AI ad creative testing at scale

The short answer

AI video ad creative testing is many variants of one brief, scored. It is not one hero render.

1. Lock one brief. One product, one job for the ad. Hooks, offers, lengths, and aspect ratios are the cells. A new product is a new test, not another cell.
2. Generate the cells as takes on one lot. A lot is an ordered list of shots. Each render of a shot is a take. The homepage says “A lot is an ordered list of shots. Each renders on its own, every render is a take, and the winner is the one you keep.”
3. Score what you can see before you spend media. Did the take deliver the cell you asked for, at a quote you accept? Mount that take. Do not call a model, or a variant, the winner of a media test you have not run.
4. Hand the mounted cut to a person. A brand approver or a client gets a review link. That workflow is the agency AI video workflow.

lotgen.ai is a takes and compare workspace, not a single-clip playground and not an avatar vendor. The homepage puts it as “Dollars and seconds, not credits,” and every render is quoted in real dollars from your own provider keys before you commit. Pricing is “Two prices. No credits.” The seat is the studio. The labs bill the clip.

How to make the ad itself is a different guide: how to make UGC ads with AI. There is no run footage here, no ranked quality score, and no testing dashboard.

What “ai video ad creative testing” means here

People search ai video ad creative testing for a workflow: how a performance team tries more than one ad from the same brief. Most of what ranks for it is avatar and UGC vendors. lotgen.ai is not an avatar tool, and this guide is not about ai ugc video generator.

Here: one brief, a matrix of cells, several takes per cell, one lot so the variants stay comparable.

A cell changes one thing you can name:

Elsewhere on this site:

A montage is a pile of unrelated prompts. A test is a matrix: the cells share a brief, so a difference has a name. If you change the product, the brand, the offer, and the model in the same file, you cannot say which change you are scoring.

Four product words, from the Academy lesson What a lot actually is:

Map the matrix onto those words. One cell is one shot. Each provider you try on that cell is another take of the same shot. The take you will stand behind is the one you mount. Homepage, “Takes compete”: “Compare takes side by side, mark the winner and keep the rest on the shelf. Each one has its price on it.” That “winner” is the take you keep. It is not a media-spend result. The scoring section below says what you must not call a winner.

How many variants before it is a test

One finished render is a sample. It becomes a test when you can point at two or more cells of the same brief and say what changed.

This is a planning matrix, not a run. No prompts were submitted and no takes were kept for it. Read it as a template, not as “we tested these ads.”

Hold the product, the length, and the aspect still. Vary hook and offer only.

Six cells is enough to call it a test of hook by offer. You can name the difference between cell 1 and cell 2 (the hook) and between cell 1 and cell 4 (the offer).

Add a 6-second cutdown of the same six only if length is a variable you mean to score. Say so. That is twelve cells, not “more creative.” Add 1:1 or 16:9 on the keep, not on every miss. The UGC how-to treats a second aspect the same way: a re-render of the keep.

Stop when the next cell changes the product or the brand. That is a new brief, and it does not belong on this lot if you still want the first six to be comparable.

One cost row

Dollars for the whole board, the 50-variant model, aspect re-renders, and the client round live on AI UGC ads cost. Here, price one cell, then use the AI video cost calculator to swap key, length, and attempts. How a Unit becomes a dollar, credit table included, is on Kling pricing explained. An 8-second Kling versus Veo bill is on Kling vs Veo cost.

Take cell 1 and render it on Kling 3.0’s official API at 720p with native audio and no voice control. On Kling’s developer pricing page, read 2026-09-21, the video list price is 1 Unit = $0.14, and the 720p row “With Native Audio x No Voice Control” is 0.9 Units ($0.126) per second. The other cost pages on this site read the same row on 2026-09-18, and it had not moved. This is Kling’s published Unit price, not an invoice and not a run.

0.9 Units/s × $0.14 per Unit = $0.126/s
$0.126/s × 15s               = $1.89 per take
$1.89 × 3 attempts           = $5.67 for this one cell

Three attempts is a model, not a measured keep rate. No provider publishes one, and there is no industry figure worth citing. Change the slider on the calculator to your own.

That $5.67 is one cell, before a second aspect and before a review round. It is not a 50-variant forecast. Multiply only the cells you will actually render, on the calculator, after the matrix is locked.

If the cell is Veo, do not multiply a per-second sticker by 15. Google’s Veo 3.1 generate page, re-read 2026-09-21, lists video lengths of 4, 6, or 8 seconds. Use the cell on Veo 3 cost per second. There is no 15-second Veo number to print.

The seat is separate from the clip. On Pricing as of 2026-09-21: Solo $99 a month. Team starts at 3 seats. $90 a seat at 3, $86 at 4, $83 at 5, and $79 a seat from 6. Reaching 6 drops every seat to $79, not only the sixth. Two seats is not a plan. “Two prices. No credits.” MCP on every plan. The labs still bill the clip on your keys.

Bulk generation without a junk folder

bulk ai video generation for ads means the same brief, several providers, one lot. It does not mean a folder of files named output_final_v7.

