

Finally, agree who owns each number. Media owns delivery and viewability. Analytics owns sessions and conversions. Finance owns spend. When a number is disputed, the owner adjudicates.

What standards and skills should the team rely on?

Measurement practice in programmatic advertising is partly codified and partly judgement. The codified part comes from industry bodies. The IAB's Measurement Centre publishes the standards and methodologies that underpin viewability, invalid traffic and cross-media measurement definitions.

Adopting those definitions makes your reporting comparable with partners and clients who use the same language. It also gives you a defensible position when a vendor's numbers disagree with yours.

The judgement part is a skills question. Someone on the team needs to understand sampling, confidence intervals, attribution modelling and the limits of observational data. The Chartered Institute of Marketing runs data and insights training courses that cover audience analytics and measurement fundamentals.

A reasonable target for a mid-sized UK agency is one person who can read a test result critically and one person who can build the dashboard. They do not have to be the same person, but both roles need cover.

Worked example: a UK homeware brand

Consider an illustrative example. A homeware brand spends £60,000 a month on programmatic display and video across two platforms. The media agency reports 12 million impressions, a 0.09 per cent click-through rate and 1,800 conversions at £33 cost per acquisition.

The client's analytics shows 1,300 conversions from the same activity. The gap of 500 comes from view-through conversions counted by the platforms and from deduplication differences. Neither figure is fraudulent; they measure different things.

The team agrees a primary metric: incremental conversions per pound, tested quarterly with a geo holdout. Day-to-day optimisation uses platform-reported conversions, discounted by an agreed factor of 0.72 to bring them in line with analytics. That factor is reviewed each quarter.

Viewability is set at a 60 per cent floor, measured post-filter. Invalid traffic runs at 1.4 per cent of impressions, within the agreed tolerance. Reporting shows spend, impressions, viewability, invalid traffic, analytics conversions and discounted conversions on one page.

After two quarters, the holdout test suggests roughly 40 per cent of the conversions attributed to display would have occurred without it. The brand shifts £15,000 a month from display into retail media and paid search, and holds the rest.

Nothing in that example required new technology. It required agreed definitions, one reconciliation factor and a test that could contradict the dashboard.

Where do teams most often slip up?

Measurement programmes fail in predictable ways. Reporting platform conversions as business outcomes is the most common. Treating viewability as a target rather than a threshold is the second.

A third is running attribution and incrementality as separate exercises owned by different teams, so the two never meet. A fourth is changing definitions mid-quarter and comparing the results anyway.

This site's review of programmatic advertising measurement mistakes covers the recurring errors in more detail, including the ones that survive audits because nobody questions the dashboard.

A fifth is ignoring consent effects. If a meaningful share of UK users decline non-essential cookies, modelled conversions carry more weight, and the model's assumptions deserve scrutiny.

How should you choose benchmarks to compare against?

Benchmarks serve two purposes: setting expectations before launch and diagnosing after. For the first, use your own historical data or published research with a clear methodology and date.

For the second, compare like with like. A video completion rate benchmark from a connected television campaign will not help a mobile display campaign. Seasonality matters too; a November retail benchmark will flatter a February campaign.

Where no reliable benchmark exists, state the assumption explicitly and label it as an assumption. A labelled estimate is more useful than a false precision.

This site's guide to programmatic advertising benchmark research covers how to read published benchmarks and where the common sampling problems sit.

Common questions

How often should measurement definitions be reviewed?

At least quarterly, and whenever a platform changes its counting rules. Platform updates to viewability or conversion windows can shift reported numbers without any change in campaign performance, so the review should check both your definitions and theirs.

Can a small team run incrementality tests?

Yes, if the test is scoped to one channel or one geography at a time. A holdout of 10 to 15 per cent of a target region is often enough to read a directional result, provided the test runs long enough to cover a full purchase cycle.

What is the minimum viable measurement stack?

An ad server or platform report, one verification vendor, one analytics property with consistent campaign tagging, and a written plan naming the primary metric. Everything else is optional until the basics reconcile.

Should view-through conversions be counted?

Count them, but report them separately from click-through conversions and apply a consistent discount when they feed optimisation. Blending the two hides the difference between a click and an exposure.