Meta Ads AI Creative: Test Generative Variations Without Losing Brand
Meta's generative AI creative features can produce more variations than any team can write by hand. This playbook shows how to control them.

Meta's generative AI creative features can produce more variations than any team can write by hand. Text generation creates up to five variations of your primary text and headline. Image generation builds diverse versions of your product shots. Background generation, visual touch-ups, add animation, enhance media text, show summaries, and a dozen other enhancements can adjust what your audience sees. The promise is more creative in less time, with Meta delivering the variations it predicts will perform best.
The risk is that by default some enhancements are on, and Meta may adjust your text and media while maintaining what it calls the core message of your campaign. Image generation can add generated text and logo overlays. Enhance CTA can surface promotions such as discounts or free shipping. Show summaries generates a summary of selling points that appears between the footer and comments. If you are in a regulated industry or running Special Ad Category campaigns, you may not have access to all generative AI features, and certain ads may carry an AI info label.
Operators who get value from Meta Ads AI creative treat it as a controlled system, not a black box. This playbook walks through what Meta's generative AI creative features actually do, where brand control slips, how to write a creative brief and brand rulebook before you turn anything on, how to set up branding presets, how to decide which enhancements to use and which to turn off, how to review every AI variation before launch, how to design tests that actually isolate what AI contributes, how to name and log variations so you can read results by creative, and how to build a refresh cadence that adds new concepts instead of just more AI variations of the same idea.
This is not about whether AI creative works. It is about running it so you know what went live, what it said, and what you learned.
What Meta's Generative AI Creative Features Actually Do
Meta groups its enhancements under Advantage+ creative. When you turn on Advantage+ creative in Ads Manager or Meta Business Suite, images and videos are optimized to versions your audience is more likely to interact with. Some enhancements may be turned on by default, but you can turn them off at any time. Availability depends on ad format, placements, whether you use a catalog, your objective, your account, your region, and your industry. Meta is rolling some features out gradually.
Here is what the generative AI enhancements actually do, grouped by what they touch.
Text Enhancements
Text generation: Creates up to five variations of your primary text and up to five variations of your headline, based on the original inputs you provide. You need at least five words in your primary text to generate variations. The headline is optional. You see suggested variations as checkboxes during ad creation. You select the ones you want, edit them, or add your own. Meta shows variations of primary text when it predicts that can improve performance, so Meta says your description should contain only nonessential information. Text generation is available for single image, single video, and carousel ads. Text variations with and without personas are available when your primary text is in English, Portuguese, or Spanish, though not all regions and languages are supported yet.
Text improvements: Keywords and phrases from your copy are displayed or adjusted to fit on or around the creative, such as overlays, footers, or prominent headlines.
Enhance media text: Creates alternative text overlays by rephrasing your original message. Meta delivers either the original or a variation. The visuals do not change. You can review variations before publishing via previews in the Enhance media text card in the Enhancements section.
Translations: Translates text, voice, and image overlays.
Visual Enhancements
Image generation: Creates diverse versions of your original creative, including generated text and logo overlays on the images. There are two modes. Product images mode keeps a specific object, such as a product or person, and generates environments and ad styles around it. Full images mode uses your reference image to inspire variations with text overlays. Image generation is available for single image ads only. Your image should be a high-resolution JPEG or PNG, recommended at least 1000 by 1000 pixels, with a minimum of 320 by 320 pixels. Image generation is not available on Advantage+ app campaigns, Advantage+ sales campaigns, or Advantage+ catalog campaigns at this time.
Visual touch-ups: Crops and expands images, generating new pixels to fit different aspect ratios.
Background generation: Creates diverse versions of product images. This enhancement is available for Advantage+ catalog ads only.
Image touch-ups: Adapts product images to 4:5 and 9:16 aspect ratios, with auto or manual selection. This enhancement is available for Advantage+ catalog ads only.
Add animation: Adds panning, zooming, or rotating to eligible images, turning them into short videos. Text in images may also be animated. This enhancement is available for single image ads only.
Add overlays: Shows text you provide along with your creative when Meta predicts it is likely to improve performance. Uses brand preferences if branding is on. This enhancement is available for single image ads.
Video touch-ups: Crops or uses AI to expand video to 9:16 or 4:5 aspect ratios, up to the first 15 seconds for expansion.
Add video effects: Applies higher contrast and more vibrant colors. This enhancement is available for video ads.
