Operations
Part of Display advertising: signal versus noise
Improving display advertising starts with tracing supply, not spend
How to improve display advertising: trace supply, placements, combinations, rendering, viewability, outcomes, fees, customer effects, and experiments in order.
What to take away
- Fix the earliest broken handoff before changing bids.
- Trace representative impressions from contract and creative through outcome and invoice.
- Use randomized lift only when its eligibility and design match the decision.
How to improve display advertising begins with a reconstruction, not an optimization score. Work the trace in this order.
- Pick a completed month. Pull the placement report from your ad server or DSP, grouped by site, app, seller, and deal.
- Sample five placementshighest spend, highest value, worst quality, unknown supply, and any flagged incident.
- Open the supply path report. Record the seller ID, reseller chain, intermediary domain, and fee.
- Check ads.txt or app-ads.txt and sellers.json. A missing or mismatched record is a defect.
- Match the creative version served with the page URL and app bundle ID, using the ad server log and the verification tag.
- Confirm the events firedimpression, viewable impression, click, and conversion. Note the measurable rate for the placement.
- Wait for the outcome window to mature. Pull refunds, cancellations, and lead quality from the CRM.
- Compare the invoice line item and fee type with the fee recorded in the supply path report.
Worked example, illustrative figures.
| Stage | Record | Source |
|---|---|---|
| Supply | Seller 4471, reseller Alpha, no ads.txt entry | DSP supply path report, ads.txt |
| Creative | Version B served, version A crop rendered | Ad server log, page screenshot |
| Page | Article page, three slots, one above the fold | Verification tag |
| Event | Impression fired, viewable impression did not | Verification report |
| Outcome | 40 clicks, 1 qualified lead after 14 days | CRM |
| Fee | Invoice 22 percent tech fee, 12 percent reseller fee; supply path showed 8 percent | Invoice, supply path report |
| Invoice | 1,180 dollars charged, working media ratio 66 percent | Reconciliation sheet |
Rank defects by position in the chain. An unknown seller or a missing ads.txt entry affects every impression on that path, so it outranks a weak creative crop. Fix the earliest broken handoff first.
Track three numbers per placement: measurable rate, viewable rate against the MRC standard of 50 percent of pixels for one continuous second, and working media ratio. Open-marketplace fees commonly fall in the 15 to 30 percent range of spend, and the contract sets the actual figure.
Use lift for the causal question
Google's current instructions for Conversion Lift in Display and Video 360 describe randomized treatment and control groups plus product eligibility and reporting conditions. It is vendor documentation. Check interference, identity, power, outcome quality, and current availability before relying on the study.
Match the experiment to the decision
NIST's explanation of design of experiments lists four purposes: comparative, screening, modeling, and optimization. Display teams should state which applies and pick a valid media design instead of treating a generic platform test as a universal answer.
Match Experiment to Decision
What is the experiment purpose?
Comparative -> compare options
Screening -> filter many factors
A documented design of experiments purpose also supports ad creative testing when the decision is which creative to keep.
| Stage | Diagnostic | Possible repair |
|---|---|---|
| Supply | Unknown seller, reseller, fee, or property | Direct path, deal change, exclusion, or disclosure |
| Placement | Unsuitable page, app, position, or refresh | Allow list, block list, verification, or publisher action |
| Creative | Broken crop, claim, disclosure, or load | Rebuild, restrict, preview, or pause |
| Measurement | Low measurable rate, inconsistent counts, or duplicates | Instrumentation, vendor, reconciliation, or definition |
| Outcome | Weak qualification, returns, or no incremental effect | Offer, audience, page, suppression, or budget |
| Economics | Hidden fee or falling marginal value | Contract, path, cap, reallocation, or stop |
- Freeze the decision, unit, primary outcome, guardrails, and stopping rule
- Keep placement, creative, audience, bid, and page changes distinguishable
- Allow downstream outcomes and refunds to mature
- Report counts, uncertainty, implementation, and practical effect
- Record spillover, outages, exclusions, and deviations
- Scale in stages with a reversible configuration
Optimization can improve the supplied signal while harming the business. Lower cost per conversion may come from existing customers; higher click rate may come from misleading creative; better viewability may concentrate costly inventory without incremental value. Pair each target with a quality or customer guardrail.
Close the review with an owned action. The team may remove a seller, repair a feed, redesign a distinct concept, change the primary event, negotiate fees, run a holdout, or stop. A report that cannot change a decision is not yet an improvement process.
Run a controlled handoff
The GOV.UK technology selection guidance recommends adaptable choices, data control, security review, and ownership-cost analysis. Use those public-service questions when assessing how to improve display advertising; they are not product endorsements. The same ownership and cost questions apply to media buying, where contract terms decide what the team can change later.
The CISA software acquisition fact sheet covers development practice, supply-chain exposure, deployment, and vulnerability management.
Assign each step to a named role and an observable finish condition. The analyst should be able to reproduce the result, while the decision owner can explain the action and stop rule.
Include one broken-data case and one revoked-access case, then record the repair, time required, and any vendor help. The written process should reflect normal operation, not a prepared demonstration.
Keep the evidence record beside the decision and set the next review date.
Common questions
What should be improved first?
Correct unsafe placements, false claims, broken destinations, corrupted data, or unauthorized access before efficiency.
Can view-through conversions guide optimization?
They can support investigation under a stated model, but they do not establish incremental causation.
When should spend increase?
Increase only when mature marginal evidence, capacity, customer effects, risk, and cash support the next step.
