TikTok ads targeting rewards a habit most advertisers learned on other platforms and then have to unlearn: building a precise audience before launch. On a recommendation-driven feed, the system finds the audience by watching who responds, and the biggest constraint you can impose on it is a narrow targeting set that starves it of signal.
That does not mean targeting is irrelevant. It means the job has changed. You set the boundaries that are genuinely non-negotiable, you feed the system good conversion data, and you do the real audience work in the creative itself.
Set the hard constraints, leave the rest open
Some settings are facts about your business and should always be applied: the countries and languages you can actually serve, the age floor your product or regulation requires, the devices your offer works on. These are not optimisation choices, and getting them wrong wastes spend in a way no algorithm can fix.
Interest and behaviour categories are a different matter. Layering several of them narrows delivery quickly, and on a platform where interest is inferred from viewing behaviour rather than declared, those categories are approximations. A broad audience with strong creative usually outperforms a tightly stacked one, because the system gets enough room to find pockets you would never have selected.
- Apply location, language, age floor and device constraints deliberately.
- Start with one interest or behaviour signal at most, or none at all.
- Avoid stacking exclusions early; each one removes learning surface.
- Widen before you narrow — narrowing a working campaign is easier than rescuing a starved one.
Creative is the targeting
On a feed driven by response, the video decides who sees it. An opening that names a situation — a job, a frustration, a moment — self-selects an audience more accurately than an interest category, because it filters on recognition rather than inference. Someone who sees their own problem in the first two seconds keeps watching, and the system takes that as instruction.
This is why running several distinct creative angles beats running several targeting sets with the same video. Each angle addresses a different audience, and the platform will tell you which angle found people worth spending on. If you want to reach a segment, make an ad that only that segment would feel spoken to by.
Custom and lookalike audiences: where precision still pays
First-party audiences are the exception to the broad rule, because they carry information the algorithm cannot infer. Customer lists, site visitors, video viewers and profile engagers describe real relationships. Retargeting these groups with a different message — not the same acquisition video — is usually the highest-return targeting decision available.
Lookalikes built from a meaningful source can work well, but the source quality decides the outcome. A lookalike from all site visitors mostly models people who visit websites. A lookalike from repeat purchasers models something worth finding. Small, high-quality sources beat large, mixed ones.
Feed the system clean signal
Targeting quality depends on conversion tracking quality. If the event you optimise for fires unreliably, or fires on something trivial, the system optimises toward the wrong people with complete confidence. Verify the event you care about actually records, and choose the deepest event that has enough volume to learn from.
Where purchase volume is too low to support optimisation, optimise for a reliable upstream event and watch the downstream rate separately. Optimising for a rare event on a small budget produces erratic delivery and expensive silence.
Audience overlap and fragmentation
Running many ad groups against overlapping audiences makes them compete with each other, raising costs and splitting the conversion data that each one needs. Consolidation is usually the fix: fewer ad groups, more creatives inside them, budget concentrated enough to exit learning.
Exclusions are worth applying between funnel stages — keep recent converters out of acquisition campaigns — but a maze of cross-exclusions between similar groups costs more in lost learning than it saves in duplication.
- Consolidate ad groups until each gets enough conversions to be judged.
- Separate acquisition and retargeting, and exclude converters from acquisition.
- Refresh creative before rebuilding targeting when performance decays — fatigue is usually creative, not audience.
Reading results without fooling yourself
Short-form video drives a lot of activity that does not appear as a click. People see an ad, do nothing, and search for the brand later. Judging targeting purely on in-platform last-click will systematically undervalue the broad campaigns that create that demand and overvalue retargeting, which harvests it.
Watch branded search and direct traffic alongside platform reporting, and use pause tests when a campaign’s contribution is genuinely in question. If total demand falls when a broad campaign stops, it was doing work the attribution model could not see.
A sensible starting structure
For most advertisers: one acquisition campaign with broad targeting inside the hard constraints, four to six creatives representing genuinely different angles, optimisation on a reliable conversion event, and a separate retargeting campaign for site visitors and engagers with a distinct message. Let it run long enough to produce data, then cut creatives rather than audiences.
The advertisers who struggle on this platform are rarely the ones who targeted too broadly. They are the ones who built an elaborate audience architecture around a single video that nobody wanted to watch.
Keep reading: Tiktok Ads · Social Media Marketing