CTA and Summary Enhancements
Enhance CTA: Can swap CTA link stickers for pre-written CTAs. You have the option to use AI to identify and show promotions with your CTA. Shop now, Learn more, and See details CTAs may get dynamically personalized information such as a percentage off or free shipping. Click Customize to see pre-written CTAs and opt in or out of promotions.
Show summaries: Generates an AI summary of selling points that appears between the footer and comments. This enhancement may be on by default for single image or video ads in Facebook Feed and Instagram Feed.
What May Be On by Default
Meta states that some enhancements may be turned on by default. Reveal details over time, show spotlights, and show summaries may be on by default depending on your ad format and placement. In the boost flow on a Facebook Page or in Meta Business Suite, Advantage+ creative is turned on by default. If you do not want Meta to automatically deliver different creative variations, you need to turn it off.
Because defaults vary, check the ad preview and review how the enhancements might adjust your ad creative before publishing. Some enhancements can only be turned off in Advanced preview and will not be visible in the main Ad creative section. To review the status of all enhancements, check both the Enhancements section and Advanced preview.

Where Brand Control Slips
Meta says the media and text you upload may be adjusted to help improve ad performance while maintaining the core message of your campaign. If branding is on, Meta tries to use your brand preferences in some enhancements. The phrase maintaining the core message is doing a lot of work. Here is where brand control can slip if you do not review and constrain what goes live.
Adjusted text and media: Text generation, text improvements, and enhance media text can rephrase your copy. Visual touch-ups and video touch-ups crop and expand images and video, generating new pixels. Add animation turns static images into short videos with panning, zooming, rotating, and animated text. The variations Meta delivers are the ones it predicts will perform best, and you will not see every variation in advance unless you review them in Advanced preview.
Generated overlays and logos: Image generation can add generated text and logo overlays to the images. Add overlays shows text you provide along with your creative. If your logo file has transparency issues, if your brand colors are not set correctly in branding presets, or if the AI misinterprets your visual style, the generated overlays may not match your brand standards.
Promotions: Enhance CTA can surface promotions such as discounts or free shipping. If you opt in to promotions, Meta may dynamically personalize the CTA with information like a percentage off. If your offer has specific terms, expiration dates, or exclusions, a dynamically generated promotion line may omit those details or imply a broader discount than you are actually running.
Summaries: Show summaries generates a summary of selling points that appears between the footer and comments. The summary is AI-generated, and you do not write it. If your ad copy includes conditional language, disclaimers, or carefully worded claims, the summary may simplify or reword them in ways that are not accurate or compliant.
AI info labels: Ad images or videos created or significantly edited with certain Meta generative AI features may include an AI info label. The label is found via the three-dot menu, then About this ad, then AI info. Before publishing, you see a message during ad creation letting you know when an AI info label will appear. In certain regions, including the European Region, California, New York, India, and Taiwan, advertisers may self-disclose use of AI, and the label appears next to the Ad label. Ads about social issues, elections, or politics must disclose photorealistic AI-created or edited media.
Regulated categories: Financial services, pharmaceutical and health, Special Ad Categories such as housing, employment, financial services or credit, and social issues, elections, or politics may not have immediate access to all generative AI features. Certain countries may also have restricted access. If you are in a regulated vertical, assume you will need to review and approve every variation manually, and test access in a safe environment before you scale.
Write the Creative Brief and Brand Rulebook First
Before you turn on any generative AI creative enhancement, write a creative brief and a brand rulebook. The brief defines what you are testing. The rulebook defines what the AI is never allowed to say or show. Both documents are for your team and for any AI tool you use to draft concepts or check generated variations. Neither document goes into Meta. They are your internal control layer.
The Creative Brief
The creative brief should answer these questions in plain language.
Offer: What are you selling, and what is the specific offer in this campaign? For example, a free consultation, a discount code, a lead magnet, a product launch.
Audience: Who is this for? Be specific. Not business owners, but contractors who bid commercial jobs and need bonding. Not parents, but parents of kids ages eight to twelve who play travel soccer.
The one claim you can prove: What is the single most important true statement you can make, and what evidence supports it? For example, licensed and insured in Texas since 2015, five-star average on Google from 200-plus reviews, ships same day if ordered before 2 p.m. Central.
