# Ad experiments: A/B test banner ads on your own site
> Rotate banner ads on your own site, count viewable impressions and clicks without cookies, and test creatives to find the one that gets the most clicks.
Ad experiments are for the ads you show on your own site: a house ad for
your own product, a sponsor's banner on a directory, a call to action in
the sidebar of your docs. Privatus Analytics shows your creatives in a
slot, counts how often each one is shown, seen and clicked, and tells you
which one gets the most clicks.
It works without cookies, without anything stored in the visitor's
browser and without a visitor id, like the rest of Privatus Analytics.
## How it works, in one minute
1. You create a **placement**: a slot of one size, such as a 300 by 250
sidebar or a 728 by 90 leaderboard.
2. You upload its **creatives**: images of that shape, each with the
address a click goes to.
3. You paste one tag where the slot should appear on your site.
4. On every view the loader picks a creative by its weight and shows it.
It counts the impression, a viewable impression once the creative has
been seen, and the click.
5. To find the best creative, you start an **ad experiment** on the
placement. When the result is clear enough it ends by itself, and the
winner can be shown to everyone.
## Plans
| | Free | Pro | Business and Enterprise |
|---|---|---|---|
| Placements and creatives | Yes | Yes | Yes |
| Rotation by weight | Yes | Yes | Yes |
| Impressions, viewable impressions, clicks and click rate | Yes | Yes | Yes |
| Ad experiments at a time, per workspace (shared with page experiments) | 1 | Unlimited | Unlimited |
| Creatives per experiment | 2 | Up to 10 | Up to 10 |
| Steer to the leader (automatic shares) | No | Yes | Yes |
| Schedule a creative, and show it on some devices or pages only | No | Yes | Yes |
| Count likely humans only | No | Yes | Yes |
| Sponsor reports | No | Yes | Yes |
Every impression is an event and counts toward your
[monthly events](/docs/billing/usage). Viewable impressions
and clicks do not.
## Add a placement
Open **Experiments**, then **Ad placements**, and press **New
placement**. Give it:
- **A name**, for you, such as "Profile sidebar".
- **A key**, which the tag on your site carries, such as
`profile-sidebar`. Lowercase letters, numbers and dashes. It is made
from the name when you leave it empty.
- **A size.** Pick a common one or type your own:
| Size | Pixels |
|---|---|
| Medium rectangle | 300 by 250 |
| Leaderboard | 728 by 90 |
| Billboard | 970 by 250 |
| Half page | 300 by 600 |
| Wide skyscraper | 160 by 600 |
| Large rectangle | 336 by 280 |
| Mobile banner | 320 by 50 |
| Large mobile banner | 320 by 100 |
| Square | 250 by 250 |
- **A label**, if you like, shown above the ad, such as "Sponsored".
Saying that something is an ad is good practice, and the law in some
places.
One size a placement. For a leaderboard on desktops and a banner on
phones, make two placements and show each with your own CSS. A slot your
page hides is never counted.
## Creatives
On the placement's page, press **Add creative**.
- **Image.** A JPEG, PNG, WebP or GIF up to 2 MB, of the placement's
shape and at least its size. Twice the size looks sharpest on screens
with dense pixels. An animated GIF stays animated. The image is turned
into WebP at the placement's size and at twice that, with its metadata
removed, and served from our media host.
- **Link.** An https address, or a page of your own site such as
`/pricing`. An outside address opens in a new tab.
- **Campaign parameters.** On by default: `utm_source` is your site's
domain, `utm_medium` is `house_ad`, `utm_campaign` is the placement's
key and `utm_content` is the creative's name. For a page of a site you
measure, the clicks then show up in
[Campaigns](/docs/features/campaigns) and in your goals. A parameter
the link already has is never replaced.
- **Alternative text**, for people who use a screen reader.
- **Weight.** Outside an experiment, creatives rotate by weight. A
creative with 70 is shown about twice as often as one with 35.
- **In the rotation.** Switch a creative off to keep it and its numbers
without showing it.
- **When and where** (Pro and up): a start and an end, devices and pages
(one a line, with one `*` for many, such as `/tools/*`). These are
checked in the visitor's browser on every view, so a start or an end is
never late.
The image is checked when you upload it, so a file of the wrong shape is
refused on the form. It is ready a few seconds later.
A creative that an experiment compares can't be deleted, because the
results need it. Switch it off instead.
## Put the tag on your site
The placement's page shows the tag. It has two parts:
```html
```
- The **script** goes once on a page, however many placements it has. It
is about 2 KB compressed.
- The **slot** goes where the ad should appear. Its size is set before
the script runs, so the page does not move when the creative arrives.
When there is nothing to show, the slot is hidden.
It works with pages that change without a reload (Turbo). The slot
config is kept at the edge for up to an hour and refreshed within
seconds of any change you make.
## What is counted
- **Impression**: the creative was shown in a slot the page displays.
- **Viewable impression**: at least half of the creative was on screen
for one unbroken second, in a tab that was in view. For a creative of
242,500 pixels or more, 30% is enough. This is the IAB and MRC rule.
- **Click**: the creative was clicked. A second click on the same
creative from the same visitor address within ten minutes is not
counted again.
- **Click rate**: clicks per viewable impression. It is the number
experiments compare.