The junk folder fails the test in two ways. You lose which file was which cell. You lose the quote that was on the take before you downloaded it. A lot keeps both: the shot is the cell, the take is the render, and the price stays on the take.

How to make the ad, including its own bulk section, is on how to make UGC ads with AI. This section is only the rule that keeps a matrix comparable.

1. One lot for the brief. The Solo plan is “unlimited lots and shots.” The seat does not meter the matrix. The labs meter each render.
2. One shot per cell. Re-render that shot when the hook is wrong. Do not re-render the other five. Academy lesson Re-rendering one shot: fixing one note costs one shot, and the take you replace stays, because you paid for it.
3. Several providers means several takes of that shot, not several folders. Homepage: compare takes side by side. Each one has its price on it.
4. Read the quote before you commit. Every render is quoted in real dollars from your own keys before you commit. A cell that would cross the spend cap does not get to become a result. Spend caps per workspace and per lot are on every plan, Solo included.

Which keys you can name

Twelve keys pay for a picture, a clip, or a voice, and they are the same twelve the cost pages price:

Google, BytePlus ARK, fal, evolink, Kling, OpenAI, MiniMax, Black Forest Labs, Alibaba, ElevenLabs, xAI and Runway.

The key is BytePlus ARK, the console that issues it, not the parent company.

The request form says thirteen platforms. The thirteenth is Anthropic, which writes copy and renders nothing.

Not every one of the twelve renders the video cell:

• ElevenLabs is voice.
• The OpenAI key pays for stills and text drafts. It is not a Sora key.
• Luma and Sora are models on fal or evolink, not keys of their own.

Which of those clip keys to use for ads is best AI video generator for ads, and it does not rank one. The product page for holding more than one provider in the same workspace is One API across your keys.

Same prompt on two models is a comparison of those two takes, not a published test. There is no run footage on this page.

What you can score without a media-spend result

You can score a matrix before anyone pays for distribution. You cannot score media performance. Those are different words, and mixing them is how a sample gets called a winner.

You can score:

• Did the take deliver the cell? Hook, offer, length, and aspect you wrote down, not a nearby substitute. Academy lesson Picking a model: every model quietly changes something about a shot it cannot make as asked, and the quote tells you which, in advance.
• What it cost. The quote before you run, and the price on the take after. Usage, Academy lesson Reading Usage: spent, kept, and discarded are three different numbers. That is a cost score. It is not a thumbstop, a CTR, a CPA, or a ROAS.
• Which take you will mount. Compare them in the lot. Mount one. Leave the rest on the shelf with their prices. The Team plan line calls this “Winners, review rounds and approvals with a name on them.” In the product, that winner is the mounted take.

You must not call a winner:

• A model. Generator choice lives on the ads page, and that page does not crown one. A take you prefer on this brief is not “the best AI video model.”
• A variant, on media math. CTR, CPA, ROAS, and thumbstop need a media-spend result you do not have yet. A cheap take is not a winning ad. An expensive take is not a losing ad. You have not bought the impression.
• The first file that finishes. Academy: the first successful take of a shot mounts itself. That is the first one that finished, not the one you would have chosen. Overrule it before you share. Export and the review link both show the mounted take.
• A ranked quality score. The product does not print one, and this guide does not invent one or dress up footage to imply measured quality.

On-pack legal copy, a claim, or a logo that must be pixel-true is a different bar. No provider page we read publishes a text-rendering guarantee. Do not treat a readable word in one frame as a pass.

How to sit with comparable takes and choose is how to compare AI video takes.

The handoff

When a cell is mounted, someone who can say yes still has to see the cut. For a brand team that is the approver. For an agency that is the client. Both of them get a review link on the lot, not a folder of every seed you already paid for.

The homepage puts it as “Clients approve by link. A review room with no login. They see the cut, never your costs, providers or prompts.”

The full workflow, including what the reviewer is allowed to see and what the verdicts are called, is the agency AI video workflow. Mount the keep before you send the link. The first finished take is not the cut.

If you post the output yourself, you still mount one take per shot. You do not need a client in the room. On the request form that is Creator: “You post the output yourself.”

If you are building an agent, pick Developer or agent builder instead: “You want the API or MCP more than the studio.” MCP is on every plan, and the how-to for that door is AI video MCP.

Agency aside: client-billable tests

Request access

lotgen.ai is invite-only while the beta is on. The door is the invitation form.

For the brand team running this matrix, the tile is Brand or in-house team: “You make ads for your own product.”

Request an invitationSeats and pricing

The paid door is the same studio, without the queue. Pricing is “Two prices. No credits.” Solo $99. Team $90 a seat at 3, $86 at 4, $83 at 5, and $79 a seat from 6. Reaching 6 drops every seat to $79, not only the sixth. Two seats is not a plan. See pricing.

To price a cell before you render it, use the AI video cost calculator. To see what an approved 15-second variant costs across keys, use AI UGC ads cost. To make the ad, use how to make UGC ads with AI.

Nothing here is a limited offer. The request form says it is “letting people in a few at a time while the beta is on.” That is a queue, not a deadline.

FAQ