Words and claims that are banned: List any words or phrases that are prohibited by regulation, by platform policy, or by your own standards. For example, no income claims, no before-and-after photos, no guarantee of results, no use of the word cure.
Required disclaimers: List any legal or regulatory disclaimers that must appear in the ad or on the landing page. For example, results not typical, consult a physician, rates subject to credit approval.
Visual rules: What colors, fonts, and logo treatments are allowed? What objects, settings, and people may appear? What must not appear? For example, logo must appear on white or light gray background only, no people in medical settings, no before-and-after images, no stock photos of handshakes.
Tone: Describe the voice in a few words. For example, confident but not boastful, helpful but not cutesy, professional but not stiff.
The Brand Rulebook
The brand rulebook is a never let AI say list. It should be specific enough that a person reviewing a generated variation can check each item in under a minute. Here is a sample rulebook for a hypothetical med spa.
Sample Brand Rulebook: Hypothetical Med Spa
Never let AI say: Guaranteed results, permanent, cure, medical-grade (unless referring to a specific FDA-cleared device by name), pain-free, no downtime, dramatic, life-changing, miracle, anti-aging (use skin rejuvenation or address specific concerns like fine lines).
Never let AI show: Before-and-after photos (we provide them, AI does not generate them), close-ups of medical procedures, needles, blood, or clinical settings, people in hospital gowns or surgical caps, any person who is not a stock model we have licensed or a real client with a signed release.
Never let AI imply: That results are typical or that every client will see the same outcome, that a treatment is risk-free or has no side effects, that a treatment is a medical necessity, that we are a medical practice (we are a med spa under the supervision of Dr. [Name], licensed in [State]).
Always require: The disclaimer Individual results may vary, the statement Consult with a licensed provider to determine if this treatment is right for you, the name of the supervising physician if the ad mentions a specific treatment.
Visual rules: Logo on white or light blush background only, no logo on dark or busy backgrounds, brand colors are blush (#F4C2C2), charcoal (#333333), and cream (#FAF9F6), no neon or primary colors, images must show a calm, spa-like environment, no harsh lighting.
Tone: Warm, professional, and educational. Not salesy, not clinical, not overly casual.
This rulebook is a checklist. When you review AI-generated text variations or image variations, you open the rulebook and check each item. If a variation violates any rule, you do not select it, or you edit it until it complies.
Where to Put Essential Information
Meta shows variations of primary text when it predicts that can improve performance, so Meta says your description should contain only nonessential information. Put your offer, your one claim, and any required disclaimer in the primary text and headline. Write your primary text and headline carefully, and review the variations Meta generates.
Set Up Branding Presets
Branding presets let you save brand preferences for identity or ad creative. Meta is rolling this feature out gradually, so you may not see it in your account yet. When you do, set it up before you turn on any generative AI enhancement.
Meta suggests presets based on previously used URLs, previous ads, your Facebook Page, or your Shopify account if connected. You confirm, edit, or remove the suggestions. You can upload your own logo.
The presets you can define are:
Colors: Used in image generation and add overlays.
Font: Used in image generation and add overlays.
Logo: Used in image generation.
Text tone: Used in text generation.
Visual style: Used in image generation backgrounds.
Toggle Apply this branding on or off. Branding is on by default for ads using a generative AI creative enhancement and off by default for other ads. You can turn it on or off at any time.
Review the suggested presets carefully. If Meta pulled a color from an old ad that is not part of your current brand palette, remove it. If the suggested text tone is too formal or too casual, edit it. The presets are inputs to the generative AI features, and garbage in means garbage out.
Upload a clean, high-resolution logo file, then check how it looks in generated images. If your logo file has a white or colored background baked in, preview carefully to confirm the logo is legible and on an appropriate background.

The Enhancement Decision Matrix
Do not turn on every enhancement and hope for the best. Decide which enhancements to use, which to test with human review, and which to turn off. The decision depends on your offer, your vertical, your risk tolerance, and how much time you have to review variations.
For each enhancement, assign one of three settings.
On: The enhancement runs, and Meta delivers variations automatically. You review variations in Advanced preview before launch, but you do not review every variation Meta might show after launch.
On with review: The enhancement runs, and you review and approve every variation before launch. You may also spot-check live ads to confirm what is being delivered.
Off: The enhancement is turned off. You are not testing it in this campaign.
Here are two hypothetical examples, framed as starting suggestions, not proven best practices.