For each of them Privatus Analytics keeps the placement, the creative,
the experiment if there is one, the time, the device, the country and
the page (its host and path, with your path masks). Nothing about the
visitor is kept.
## Bots
Every plan leaves out what the
[basic bot filter](/docs/features/bot-filtering) catches: known bots,
headless browsers and visits from data centers. They are never counted
and never billed.
On Pro and up, every hit also gets a
[bot score](/docs/features/bot-detection#what-we-store), and its band
(likely human, unsure or likely bot) is kept with it. Tick **Likely
humans only** on a placement or an experiment to count only what the
score calls likely human.
## Test creatives against each other
On the placement's page press **Test these creatives**, or open
**Experiments** and press **Test ads**.
1. Pick two or more creatives of the placement whose images are ready.
The first is the original.
2. Give each a share, or leave the shares empty for an even split.
3. Choose when it ends: the shortest and longest duration, the viewable
impressions each creative needs (1,000 by default) and the
confidence.
4. On Pro and up, limit the test to some devices or pages. Elsewhere the
placement keeps its usual rotation.
5. Press **Start**.
The results show each creative's impressions, viewable impressions,
clicks and click rate, its uplift against the original and the chance
that it really is better. A winner is called the same way as in a
[page experiment](/docs/features/experiments): once the shortest
duration has passed, every creative has its viewable impressions and one
beats the original with the confidence you chose.
When a creative wins and **Send everyone to the winner** is on, the
placement shows the winner on every view until you turn it off. Then it
rotates its creatives again.
A placement runs one experiment at a time. Ad experiments count toward
the same limit as page experiments.
### Steer to the leader
On Pro and up, choose **Steer to the leader**. This makes the experiment
a multi-armed bandit: once a day, more views go to the creative most
likely to have the best click rate, and every creative keeps at least
10%, so one that started badly can still recover. Nothing moves until
every creative has 300 viewable impressions. The method is Thompson
sampling, explained under [The method](#the-method).
## The method
This is what happens behind the numbers, in brief.
**Picking a creative.** Every placement has a small list of its creatives
and their shares, which the script fetches once per page. On every view
the script draws a random number and picks a creative in proportion to
its share: with shares of 70 and 30, the first is shown about seven times
in ten. Nothing remembers which creative a visitor saw, so a visitor can
see a different one on the next page. That is why the results count
views, not people.
**Comparing the creatives.** Each creative is compared with the original
on its click rate: clicks divided by viewable impressions.
Privatus Analytics uses the same statistics as for
[page experiments](/docs/features/experiments#the-method):
- A **two-proportion z-test** gives the p-value: how likely a difference
this big would be if the two creatives were really the same.
- With more than two creatives, **Holm's method** raises the p-values, so
more comparisons don't mean more false winners.
- The **likely range** is the difference in click rate plus or minus the
margin for the confidence you chose.
- The **chance to beat the original** treats each creative's real click
rate as unknown, describes what it could be from the clicks and views
so far (a Beta distribution), and works out how often the creative
would come out ahead.
A winner is only called once the shortest duration has passed, every
creative has its viewable impressions and one beats the original at
your confidence. The check runs once an hour.
**Steering.** With "Steer to the leader", once a day the creatives' chance
of being the best is worked out by drawing 4,000 possible click rates for
each from the same Beta distributions and counting how often each one is
the highest. This is known as Thompson sampling, a kind of multi-armed
bandit. Each creative keeps 10%, and the rest is shared in proportion to
those chances. The new shares reach your site within seconds.
## Sponsor reports
On Pro and up, a placement's page can make a link to a live report for
someone with no account, such as a sponsor. It shows the impressions,
viewable impressions, clicks and click rate of the placement's creatives,
or the ones you choose, in total and day by day for the last 7, 30 or 90
days. What the bot score calls likely bots is left out.
The link can expire on a date, and deleting it stops it working at once.
If the workspace moves to a plan without sponsor reports, or the site is
locked by its plan, the link stops working but is kept: an upgrade brings
it back. The report says nothing about your visitors.
## Ad blockers
Some ad blockers hide slots and block the script. Privatus Analytics
does not try to get around them. A view that is blocked is never counted
and never billed, so the numbers you see are those of the people who saw
the ads.
## Privacy
The script sets no cookie, uses no local storage and reads no
fingerprinting API. A creative is picked at random by its weight on every
view, so nothing has to remember who saw what. The visitor's address and
browser are read in memory to leave out bots and to tell the device and
country, then dropped. See [what we store](/docs/privacy/data-inventory).
## API and MCP
Placements, creatives, ad experiments and sponsor links have the same
operations in the [API](/docs/api) and over [MCP](/docs/mcp) as in the
app: `ad_placements_list`, `ad_placements_get`, `ad_placements_create`,
`ad_placements_update`, `ad_placements_embed`, `ad_creatives_create`
(with `image_url` to have the image downloaded), `ad_creatives_update`,
`ad_report_links_create` and `experiments_create` with `kind: "ad"`.
Over MCP nothing is deleted: deleting a placement, a creative or a link
is done by a person in the app.
## Limits
- Images up to 2 MB, from 10 to 2,000 pixels a side.
- Up to 10 creatives an experiment, and 20 page patterns a creative.
- One click per creative and visitor address in ten minutes.