Why Brand-Safety Tools Built for Programmatic Ads Often Misjudge Influencer Content

What "brand safety" means in programmatic advertising — and why influencer content is a different problem

Brand-safety and ad-verification vendors such as DoubleVerify and Integral Ad Science (IAS) were built for one specific job: keeping a programmatic display or video ad from appearing next to unsuitable content — violence, hate speech, pirated material — by automatically classifying the text and context around each ad placement, often before the bid is even placed. The term comes from the IAB (Interactive Advertising Bureau) and the rest of the programmatic ecosystem, where a single ad can run across thousands of unknown web pages a second and no human can approve each placement individually.

Influencer and creator content is a different problem. It isn't a third-party page the advertiser has never seen — it's typically a single, often long-form video, made by a named person the brand deliberately chose to work with. Yet more brands and their media agencies are increasingly pointing the same vendors and the same underlying classification models at influencer content, to judge whether it's "safe" to advertise alongside or be associated with.

It's also different from the checks this Academy already covers: whether the audience is real (see how to spot fake followers), whether it matches your target market (see how to check an influencer's audience before you pay), or whether some of the engagement is coordinated rather than organic (see engagement pods). This article covers a fourth, separate layer: how third-party technology assesses the content itself for ad suitability — not the creator, and not the audience.

Why the same models struggle specifically with creator content

The core problem, per a detailed June 2026 review of industry complaints, is that brand-safety vendors do not distinguish between content that discusses a topic and content that promotes it. The distinction matters — a veteran speaking openly about PTSD, a nurse explaining cancer symptoms, or a journalist covering a geopolitical conflict is not the same as content glorifying violence or illness. But a classification model built to scan keywords and page-level context often doesn't catch that difference, particularly in long-form video, where meaning sits in tone and framing rather than in any single word.

This isn't a marginal issue. DoubleVerify reported $656.8 million in 2024 revenue and IAS reported $530.1 million the same year — both figures from the companies' own earnings releases. A meaningful part of that infrastructure is still built on a legacy from web and display advertising: page-level categorisation and keyword-based content analysis, not a model trained specifically to interpret a 90-second clip from a named creator.

Vendors are now building creator-specific tools — but the core critique persists

It's worth being precise here: vendors aren't standing still. DoubleVerify expanded its AI-powered "Universal Content Intelligence" engine — which, per the company's own description, analyses video, image, audio, speech and text — to cover Meta Threads globally in October 2025, having previously extended similar coverage to TikTok in April 2024. That's a genuine attempt to build a model trained on social and creator content rather than only web pages.

Independent critique from 2026 argues the context-blindness persists even with the newer technology, pointing to independent auditors such as Adalytics, who have documented for several years cases where ads ran alongside content the verification tools should have caught. Make Influence has not independently verified Adalytics' specific cases and reports them here as industry critique, not as an independently confirmed fact.

Two different checks — and why one doesn't replace the other

It's easy to conflate this article's subject with the manual brand-safety review you run on an influencer before signing. They're related, but they answer different questions:

See the brand-safety checklist for influencer partnerships for the manual process. The two checks answer different questions and shouldn't replace each other.

How to review a vendor's methodology before trusting a flag

If your agency or media buy already runs a brand-safety vendor against influencer or social content, it's worth asking a few concrete questions before letting an automated flag decide whether a partnership or a campaign gets pulled:

• What training data does the classification rely on? Is the model trained specifically on social/creator content, or is it a general web model applied outside its original purpose?
• When was the model last updated for the specific platform and content format you're using it on?
• How does it handle creator-level context — does it distinguish a single video from a creator's overall history and tone?
• What appeal process exists if a creator or brand believes a flag is wrong?

This is Make Influence's own recommendation on what's worth asking — not a standardised industry process, and not a guarantee of an accurate result regardless of the answers.

A hypothetical example

The example below is invented and for illustration only. It is not a real Make Influence customer.

A nurse-creator posts an informative video about early testicular cancer symptoms as part of a health brand's campaign. An automated brand-safety flag, trained to treat the keyword "cancer" as a risk topic without further context, marks the content "sensitive/unsuitable" and excludes it from part of the media budget — even though the content is actually driving a high completion rate and positive comments. Had the brand had a manual appeal process, or known to ask the vendor how the model distinguishes health information from scare content, the misclassification could have been corrected before the budget was reallocated.

Common mistakes

• Treating an automated "unsafe" flag as an objective fact rather than one model's interpretation, which can be wrong in either direction.
• Using the same vendor and thresholds for influencer content as for programmatic display without asking whether the model is actually trained for the format.
• Conflating automated content scanning with the manual brand-safety vetting of the creator themselves — they answer different questions, see the table above.
• No appeal process — leaving an incorrect flag unchallenged because nobody asked the vendor how to dispute it.

Make Influence's operational perspective

Make Influence doesn't use automated brand-safety scanning tools of the kind described in this article to assess creator content — our own vetting is the manual process described in the brand-safety checklist, combined with results tracked per collaboration. That's an operational judgement on our part, not a claim that automated scanning never has value — for brands buying significant volumes of programmatic inventory around creator content, it can still be one signal among several, provided you ask the methodology questions listed above.

Our experience is that the real cost sits less often in the content an automated model correctly flags, and more often in the time and budget a misclassification costs when nobody looks closer at it.

FAQ

Is this the same as the brand-safety checklist for influencer partnerships?

No. The brand-safety checklist is about assessing an influencer's public conduct and history before you sign. This article is about automated vendor tools that continuously scan the content itself for ad adjacency — a different layer that typically comes into play after a partnership is already live.

Does Make Influence use DoubleVerify or IAS?

No, not for our own creator vetting — see the operational-perspective section above. We can't speak to whether individual brands or agencies we work with use these tools in their own media buying.

Does this mean automated brand-safety tools are worthless for influencer marketing?

No. They can still catch genuinely problematic content at a scale manual review can't reach. The point is that an automated flag shouldn't stand alone as a verdict without a human, context-aware review — particularly for content sitting close to the line between "discusses" and "promotes".

Where can I read more about the vendors' own technology?

Both companies describe their methodology in their own press releases and on their company sites, but neither publishes the full technical detail of training data or model architecture, so an outside assessment of accuracy is inherently limited.