Hypothetical Local Service Business Example
Imagine a plumbing company running lead generation campaigns in a single metro area. The offer is a free estimate. The audience is homeowners with properties built before 1980. The company has a strong reputation and wants to protect it.
Text generation: On with review. Generate up to five variations of primary text and headline, review each one against the brand rulebook, select two or three that are accurate and on-brand.
Text improvements: Off. The company does not want Meta adjusting keywords or phrases without approval.
Enhance media text: Off. Same reason.
Image generation: On with review. Generate variations of a photo of a branded truck in front of a home. Review each generated image for logo clarity, color accuracy, and to confirm no inappropriate objects or text overlays were added. Select one or two variations to test.
Visual touch-ups: On. Allow Meta to crop and expand images to fit placements, because the original image is simple and cropping is unlikely to introduce errors.
Add animation: Off. The company prefers static images and does not want text in the image animated without review.
Add overlays: Off. The company will write its own overlay text.
Enhance CTA: Off, promotions off. The CTA is Schedule your free estimate, and the company does not want Meta adding discount language.
Show summaries: Off. The company does not want an AI-generated summary of selling points appearing below the ad.
Translations: Off. The campaign is English only.
Hypothetical Ecommerce Brand Example
Imagine an ecommerce brand selling home decor. The offer is 20% off sitewide with code FALL20. The audience is broad, targeting interests in interior design and home improvement. The brand has a library of high-quality product photos and wants to test more creative variations than the team can produce manually.
Text generation: On with review. Generate up to five variations of primary text and headline, review for accuracy and brand voice, select three to four variations.
Text improvements: On. Allow Meta to adjust keywords and phrases to fit placements, because the brand has tested this before and found the adjustments acceptable.
Enhance media text: On with review. Generate alternative text overlays, review in the Enhance media text card, approve variations that match the promotion and brand tone.
Image generation: On with review. Generate variations of product images in different environments. Review each generated image for product accuracy, color accuracy, logo clarity, and to confirm generated text overlays match the offer. Select three to five variations per product.
Visual touch-ups: On. Allow Meta to crop and expand images to fit placements.
Add animation: On. Allow Meta to add panning and zooming to static product images, turning them into short videos. Spot-check a few live ads to confirm animations are smooth and on-brand.
Add overlays: On with review. Provide overlay text such as 20% Off, Free Shipping, Shop Now, review in Advanced preview to confirm overlays are legible and placed well.
Enhance CTA: On, promotions on. Allow Meta to personalize the CTA with 20% off or Free shipping, because the promotion is straightforward and the brand has confirmed the terms are accurate.
Show summaries: Off. The brand prefers to control the message and does not want an AI-generated summary.
Translations: Off. The campaign is English only.
These are starting suggestions. Your own matrix will depend on your vertical, your risk tolerance, and what you learn from testing.
Where to Check Enhancements
During ad creation, go to Ad setup, then Create ad, then choose Single image or video or Carousel, then Ad creative, then Set up creative, then Enhancements. You can turn off individual enhancements or click Customize.
Some enhancements can only be turned off in Advanced preview and will not be visible in the main Ad creative section. To review the status of all enhancements, check both the Enhancements section and Advanced preview. In Advanced preview, go to the Placements tab to see how enhancements vary by placement, and go to the Advantage+ creative tab to see arrows on each enhancement card that show possible variations. You can click Customize and turn enhancements on or off in Advanced preview.
After publishing, go to Ads, select your ads, click Edit, and look for the Advantage+ creative section. It shows the number of enhancements turned on. Click Edit to turn off individual enhancements. Also check Advanced preview. Save your changes, then click Publish changes.
Meta offers a bulk edit option for Advantage+ creative enhancements. If you are managing many ads and want to turn off a specific enhancement across all of them, bulk edit saves time.
You can find Meta's list of Advantage+ creative enhancements in the Business Help Center, along with descriptions of what each enhancement does and which ad formats and placements support it.
Review Every AI Variation Before Launch
Do not publish an ad with generative AI enhancements turned on until you have reviewed every variation you selected. Meta shows you suggestions during ad creation, but you must actively review them. Here is how to review each type of enhancement.
Text Generation
Text generation suggestions appear as checkboxes during ad creation. Check the box next to each suggested variation you want to use. You can edit any variation by clicking into the text field. You can also click Add text option to add your own variation. Do not check a variation just because Meta suggested it. Read each one against your brand rulebook. Check for accuracy, tone, banned words, required disclaimers, and spelling.
