Make and Test AI Ad Variants With Atria and Creatify

The short answer: if your Meta account needs six fresh ads a month to replace tired winners and your designer can make two, an agent can draft the other four, on one condition that never bends: no claim reaches an ad until a person has matched it to a source. On monthly billing the tools come to $174 a month, Atria Core at $159 (shown as a promotion against a listed $219) and Creatify Standard at $15 (Atria pricing, Creatify pricing, checked on 26 October 2026), and Meta's own ads MCP server creates every ad paused until you set it live (Meta Business Help). For example, if one variant takes 3 hours from brief to render by hand and checking an agent's draft takes 45 minutes, four variants give back 4 x 135 = 540 minutes a month, or 9 hours.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

Generative tools are good at the confident sentence. In an ad that is the danger: "dermatologist approved", written because it sounds right and rendered before anyone reads it. This build checks claims where a check costs least, before the render, and looks again after it.

Before you build:

1. The bill and the allowances. Atria Core includes 600 MCP calls a month, and each call beyond that costs 1 AI credit (Atria MCP article). Creatify's free plan turns 10 credits into up to 2 video ads, so 10 / 2 = 5 credits a video, and at that rate Standard's 55 credits make 55 / 5 = 11; the app shows the real cost before each generation (Creatify pricing).
2. A proof file you already half have. One sheet of every claim the brand may make, each with its source: the product page, a test report, the review export, the price list. The agent may quote only from this sheet, by line number.
3. The accounts. Access to the Meta ad account for whoever connects Meta's ads MCP server at https://mcp.facebook.com/ads (setup), plus Atria and Creatify logins.
4. The number that decides it. Cost per purchase, read from your shop orders, for the current champion ad over the last 30 days. Meta's A/B test will judge on cost per result (Meta A/B testing); your shop is the second opinion.

The free version, enough for two or three variants

Research what competitors run in Meta's Ad Library, which is free, and your own winners in Ads Manager. Paste the proof file and the winning ad's script into a free chat assistant and ask for six new openings, each citing a proof line. Produce the two strongest with your designer or a phone. Creatify's free plan gives 10 credits a month with a watermark on exports (Creatify pricing), enough to see a format before paying for it.

Test inside Meta, for nothing. An A/B test splits the audience so nobody sees both versions, and Meta advises against informal testing by switching ad sets on and off (About A/B testing). Its guidance asks for one variable, at least 7 days and at most 30 (Meta test guidance). The claims check is a spreadsheet with three columns: line, source, approved by.

What the paid build adds is grading across every ad you run, generation in volume, and an agent that writes the test for you. It does not add judgement.

The build: Atria reads, Creatify renders, Meta holds the test

1. Trigger: Monday, the ads that nearly win. In Claude, add a custom connector named Atria with https://api.tryatria.com/mcp and sign in (Atria MCP article). Ask for the ads in Radar's High iteration potential tab and the grade holding each one back. Radar scores each ad of the past 30 days for hook, retention, click-through and conversion, measured against the goal you set (Radar). The agent reads; Atria asks for confirmation before anything that would change or delete data.
2. The agent drafts against the proof file. Ask it to keep the body of the chosen ad, write six openings for the weak hook, cite a proof line for every claim, and pull three reference ads with a similar opening from Atria's library of over 25 million ads, which the same connector searches (Atria MCP article). The output is a table: opening, proof line, reference ad.
3. Human check one, before a credit is spent. The person who owns claims strikes any opening without a source line, any line where an actor speaks as a customer ("I lost four kilos"), and any price or discount that differs from the landing page, which TikTok's policy does not allow (TikTok misleading content). Initials go in the sheet, beside each line that survives.
4. The agent renders only what was approved. Add Creatify's MCP server, https://api.creatify.ai/mcp, as a second connector. Its Hook multiplier preset holds the body fixed and writes six alternative openings, Headline multiplier does the same for a static visual, and each turn reports the credits it used (Creatify MCP). For creator-style talking ads, Arcads runs an MCP server at https://mcp.arcads.ai, included on every plan, and its skills plugin adds a /clone-hook command (Arcads MCP, Arcads pricing).
5. Human check two: the renders and the labels. Watch each render with sound, read the captions, and compare the product on screen with the product you ship. Then label: the rules are in the next section.
6. The agent builds the test, paused. With Meta's ads MCP server connected, ask for a new ad set that mirrors the champion's settings with the new ad as the only difference, and an A/B test between the two on cost per purchase. The server's ads_create_ad_set and ads_create_ad tools create both paused, and ads_experiment_abtest_create_test defines the test cells (ad tools, test tools).
7. A person sets it live. Preview every placement in Ads Manager, give both cells the same budget, as Meta recommends, and activate.
8. The agent reads, the person decides. The agent fetches the result with ads_experiment_abtest_get_test and writes the note. Meta picks a winner by simulating outcomes tens of thousands of times and reports it with a confidence percentage (how winners are determined).

