mardab96/tiktok-ads-claude-skills

Spark vs Brand Ad Split - TikTok Ads

Use this skill when

A team sees different results from Spark Ads (an existing creator post run as an ad) and brand-handle ads and wants to know whether format is the reason. Format is unproven by default, because the two usually differ in creator, concept, offer, and age. This skill checks that before comparing anything.

Common requests:

• "Do Spark ads beat our brand-handle ads?"
• "Our creator ad has half the CPA. Is that Spark or the creative?"
• "Should we move budget to boosted creator posts?"
• "Set up a fair test between the two formats."

When not to use

• A single ad's opening is weak -> hook-rate-scorer-tiktok-ads.
• An ad is tiring over time -> creative-fatigue-detector-tiktok-ads.
• A creator brief or rights wording is needed -> ugc-brief-writer-tiktok-ads.
• CPA moved in the whole account -> cpa-spike-diagnosis-tiktok-ads.

Required input

Ad-level export from Ads Manager Reporting for a stated date range, one row per ad. Column labels vary; keep them as labelled in your Ads Manager and map them in the first reply.

Field	Note
ad id, ad name	Never infer format from the name
Format (Spark or brand-handle) and ad identity or post source	From ad metadata or a screenshot of the ad detail
Creator, for Spark ads	As shown on the identity
campaign name, ad group name, objective, optimization event	Matching keys
Audience summary, placements, bidding strategy as labelled, daily budget	Matching keys
Offer per ad (user-stated), concept, length	Matching keys
start date, end date or date	Matching keys
spend, impressions, clicks, CTR, CPM, CPC	Delivery
reported conversions, CPA, conversion value, frequency	Outcome
2-second video views, 6-second video views	Optional, for the hook read

Also needed: the conversion lag, attribution setting, and a screenshot of each Spark ad's authorization status and expiry as shown in your account. Optional: engagement columns (likes, comments, shares) if the export carries them, creative previews, change log.

Analysis workflow

1. Confirm format per ad from metadata. No metadata means needs_data for that ad, and it stays out of the comparison.
2. Normalise the window: same dates, same currency and timezone, same attribution setting. Drop rows inside the conversion lag.
3. Build candidate pairs. A pair is two ads, one per format, matched on all eight dimensions in the table below.
4. Compute for each arm: spend, impressions, CPM, CTR, CPC, conversion rate (reported conversions / clicks), CPA, frequency, 2-second view rate where present. Always print the sample next to the metric.
5. Apply floors: 3,000 impressions per arm before any hook or CTR read (T02); 10 reported conversions per arm before any CPA read (the floor T04 uses for its windows, borrowed here as a pack heuristic); exclude arms inside about 25 results or 7 days of entry (T05); check T01 if either ad is old.
6. Run the confounder pass: creator versus brand concept, offer, audience size, ad age, budget skew, social proof visible on the ad, view-versus-click attribution.
7. Assign a verdict per pair. If no pair is clean, list the mismatches and design the matched test below.
8. Record authorization per Spark ad: status, scope, expiry, owner. Unknown is needs_data.

Matching dimensions

Dimension	Must match	If it does not
Objective and optimization event	Identical	Not comparable; stop
Audience	Same targeting summary and size band	Label mismatch; verdict capped at possible
Offer	Same price, discount, and end date	Offer becomes the explanation
Placements	Same set	Placement mix can drive CPM
Bidding strategy label and budget scale	Same label, similar daily budget	Spend skew changes the sample
Dates	Overlapping window, same lag treatment	Seasonality and promo timing leak in
Concept and length	Same script and cut, or documented difference	Creative is the confounder, not format
Age and learning stage	Similar days live, both past T05	Younger ad carries learning noise

Matched test design

When no clean pair exists, propose one variable: format. Use the same video: the creator's post as the Spark ad, and the identical footage uploaded as the brand-handle ad. Keep objective, optimization event, audience, placements, bidding label, offer, and schedule identical, with equal budgets, in separate ad groups. Read after both arms pass about 25 results or 7 days (T05) and the lag has cleared. State the limits: two ad groups with the same targeting can compete in the same auctions, and the Spark post may carry engagement history the upload lacks. Starting it requires approval.

Reading metrics across formats

Some gaps come from how the two formats are built, so read them as questions, not as results.

Metric	What a gap may reflect	How to check
CPM	Audience, placement mix, and competition at the time	Same audience band and placements (matching dimensions)
CTR	Creative and creator pull, plus the profile and engagement surface a post carries	Compare only with CPA and conversion rate beside it
Conversion rate and CPA	Offer, landing page, creator trust	Same landing page and offer, lag cleared
Likes, comments, shares on a Spark post	Organic history of the post, not only paid delivery	Do not treat as an ad-quality score

A cheaper CTR with a worse CPA points at a landing or offer problem, not at the format. Route that to landing-page-match-tiktok-ads.