Image Generation
On the image generation screen, you see a set of generated images. Click Generate new images to create a new set, which overrides the last set. Click the pencil icon to edit an image with an instruction, such as change background to light gray or remove text overlay. Click Refresh image with edits to apply your instruction. Click Revert to original to go back to your uploaded image. Click Save to keep the images you like. Select the variations you want to use and click Next, or click Skip and continue if you do not want to use any generated images.
Review each generated image for logo clarity, color accuracy, product accuracy, generated text overlays, and background appropriateness. If the generated image adds text you did not write, read that text carefully. If it adds a logo overlay, check that the logo is not distorted, cut off, or placed on a background that makes it unreadable. If the generated image shows people, check that hands, faces, and body proportions look natural. If the generated image shows your product, check that the product is recognizable and that colors and details match the real product.
Enhance Media Text
Enhance media text previews appear in the Enhance media text card in the Enhancements section. Click through the previews to see the alternative text overlays Meta generated. The visuals do not change, only the text. Review each text overlay for accuracy, tone, and compliance with your brand rulebook.
Advanced Preview
In Advanced preview, go to the Advantage+ creative tab. Each enhancement has a card with arrows that let you step through possible variations. Click the arrows to see what Meta might deliver. Check the Placements tab to see how enhancements vary by placement. For example, a text overlay that looks fine in Feed may be cut off in Stories.
Review Checklist
Use this checklist for every AI-generated variation before you approve it.
Claims: Does the variation make any claim you cannot prove? Does it imply a guarantee, a typical result, or a promise you do not make?
Prices and offers: If the variation mentions a price, discount, or offer, is it accurate? Are the terms correct? Is the expiration date correct?
Spelling and grammar: Are there typos, awkward phrases, or grammatical errors?
Brand colors: If the variation includes color, do the colors match your brand palette?
Logo: If the variation includes your logo, is it clear, undistorted, and on an appropriate background?
Product accuracy: If the variation shows your product, is the product recognizable? Are colors, shapes, and details accurate?
People and hands: If the variation shows people, do hands, faces, and body proportions look natural? Are there extra fingers, missing fingers, or distorted features?
Cultural fit: Does the variation fit the cultural context of your audience? Are there symbols, gestures, or settings that could be misunderstood or offensive?
Legal disclaimers: If your brand rulebook requires a disclaimer, is it present and legible?
If a variation fails any item on this checklist, do not select it, or edit it until it passes.
API Users Note
If you are using the Marketing API to create ads with text generation, and you opt in via the creative features spec, the ad status is set to PAUSED by default so you can review the generated suggestions before setting the ad to ACTIVE. You can read the suggestions in the asset feed spec. If the suggestions are not acceptable, create a new ad or creative without opting in to text generation. Meta makes no warranties on the completeness, reliability, or accuracy of suggested text, generated backgrounds, or expanded images, and advertisers are responsible for previewing ad creative featuring AI-generated creatives before publishing their ads.
Design Tests That Actually Isolate AI
If you turn on five generative AI enhancements at once, run the campaign for two weeks, and see better results, you have learned that something worked. You have not learned which enhancement contributed, whether any enhancement hurt performance, or whether the improvement came from AI or from some other change in your account, your offer, or your audience.
To learn what AI creative contributes, design tests that isolate one variable. Meta offers two testing structures: creative testing inside an existing campaign, and A/B testing. Choose the structure that fits your question.
Creative Testing in Ads Manager
Creative testing lets you compare up to seven creative variants inside an existing campaign. Meta provides delivery to new test ads while your existing ads keep running. The advantage is that you do not need to merge test ads into another campaign afterward, and the delivery system retains learnings from the test. The disadvantage is that creative testing is supported only in campaigns using the Highest volume bid strategy, and results do not include a confidence level.
To set up a creative test in Ads Manager, go to the ad level, find the Creative test section, and click Set up test. Choose two to seven copies of the ad. After you confirm setup, update the creative for each test ad. For example, you might create one test ad with your original human-written copy and image, and a second test ad with the same copy and image but with text generation turned on and three AI variations selected.
