A/B test your ad set-up on live traffic
Experiments is the A/B testing tool in Fusion. You duplicate an ad set-up (a tag config), change one thing, and split live traffic between the versions. Fusion measures each one on experiment RPM, alongside viewability, fill, latency and bid rates, and switches on the winner in one click. No developer and no custom code are needed.
From one idea to the winning config
- 01Duplicate a config and change one thingCopy the tag config you have live and change the setting you want to test: position, sizes, SSPs, timeout, floors or refresh. Leave the copy disabled; the experiment switches it on for its share of traffic.
- 02Split the trafficGive each config a share, for example 20% each, while the other 60% keeps your current set-up. Narrow the test by country, region, device, browser, operating system, screen size, audience or URL.
- 03Schedule it and let it runSet a start and end date, or no end date. Run for at least two weeks: new units and sizes are new inventory, and buyers take time to bid on them.
- 04Compare and activate the winnerThe report ranks configs on experiment RPM. Activate the winner and Fusion switches off the others and rebuilds your tag, live within 10 minutes. Or turn on auto-activate before it starts.
Ad experiments publishers run in Fusion
Anything in a tag config or its ad stack can be tested: where ads go, which sizes and demand partners they use, and how they load.
| Test | Example | What you change | What to watch |
|---|---|---|---|
| Ad position | In-image ads "in" the image vs "under" it | Targeting and position in the config | Experiment RPM, viewability |
| Page layout | Two MPUs spaced out vs one double MPU | Unit type and placement | Experiment RPM, viewability, fill |
| Ad sizes | Adding creative sizes to a placement for more competition | Unit sizes for mobile, tablet and desktop | Experiment RPM, fill. New sizes ramp up, so allow two weeks |
| Prebid timeout | A shorter or longer bidder timeout | Timeout in a detached copy of the ad stack | Ad latency, bid rate, fill |
| SSP line-up | Removing the weakest SSPs | SSPs in a detached copy of the ad stack | Bid rate, win rate, latency. See SSP Insights |
| Bid throttling | Throttling on vs off, or different settings | Bid throttling in the ad stack | Throttle rate, bid rate, latency |
| Floor prices | AI Dynamic Floors vs your current floors | Floor model in the config | Experiment RPM, fill |
| User ID solutions | With and without an ID module | User ID modules in the ad stack | Experiment RPM, bid rate |
| Lazy loading | Loading ads closer to or further from the viewport | In-view buffer in the config | Viewability, unit latency, experiment RPM |
Six rules for an ad experiment you can trust
Change one thing
The report lists the settings that differ between configs. Keep that list to one item and you know what caused the result.
Run for two weeks or more
New units and sizes start as new inventory, and buyers ramp up their bids over time. Weekday and weekend traffic also differ.
Test on a slice of traffic
Put 20% on each config and leave the rest on your current set-up. That traffic isn't measured, so a losing test costs little.
Test where it matters
Filter to the traffic you care about, such as UK mobile visitors, rather than averaging it with everything else.
Look past RPM
Experiment RPM picks the winner, but check viewability, fill and latency too. A small gain may not be worth a slower page.
Keep testing
Leave the end date blank and move the winner to 90%, keeping 10% on the other config in case seasonality changes the answer.
What the experiment report shows
Each experiment is Scheduled, Live or Completed, with a progress bar while it runs. Each config has its own colour, and the report lists the settings that differ between them.
| Metric | What it tells you |
|---|---|
| Experiment RPM | Revenue per 1,000 times each config was served in the experiment, normalised for its share of traffic, so an 80/20 split still compares fairly. This picks the winner. |
| Gross revenue (cumulative) | Revenue added up day by day, adjusted to 100% of traffic so you can forecast. Toggle it off to see raw revenue. |
| Gross revenue (by day) | How close or far apart the configs are each day. |
| Viewability | Viewable impressions as a share of impressions. Worth checking if advertisers buy on viewability. |
| Fill | Ads loaded as a share of ad requests. Lower fill can still mean more revenue when floors rise. |
| Ad and unit latency | Median time to load the ad and the unit, in milliseconds. |
| Bid rate, win rate, throttle rate | How SSPs responded to each config: bids per request, wins per bid and requests held back by bid throttling. |
| Heavy ad intervention | The share of ads Chrome removed for using too much CPU or data. |
| Data table | The raw daily numbers, with filters, grouping and export. Save a view for other people on your team. |
Where publishers got to on Fusion
Results are for each publisher's whole Fusion set-up, where testing is one part of the work.
A/B testing ads: questions, answered
How do publishers A/B test ads?
Why not just compare this week with last week?
How long should an ad experiment run?
What is experiment RPM?
How does Fusion pick the winner?
What traffic split should I use?
Can I test on only some visitors?
Can I change an experiment once it is live?
Do experiments need a developer or changes to my site?
Does it work with my own Google Ad Manager?
Not sure what to test first?
Get a free Revenue Healthcheck: your ad set-up, header bidding and Core Web Vitals scored and benchmarked against similar publishers, with the changes worth testing.