> For the complete documentation index, see [llms.txt](https://docs.elevateab.com/elevate-helpcenter/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.elevateab.com/elevate-helpcenter/experiment-setup/traffic-isolation.md).

# Traffic Isolation

***

Traffic isolation sets what percentage of eligible visitors enter an experiment. The rest continue to your normal store experience or enter other isolated experiments.

It does two jobs at once. It caps your exposure to a risky change, and it makes concurrent experiments **mutually exclusive**, so no visitor is ever counted in two of them.

### Does Elevate A/B Testing support mutually exclusive experiments?

Yes. Traffic isolation is that feature. If you have used another experimentation platform, you have met this control under a different name.

| Term you may know              | Where it comes from                                                               |
| ------------------------------ | --------------------------------------------------------------------------------- |
| Mutually exclusive experiments | The statistical description, and the common name across experimentation platforms |
| Exclusion groups               | Optimizely                                                                        |
| Layers, or universes           | Google's overlapping experiment infrastructure, and platforms modelled on it      |
| Traffic isolation              | Elevate A/B Testing                                                               |

Mutual exclusivity is the guarantee: a visitor placed in one isolated experiment cannot enter another. Traffic isolation is how Elevate A/B Testing delivers it, by reserving each experiment a named slice of the traffic pool and drawing each visitor into at most one slice.

The reserved slice is also a cap, which is the second job. A platform that only offers mutual exclusivity splits your traffic between concurrent experiments. Traffic isolation lets you reserve 10 percent for one experiment and leave the other 90 percent entirely untouched, which is what you want when the change is risky rather than when it merely conflicts.

### Why mutual exclusivity matters

Two experiments running on the same page, both entering the same visitors, produce results neither of you can attribute.

Say you are testing a new hero image and, at the same time, a new add to cart button. A visitor can see the new hero with the old button, the old hero with the new button, both, or neither. Each experiment reports its own variation against its own control, but every one of those numbers now contains the other experiment's effect as noise. If the two changes interact, and layout changes frequently do, both results are wrong in ways the significance calculation cannot see or warn you about.

Isolating each experiment to its own slice removes the overlap. Experiment A measures only visitors who never saw Experiment B, and the reverse. Each result is clean, at the cost of each experiment running on less traffic.

> **Note:** Experiments running on unrelated pages, or on non-overlapping audiences, do not need isolation. A homepage experiment and a checkout experiment cannot contaminate each other's page level metrics in the way two experiments on the same template can. Isolate when the surfaces overlap, not by default.

### Isolation, allocation, and targeting are three different controls

They are configured separately and they answer different questions. Confusing them is the most common setup error on this page.

| Control                | Question                                                            | Example                                  |
| ---------------------- | ------------------------------------------------------------------- | ---------------------------------------- |
| **Audience targeting** | Which visitors are eligible?                                        | Mobile visitors in the US                |
| **Traffic isolation**  | What share of eligible visitors enter this experiment, exclusively? | 10 percent of them                       |
| **Traffic allocation** | How are entrants split across control and variations?               | 50 percent control, 50 percent variation |

They apply in that order. With the settings above, a US mobile visitor has a 10 percent chance of entering the experiment, and if they enter, a 50 percent chance of seeing the variation. Everyone else sees your store unchanged.

### What isolation changes

By default, every eligible visitor enters the experiment. If they land on a page with an active experiment and match your audience rules, they are in, including when another experiment is also running there.

With isolation on, only your chosen percentage enters, and that percentage is reserved.

| Setting                 | Result                                                                                                                                     |
| ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
| Isolation off           | All eligible visitors are split across control and variations, and may also be in other non-isolated experiments                           |
| Isolation at 10 percent | 10 percent of eligible visitors enter this experiment and no other isolated experiment. The other 90 percent always see the original page. |

### When to use it

* **Making concurrent experiments mutually exclusive.** Two experiments on the same page, or on the same template, need isolation to produce attributable results.
* **Limiting risk on high-traffic pages.** A bold change on a homepage taking 50,000 visitors a day is a different proposition at 10 percent than at 100 percent.
* **Gradual rollout.** Start small, confirm nothing is broken, raise the percentage.
* **Protecting revenue in peak periods.** During BFCM or a major campaign, capping exposure caps the downside.