Choose the amount of your existing budget to spend across the test ads. Meta suggests no more than 20% of your existing budget to reduce potential impact on overall performance. Meta aims to spend that amount per day on the test ads. Set the test duration and choose the metric you want to use to compare test ads, ideally the one closest to the result your campaign optimizes for. Test ads start running on the ad set schedule when you publish. The test schedule only determines when results are measured.
At the end of the test, you receive results by email. You can also view results by clicking the beaker icon for charts view, or in the Experiments tool. The test makes no automatic changes. Test ads keep running until you turn them off. Results identify the top performing test ads against the comparison metric, plus additional metrics.
A/B Testing
A/B testing compares two versions by changing one variable. Each version is shown to a separate segment of your audience, and nobody sees both. You can create an A/B test by duplicating a campaign, ad set, or ad and changing a variable, or by comparing two existing campaigns or ad sets. Meta recommends using the same budget for both versions. A/B tests measure cost per result or cost per conversion lift.
Meta does not recommend informal testing such as turning ad sets on and off manually. That approach leads to inefficient delivery, unreliable results, and overlapping audiences. Use the Experiments tool in Ads Manager or create the A/B test in Ads Manager.
Best practices for A/B tests: test one variable, have a measurable hypothesis, make sure your audience is large enough and not being used by other campaigns at the same time, run the test for at least seven days, and allocate enough budget to produce enough results. Meta allows tests to run from one to thirty days, but recommends at least seven. If your customers typically take more than seven days to convert, run the test longer.
What to Compare
To isolate what AI contributes, compare a control version with no AI enhancements against a test version with one AI enhancement turned on. For example:
Text generation test: Control ad with human-written primary text and headline. Test ad with the same image but text generation turned on and three AI variations selected.
Image generation test: Control ad with your original product photo and human-written copy. Test ad with the same copy but image generation turned on and three AI-generated image variations selected.
Enhance CTA test: Control ad with a standard CTA such as Learn more. Test ad with enhance CTA turned on and promotions opted in.
Add animation test: Control ad with a static image. Test ad with the same image but add animation turned on.
If you stack multiple enhancements in the test version, you will not know which one drove the result. For example, if you turn on text generation, image generation, and add animation all in the same test ad, and that ad performs better, you do not know whether the lift came from better copy, better images, animation, or some combination. You also do not know whether one enhancement hurt performance but was offset by another.
Run one test at a time. For a creative test, the test ads keep running in the existing campaign until you turn them off, so turn off the losers and keep the winners. For an A/B test, keep the winning version and turn off the other. Then design the next test.
Naming Conventions and a Variation Log
If you run ten ads with AI enhancements turned on, and you export results a month later, you need to know which ads used which enhancements and which AI variations were selected. Without a naming convention and a log, you are guessing.
Here is a naming pattern that works. Adapt it to your own taxonomy.
Pattern: YYYYMMDD_Offer_Angle_Format_Source_Enhancements
Example: 20261101_FreeEstimate_FastResponse_SingleImage_Human_None
Example: 20261101_FreeEstimate_FastResponse_SingleImage_AI_TextGen-ImageGen
Example: 20261108_20PercentOff_FallSale_Carousel_Human_None
Example: 20261108_20PercentOff_FallSale_SingleImage_AI_TextGen-EnhanceCTA
The date is the launch date. The offer is the hook or the promotion. The angle is the message or the concept. The format is single image, single video, carousel, or collection. The source is Human or AI. The enhancements field lists which AI enhancements are turned on, separated by hyphens, or None if no enhancements are used.
Keep the naming convention consistent across your account. If you abbreviate text generation as TextGen in one ad and TxtGen in another, your exports will be harder to read.
The Variation Log
In addition to the ad name, keep a log of which AI variations were selected for each ad. The log can be a simple spreadsheet with columns for ad name, date launched, enhancement used, number of variations selected, and notes. For example:
Ad name: 20261101_FreeEstimate_FastResponse_SingleImage_AI_TextGen
Date launched: 2026-11-01
Enhancement: Text generation
Variations selected: 3 (Primary text variations 1, 3, 5 from the suggestion list)
Notes: Variation 2 used the word guaranteed, rejected. Variation 4 was too casual, rejected.
When you export results and see that this ad performed well, you can go back to the log and see which text variations were live. If you want to iterate, you know which angles worked and which to avoid.
Read Results by Creative
Meta optimizes delivery at the ad level, but when you use AI enhancements that generate multiple variations, Meta decides which variation to show to which person. You do not control variation-level delivery. That means you should treat the AI variations in a single ad as a package, not as separate tests.
