Twenty clients, four creatives each, refreshed four times a year: that is 320 videos. Few agencies run this arithmetic before buying an AI video tool. Without the number, you cannot tell whether your workflow is slow, thin or generic.
This audit gives you a target, a scorecard and a pre-publish check. It builds on a sponsored post from Leadzai on Search Engine Journal, which describes a pipeline built around each client's real photos. Because the post is vendor content, treat its framing as a pitch. Its planning logic is useful, but its statistics need caution.
Read the source with the right skepticism
The post cites three sets of figures:
- 49% of U.S. consumers have used TikTok as a search engine (attributed to an Adobe Express survey).
- 86% of advertisers use or plan to use generative AI for video creative.
- A WARC study conducted with TikTok found that 90% of 400 marketing leaders use AI as a core creative tool, while only 45% reported significant quality improvements.
All of these reach you second-hand through a sponsor, and none is verified here. Find the original reports before you put any of them in a client deck.
The post also claims that ChatGPT and AI Overviews do not cite TikTok videos today, while YouTube is what AI cites. That claim is testable, not settled, so check it yourself on the queries your clients care about (see the last section). This guide promises no rankings, traffic or performance gains.
Step 1: Size your annual video target
The source gives a simple structure. Pick a creative count per campaign, set a refresh cycle, multiply by active accounts, then hold your workflow to the total. Its own example is four creatives, four refreshes and 300 clients, which gives 4,800 videos.
The formula
Creatives per campaign × refreshes per year × active accounts = annual videos.
The source starts at four creatives because TikTok's recommended range is three to five per ad group. It suggests refreshing quarterly at minimum, and sooner when click-through or conversion rate drops, which it calls the first sign of creative fatigue. Confirm the current range in TikTok's own documentation before you commit.
Worked example: a 20-client agency
These numbers are illustrative, not benchmarks.
- Base case: 4 creatives × 4 refreshes × 20 accounts = 320 videos a year. That is roughly 27 a month, or six to seven per working week.
- Lean case: 3 creatives × 4 refreshes × 20 accounts = 240 videos.
- Heavy case: 5 creatives × 6 refreshes (every two months) × 20 accounts = 600 videos.
- Multi-campaign adjustment: if three clients run two separate campaigns, add one extra campaign's worth for each (4 × 4 × 3 = 48 videos). The base case becomes 368.
The range from 240 to 600 matters more than any single figure. Pick the case that matches your clients' actual ad budgets, then compare it with what your team shipped over the last 12 months. If the gap is large, fix the workflow before you add accounts. The source gives the same advice.
Keep one inference in mind. Per-channel cuts (TikTok, Reels, Shorts) multiply your exports, not your storylines. Count storylines for the target, then count exports separately for scheduling.
Step 2: Score your workflow on three failure points
The source names three places a workflow breaks: bandwidth, volume and relevance. Score each question from 0 (fails often) to 2 (reliable). Each category has three questions, so the maximum is 6. This scoring is a working method, not an industry standard. See read about calculator page audit: is chatgpt's tool builder a threat? for additional background.
Bandwidth
Bandwidth is how much video your team and your clients can actually produce. The source notes that traditional ads involve scripts, shoots and edits that can take weeks for a single client.
- Can you deliver a first cut for any client within one working week, without a client shoot?
- Does a feedback round fit in one sitting, or does it take days?
- Does production stall when a client is slow to send assets?
Volume
Volume is whether you can hit the Step 1 target at consistent quality. The source argues that AI helps with volume but not quality, citing the WARC figures above (unverified).
- Over the last quarter, did you ship at least the refreshes you planned?
- Do you keep a library of hooks and angles, or start from zero each time?
- Can you launch a new variant (new hook, same storyline) without a new production?
Relevance
Relevance is whether the video could only belong to this client. The source describes the failure as "ads that could belong to any business."
- Does every video start from the client's own photos, logo and business details?
- Do reviewers reject footage that shows the wrong dining room, storefront or product?
- Does each client have a written storyline that sets what must not change?
Read the result
Add up each category. The lowest score marks your weakest stage.
- Low bandwidth points to intake and approvals.
- Low volume points to templates and reusable storylines.
- Low relevance points to inputs, which is where generic footage comes from.
If two scores tie, fix relevance first. Scaling generic video only produces more generic video. We cover related ground in how to build an ai visibility prompt set by tag (template) in depth.
Step 3: Test AI outputs against each client's real photos
The source describes a pipeline in which a multimodal model reads the client's photos and business information. The model then writes a storyline with a hook, problem, narrative and call to action. It also maps each scene to a reference image and lists the key elements that must never change. Finally, a video model that accepts reference images generates the footage.
The source calls reference images "non-negotiable" for local businesses. The vendor has a stake in that claim, but the logic holds. In the source's words, a video advertising your restaurant that shows a restaurant that isn't yours is worse than no video at all. Whichever tools you use, verify each output yourself.
Pre-publish checklist for every AI video
- Reference-image match. Place the client's source photos next to key frames. Check that the space, products, signage and people match. Reject anything that shows a different location or product.
- Must-not-change elements. Check the list from the storyline: logo, shopfront, signature dish, uniform colors, price or offer. Confirm each one appears correctly and stays consistent between shots.
- Text and claims. Read every frame for on-screen text, offers, addresses and phone numbers, and confirm they are accurate.
- Captions. Add captions in the client's language and check them against the script. The source says TikTok suggests five to 10 words of on-screen text per second. Check TikTok's current guidance before you rely on that figure.
- AI disclosure. The source says labeling is either set by the advertiser at upload or applied automatically by the platform. It adds that some models embed invisible watermarks, such as Google's SynthID. Make the label a default in your upload checklist, not a judgment call.
- Per-channel cuts. The source recommends vertical 9:16 at 15 to 20 seconds as the base. Then check each channel:
- TikTok gets a sound-on, native-feeling cut.
- Reels works without sound, with captions carrying the story.
- Shorts and Performance Max have the aspect ratios they need (16:9, 1:1 and 9:16).
Also read: how to audit google ads cpc increases: a worked example
The source also says Google's help documentation warns that auto-generated Performance Max videos may feature a different product than your landing page. Supply your own versions. 7. Client approval. Send the check results with the video, not just the file.
Log a pass or fail for each item. After a few weeks, the failures show which input or model needs attention.
Keep the workflow from depending on one model
The source says OpenAI shut down the Sora 2 video API with no direct replacement, and that Google retired Veo 2 and Veo 3.0. These are the sponsor's statements, so check current model status on each vendor's site before you plan around them. The underlying risk is real whichever list is accurate: models change faster than agency processes.
The source's advice is to own the durable assets: storyline guidelines, brand inputs, feedback history and a hook library. Treat the video model as a swappable component.
Each quarter, run one client's storyline and reference images through a second generator, then compare both outputs against the checklist above. If your process only works with one tool, you have found a dependency.
Test the AI search claim yourself
The source suggests that one storyline can serve both TikTok search and AI citations. Treat that as a hypothesis. For each client, pick five queries and run them in the AI search products your clients' customers use. Record which sources appear, then repeat monthly and note changes. Keep the log honest: a video that is not cited is a finding, not a failure.
What to expect next
Your first pass will probably show one stage scoring clearly lowest, and a target larger than your current output. Start by fixing that stage on three pilot clients. Run the checklist on every video and record the failures.
After a month of logs, you will know whether your problem is capacity, repeatable templates or input quality. You can then set a target from your own data rather than a vendor's figures.