Authorization status watch

Record status as shown in the account, as labelled in your Ads Manager. This skill never approves or extends it.

Status	Verdict	Next step
Recorded, scope covers the run, expiry after the test ends	Proceed with the comparison	Note the expiry date in the record
Recorded, expiry inside the test window	monitor	Owner decides on extension before the read
Not recorded or unreadable	needs_data, hold on budget moves toward Spark	Owner records it in writing
Revoked or lapsed during the test	hold	Stop the comparison; mark the arm incomplete

Decision rules

Condition	Verdict	Next step
Format not confirmed from metadata	needs_data	Supply ad detail screenshots
Pair unmatched on one or more dimensions	Format effect possible; mismatch itself confirmed	Matched test
Matched, floors met, CPA and CTR agree in direction in one window	possible	Repeat in a second non-overlapping window
Same, agreeing across two non-overlapping windows [pack heuristic]	likely	Matched test before shifting budget
Matched test result with both arms past T05 and lag	confirmed for that test	approval_needed for a budget decision
Either arm under the impression or conversion floor	monitor	Wait or widen
Either arm inside a learning window (T05)	hold	learning-phase-guard-tiktok-ads
Authorization missing, unknown, or near expiry	hold on extending Spark spend; needs_data	Owner records status

Output format

### Format map
| Ad id | Format (from metadata) | Creator or identity | Evidence | Authorization status |
|---|---|---|---|---|

### Matched split table
| Pair | Dimension matches (8) | Arm | Spend | Impr | CPM | CTR | CPA | Conv | Confounder | Verdict | Next controlled test |
|---|---|---|---|---|---|---|---|---|---|---|---|

### Authorization record
| Spark ad | Scope | Expiry | Owner | Status |
|---|---|---|---|---|
(as shown in the account; not approved by this skill)

### Decision table
| Finding | Evidence | Verdict | Severity | Confidence | Business impact | Next step |
|---|---|---|---|---|---|---|

### Data gaps and approval needed
[missing item] -> [verdict prevented]. [budget move or Spark extension] -> [named owner].

Practical example

Illustrative numbers only. 14 days, lag 7 days cleared, purchase objective, same audience band and placements.

Pair 1 (unmatched): Spark creator ad, spend 1,240, impressions 210,000, CPM 5.90, clicks 2,310 (CTR 1.10%), 31 conversions, CPA 40.0, carries a 15% code. Brand-handle ad, spend 1,180, impressions 148,000, CPM 7.97, clicks 1,030 (CTR 0.70%), 19 conversions, CPA 62.1, no code. Different offer and different concept. Verdict: format possible, offer mismatch confirmed, severity medium, impact spend.

Pair 2 (matched on all eight): Spark spend 620, impressions 98,000, clicks 735 (CTR 0.75%), 9 conversions, CPA 68.9. Brand spend 600, impressions 91,000, clicks 664 (CTR 0.73%), 9 conversions, CPA 66.7. Impressions clear 3,000, but 9 conversions is under 10 per arm, so the CPA read is monitor; the CTR gap is negligible.

Conclusion: the headline gap in pair 1 is offer and concept, not format. Next: the matched test with one video in both formats. Authorization: the Spark ad expires within the next window; owner status needs_data, so extending Spark spend is hold.

Common failure modes

• Reading a Spark ad's lower CPA as a format effect when it carries a different offer.
• Comparing arms with different learning stages or ad ages.
• Pooling many unmatched ads into a format average that hides one outlier.
• Inferring format from the ad name.
• Letting social proof or comments on the creator post explain results without labelling it.
• Comparing before the conversion lag clears, so recent spend looks worse.
• Running both arms as ad groups with identical audiences and ignoring auction overlap.
• Extending Spark spend after the authorization has lapsed or was never recorded.

Guardrails

• Do not declare a format winner from unmatched pairs; format stays possible until the eight dimensions match.
• Do not infer Spark or brand-handle status from an ad name.
• Do not compare CPA with fewer than 10 reported conversions per arm or hook metrics under 3,000 impressions.
• Do not mix offers, audiences, or bidding labels inside one pair without naming the mismatch.
• Do not approve, extend, or request a Spark permission; record status only.
• Do not recommend a budget shift before a controlled test; stop at an approval-ready proposal.
• Do not hide the conversion lag, attribution setting, or date window.