When you read results, break down by ad. In Ads Manager, go to the Ads tab, select the date range, and add columns for the metrics you care about. The metrics you care about are the ones your campaign optimizes for. If you are optimizing for leads, look at cost per lead and lead volume. If you are optimizing for purchases, look at cost per purchase and return on ad spend. Do not judge an ad on click-through rate alone, because clicks do not pay the bills. Judge on the business outcome.
Also look at downstream lead or sale quality. If you are running lead generation campaigns, check how many leads from each ad turned into qualified conversations, booked appointments, or closed sales. If your CRM (for example GoHighLevel, or Sumo AI CRM, which is built on Votel AI) stores a lead source and ad name for each lead, you can tie downstream results back to specific ads.
Remember that variation-level delivery is decided by Meta. If you selected three text variations and one image variation, Meta will show different combinations to different people based on what it predicts will perform best. Do not assume you can isolate the effect of one AI variation inside an ad; treat the selected variations as a package and record which ones were live. Screenshots and the variation log tell you what ran, and overall ad performance tells you whether the package of variations worked.
Do not call a winner too early. Wait until the scheduled end of the test before you read results. Creative test results do not include a confidence level. If you turn off ads or change budgets mid-test, you invalidate the results.
Refresh Cadence
Creative fatigues. Even AI-generated creative fatigues. Build a cadence for reviewing performance and refreshing creative.
A starting suggestion: review performance weekly, and refresh creative monthly or at the end of each test. Weekly reviews let you catch problems early. Monthly refreshes keep creative from going stale. If you are in a fast-moving vertical like ecommerce, you may need to refresh every two weeks. If you are in a slow-moving vertical like B2B services, you may be fine refreshing every six weeks.
At each refresh, decide whether to retire, iterate, or promote each ad.
Retire: The ad is not performing, and there is no clear fix. Turn it off.
Iterate: The ad shows promise, but something is off. Change one variable and test again. For example, if the AI-generated text variations are getting clicks but not converting, try new text angles with the same image. If the AI-generated image variations are converting but the cost per result is too high, try a different product or a different visual style.
Promote: The ad is a winner. Keep it running, increase budget if you have room, and use the concept as a template for new ads.
Do not just generate more AI variations of the same concept. If you have been running three AI-generated text variations of We install water heaters fast for a month, and performance is flat, generating five more variations of the same message will not help. Try a new concept. For example, Licensed and insured since 2015, or Same-day service available, or We fix what other plumbers mess up. Then use AI to generate variations of the new concept.
The goal of AI creative is not to eliminate the need for new ideas. The goal is to test more variations of good ideas faster than you could by hand.
Where ChatGPT, Claude, and AI Agents Fit
Meta's generative AI creative features run inside Ads Manager and generate variations based on the inputs you provide. ChatGPT, Claude, and AI agents like Muse for Small Business run outside Ads Manager and help you with briefing, drafting, checking, naming, and summarizing. They should not publish ads on their own. A person approves what goes live.
Briefing
Use ChatGPT or Claude to draft your creative brief and brand rulebook. Give the AI your offer, your audience, your one claim, and your tone, and ask it to write a brief. Review and edit the output. The AI can help you think through what to include, but you own the final document.
Generating Human-Written Variant Drafts
Before you turn on Meta's text generation, use ChatGPT or Claude to draft five to ten human-written copy variants. Give the AI your brief and your brand rulebook, and ask it to write primary text and headline variations. Review each one, edit for accuracy and tone, and select the best two or three to test. These human-written variants become your control in tests against Meta's AI-generated variants.
Checking AI Variations Against the Brand Rulebook
When Meta generates text variations, copy them into ChatGPT or Claude along with your brand rulebook, and ask the AI to check each variation for compliance. The AI can flag banned words, missing disclaimers, unsupported claims, and tone mismatches faster than you can read line by line. You still review the flagged items yourself, but the AI does the first pass.
Naming
Use ChatGPT or Claude to generate ad names that follow your naming convention. Give the AI the pattern, the offer, the angle, the format, the source, and the enhancements, and ask it to write the name. This is especially useful if you are launching dozens of ads at once and do not want to type each name by hand.