The per-tool walkthroughs sit in how to use Atria and how to use Creatify, and the products in the Atria and Creatify entries. If competitor research is the gap rather than your own account, Foreplay connects to Claude the same way (Foreplay MCP).

Labels for AI-made ads, by platform and by law

• Meta labels ad images created or edited with third-party AI tools with AI info on its own, and advertisers running ads about social issues, elections or politics must disclose AI-made images, video or audio themselves (Meta Help Centre).
• TikTok requires the toggle "This ad contains AI-generated content" for media that is fully AI-generated or significantly modified; the toggle resets when a campaign is duplicated (TikTok disclaimers), and undisclosed AI content gets the ad rejected or restricted (TikTok policy).
• Google documents its "Altered or synthetic content" checkbox in the Political content policy, for election ads that realistically depict people or events (Google Ads policy).
• The EU AI Act requires whoever deploys AI to make a deep fake to disclose it, at the latest at first exposure (Article 50(4) and (5), Article 50). Article 3 defines a deep fake as AI content resembling real people, objects, places or events that a viewer would wrongly take for authentic (Article 3). The Act applies from 2 August 2026, and generators already on the market before then have until 2 December 2026 to mark their output as AI-made (Article 113, Article 111).

The monthly bill, side by side

Prices checked on 26 October 2026 on the Atria, Creatify, Foreplay and Claude pricing pages and the Arcads pricing article.

How to know it worked

• The measure. Cost per purchase per variant from your shop, with each ad carrying its own utm_content, set beside Meta's cost per result for the same days.
• The baseline. The champion's 30-day cost per purchase, written down before round one.
• The holdout. Meta lets you add a holdout to an A/B test and compare cost per incremental purchase (how winners are determined); use it on the round where a lot of budget follows the verdict.
• The process number. The share of drafted openings struck at check one. If it passes half, the proof file is thin, not the agent.

Where the credits go

A generator earns on every render an agent requests, approved or not, and Creatify reports the credits on each turn (Creatify MCP). That is why the claims check sits before the Hook multiplier runs: a line struck in a spreadsheet costs nothing, while a line struck after its render has already been paid for in credits.

FAQ

Will Meta label ads I made with Creatify or Arcads?

For ad images, yes: Meta applies AI info to images created or edited with third-party AI tools, shown under About this ad and sometimes beside Sponsored (Meta Help Centre). Political and social issue advertisers still self-disclose.

Can the agent put an ad live by itself?

Not in this build. Meta's server creates ads paused and asks for your authorisation through the agent for any action (Meta Business Help), and where rules are available one can deny delivery status changes outright (rules). A person activates in Ads Manager.

Is an AI actor a deep fake under the EU AI Act?

It is when the video resembles a real person and would falsely appear authentic (Article 3); then Article 50(4) requires the deployer, here the advertiser, to disclose it. Arcads' own guidance says AI actors are not real customers and must not be presented as verified buyers (Arcads testimonial guide).