> ## Documentation Index > Fetch the complete documentation index at: https://docs.instantcampaign.ai/llms.txt > Use this file to discover all available pages before exploring further. # Synthetic panel (screening) > Screen subject-line variants with AI personas built from your audience's statistics before you run a real A/B test. The synthetic panel is a **screening aid**. AI personas, built from statistics about your audience, rate your subject-line variants before you send anything. Use it to drop weak variants. It is not a prediction, and it never picks a winner. A real [A/B test](/guides/audience-and-campaigns/ab-testing-and-scheduling) decides. In published studies, synthetic panels agreed with real respondents about 79% of the time, and they exaggerated differences between segments 2–4×. Treat the scores as a hint about which variants are weak, not as a forecast of your open rate. ## Where to find it * **Plan page**: the **Synthetic panel (screening)** card at the bottom of a campaign plan. Email 1's subject is variant V1. Add up to three alternative subjects with **Add an alternative subject**. * **Send wizard**: on the **A/B test** step, when the test type is **Subject line**, the card screens the variations you typed. Click **Run panel**. A run takes a few seconds. ## What the personas are built from The panel only uses **aggregate statistics** about the audience you picked. It never uses a contact's details. * how many contacts are in each engagement group: opened or clicked in the last 30 days, 31–90 days ago, more than 90 days ago, or never * the countries in each engagement group * the audience's most common languages and time zone * which mail apps your workspace's opens come from * how many contacts completed a goal or made a purchase recently * your recent average open and click rates, and what your real A/B tests found * for up to three dropdown or yes/no custom fields, the share of each option among contacts that have a value (for example **Plan tier: free 70%, pro 30%**) From this the panel makes between one and eight personas, for example **Highly engaged, United Kingdom** or **Dormant 90+ days, France**. Each persona stands for a share of the audience. A country that is less than 3% of the audience is grouped with other countries. If your audience has fewer distinct groups, you get fewer personas. The panel never adds made-up personas. Only dropdown and yes/no custom fields are used, and only as shares of their options. Text, number, date and JSON fields are never used, and neither is any field marked [sensitive](/guides/audience-and-campaigns/custom-fields#sensitive-fields). ## Reading the results For each variant you see: | Item | Meaning | | - | - | | **Screening score** (0–100) | Open and click intent, lowered by unsubscribe risk, weighted by each persona's share of the audience | | **Spread** | How much the personas disagree. High spread means the variant works for some groups and not others | | **Persona range** | The lowest and highest persona score | | **Open / click / unsubscribe risk** | The weighted intents behind the score | | **Reasons** | One sentence each from the personas that stand for the most contacts | If the gap between the top variants is smaller than their spread, the card says so. Treat those variants as equal. ## Then run a real test Keep the strongest variants and click **Create real A/B test**. The send wizard opens on the A/B step. ## Cost and limits * One run uses **one AI text generation** from your plan's allowance. Running the same personas on the same variants again within 7 days shows the saved result and costs nothing. * Up to four variants and eight personas per run. * You need permission to edit campaigns to run the panel. Anyone who can view the campaign can see the last run. * Spark and the MCP server can read the last run with `get_synthetic_panel_results`. They cannot start a run. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.