Summarizing Exported Results
When you export results from Ads Manager, the export includes the columns you selected. Use ChatGPT or Claude to summarize the results by creative. Upload the CSV, tell the AI which metrics matter, and ask it to rank ads by performance and flag any anomalies. Review the summary, then dig into the details in Ads Manager. Do not trust the AI summary alone. Exported numbers should be checked against Ads Manager to confirm accuracy.
AI Agents
Meta offers two AI tools that connect to your ad account. Meta AI Business Assistant is a separate Meta tool in Ads Manager for campaign analysis and account issues.
Muse for Small Business is a separate tool that can connect to your Meta ad account and help analyze your ads and content performance and draft next week's campaign for review. Muse asks for your approval before it publishes, sends, spends, or shares, and you review its drafts before anything goes live. The value of Muse is in the brief and the first draft, not in handing over control.
If you use an AI agent built on ChatGPT, Claude, or another platform, the same rule applies. The agent drafts, checks, and summarizes. A person approves every live change. Do not connect an AI agent to your ad account with publishing permissions unless you have tested it thoroughly in a safe environment and you have approval gates in place.
Common Mistakes
Here are the mistakes operators make most often when they start using Meta Ads AI creative, and how to avoid them.
Turning on every enhancement without reviewing what they do. Some enhancements may be on by default. If you do not check Advanced preview before publishing, you may not know what Meta is adjusting. Review the status of all enhancements in both the Enhancements section and Advanced preview before you launch.
Not writing a brand rulebook. If you do not define what the AI is never allowed to say or show, you cannot review variations consistently. Write the rulebook first, then turn on enhancements.
Selecting every AI-generated variation Meta suggests. Meta generates suggestions, not approved copy. Read each variation against your brand rulebook. If a variation violates a rule, do not select it.
Stacking multiple enhancements in a single test. If you turn on text generation, image generation, and add animation all at once, you will not know which enhancement contributed to the result. Test one enhancement at a time.
Not logging which variations were selected. If you do not keep a variation log, you will not be able to tie results back to specific creative decisions. Log which variations went live for every ad.
Judging results on clicks instead of the business outcome. High click-through rate does not matter if the clicks do not convert. Judge ads on the metric your campaign optimizes for, and on downstream lead or sale quality.
Calling a winner too early. If you turn off ads or change budgets before the test is complete, you invalidate the results. Wait until the scheduled end of the test before reading results. Creative test results do not include a confidence level.
Generating more AI variations of a concept that is not working. If a message is flat, generating five more variations of the same message will not fix it. Test a new concept.
Not refreshing creative. Even AI-generated creative fatigues. Build a cadence for reviewing performance and refreshing creative monthly or at the end of each test.
Assuming the AI will follow your brand guidelines without checking. Branding presets help, but they are not foolproof. Review every AI-generated variation before launch, and spot-check live ads to confirm what is being delivered.
Launch Checklist
Before you publish any ad with generative AI enhancements turned on, run through this checklist.
Creative brief written and approved.
Brand rulebook written and approved.
Branding presets reviewed and edited.
Enhancement decision matrix completed: for each enhancement, you have decided On, On with review, or Off.
Enhancements checked in both the Enhancements section and Advanced preview.
Every AI-generated text variation reviewed against the brand rulebook.
Every AI-generated image variation reviewed for logo clarity, color accuracy, product accuracy, generated text overlays, people and hands, and cultural fit.
Every enhance media text variation reviewed for accuracy and tone.
Advanced preview reviewed: stepped through all enhancement cards, checked Placements tab.
Ad name follows naming convention and includes source and enhancements.
Variation log updated with which variations were selected and any notes.
Test design documented: control version, test version, variable being tested, duration, comparison metric.
Budget allocated: for creative tests, no more than 20% of existing budget suggested; for A/B tests, same budget for both versions.
Legal disclaimers present and legible if required.
AI info label reviewed if applicable.
If every item is checked, publish. If any item is not checked, do not publish until it is.
Implementation Help
Meta's generative AI creative features are powerful, but they are not plug and play. You need a brief, a rulebook, branding presets, an enhancement decision matrix, a review process, a test design, a naming convention, a variation log, and a refresh cadence. That is a system, not a switch.
SumoGrowth is an AI implementation and consulting agency that designs, builds, and maintains AI growth systems. If you want help setting up this system (brief, rulebook, presets, review process, test design, naming, log), we can help.
by Darius Gaynor