### How it works

1. A visitor arrives on a page with one or more active experiments.
2. Audience targeting decides which experiments they are eligible for.
3. Among eligible experiments with isolation enabled, Elevate A/B Testing decides which one, if any, the visitor enters, based on the isolation percentages.
4. If the visitor enters an isolated experiment, they are assigned a variation by that experiment's traffic allocation, and they are excluded from every other isolated experiment.
5. If they fall outside every isolated slice, they enter any non-isolated experiments running on that page.
6. On return visits, the visitor stays in the experiment and variation they were first assigned.

Isolation decides whether a visitor is in the experiment. It never decides which variation they see.

Step 5 is the one that surprises people. Isolation makes isolated experiments exclusive of one another. It does not shield an isolated experiment from a non-isolated one. If you are isolating for mutual exclusivity, isolate every experiment on that surface, not just the one you are worried about.

### Enable traffic isolation

Traffic isolation is set during experiment creation, below the traffic allocation section.

1. Expand **Traffic Isolation**.
2. Toggle it **on**.
3. Set the percentage using the slider.

The slider shows how the full traffic pool is divided:

| Segment                | What it represents                                   |
| ---------------------- | ---------------------------------------------------- |
| **Current Experiment** | The share you are allocating to this experiment      |
| **Other Experiments**  | Share already reserved by other isolated experiments |
| **Remaining**          | Share entering no isolated experiment                |

The default is **10 percent**, which is a reasonable starting point for a single experiment. When you are isolating several experiments to keep them mutually exclusive, size each slice from the traffic that experiment needs rather than accepting the default.

### Running several isolated experiments at once

Isolated experiments share one 100 percent pool. If Experiment A takes 10 percent and Experiment B takes 15 percent, 75 percent of traffic remains for non-isolated experiments and normal browsing.

Each visitor enters at most one isolated experiment. Assignment is random: a 10 percent chance of A, a 15 percent chance of B, a 75 percent chance of neither.

This is the mutual exclusivity guarantee in practice. Experiment A's results contain no visitor who saw Experiment B.

> **Important:** Total allocation across isolated experiments cannot exceed 100 percent. The slider caps your available range based on what other experiments have already reserved.

#### Splitting the pool between experiments

Mutual exclusivity is not free. Three experiments sharing the pool each get roughly a third of the traffic they would have had alone, and each takes roughly three times as long to reach an answer.

That is the trade you are making, and it is usually the right one. Three clean results in nine weeks beat three contaminated results in three. But it does mean the number of experiments you can run at once is set by your traffic, not by your ambition. If you cannot give each isolated experiment enough traffic to finish, run them in sequence instead.

### Why is my isolated experiment taking so long?

Because isolation reduces your sample proportionally. At 10 percent, an experiment needs roughly ten times the calendar time it would need at full traffic to reach the same confidence.

Work out the arithmetic before launching. If a page takes 2,000 visitors a day and your experiment needs 20,000 per arm, full traffic reaches that in about three weeks. At 10 percent it takes most of a year, which means the experiment will never finish and the exposure you avoided cost you the answer.

Isolation is a tool for limiting risk on high-traffic pages and for keeping concurrent experiments clean. On low-traffic pages it converts a slow experiment into an impossible one.

### Things to keep in mind

* **Set it before launching.** You can change the percentage on a running experiment, but visitors entering under two different allocation regimes complicate the result. Where you must raise it, treat the increase as a restart point.
* **Isolate every experiment on the surface, or none.** Isolating one of two overlapping experiments does not make either exclusive.
* **Not available for personalizations.** Isolation applies to experiments only. A personalization always serves the personalized experience to every matching visitor, which means a personalization can overlap an experiment.
* **Assignment persists.** Once a visitor is placed in an isolated experiment, they stay in it across sessions. They will not drift into a different experiment on a later visit.
* **Isolation does not reduce script activity.** Every visitor to the page is still evaluated. Isolation limits who is affected and measured, not who is assessed.

### Next steps

* [Audience Targeting](https://claude.ai/elevate-helpcenter/operate/experiment-setup/audience-targeting) for deciding which visitors are eligible before isolation is applied
* [Experiment Lifecycle](https://claude.ai/elevate-helpcenter/operate/experiment-setup/experiment-lifecycle) for what changes when an isolated experiment moves between states
* [Statistical Significance](https://claude.ai/elevate-helpcenter/analyze/reports/statistical-significance) for working out whether your isolated share can reach an answer in reasonable time
* [Ending an Experiment](https://claude.ai/elevate-helpcenter/operate/experiment-setup/ending-an-experiment) for concluding an experiment that will not reach significance at its current share