AI gets it wrong 40% of the time. Marketers are publishing anyway.
This week's newsletter practically wrote itself around one theme: AI is everywhere, but accuracy is nowhere near where it needs to be.
NP Digital tested 600 prompts across every major LLM. The best one got it right 60% of the time. That's the best. Meanwhile, half of us are hitting errors multiple times a week and a third of us have let mistakes go live. We dig into what that means for your team and your workflow.
On the news side, Perplexity is pulling away from ads, OpenAI is leaning in, Reddit is turning threads into shopping carts, and Accenture is now tying AI adoption to promotions. Plus, we spotlight Firsthand, a company founded by the people behind DoubleClick, AppNexus, and FreeWheel who are betting that the ad unit as we know it is done.
Let's get into it.
— Vas
This Week's Signals
Creative workflows
Figma x Anthropic: Code to Canvas. Claude Code now exports directly into editable Figma designs. No more rebuilding AI-generated UI from scratch, import it, tweak it, ship it. This was already a hack workflow at most teams. Now it's official. (CNBC)
Advertising ecosystem
Perplexity is backing away from ads. After being one of the first AI companies to run ads inside chat, Perplexity now says in-chat ads erode trust. (Pymnts)
Reddit turns conversations into commerce. A new AI-powered feature converts community product recommendations into shoppable carousels with pricing and retailer links. Reddit's bet: purchase intent lives in threads, not search bars. (Search Engine Land)
Other AI signals
Accenture is tracking AI usage by employee, weekly. Senior staff adoption is now tied directly to promotion decisions. When a 700,000-person consultancy makes AI usage a leadership KPI, it's no longer optional. (The Register)
RentAHuman.ai went from joke to phenomenon. After OpenClaw blew up, tens of thousands flooded the site hoping an AI would hire them. Skills posted, rates public, peak pricing hit $500/hour. One API call now summons an actual human. The punchline writes itself. (Computerworld)
Google launched Lyria 3 inside Gemini. Users can generate 30-second songs from text, photos, or video prompts. Tracks carry SynthID watermarks and plug into YouTube Shorts via Dream Track. AI-generated music just got a distribution channel. et's see if users will lean in (TechCrunch)
Framework: AI Is Wrong Almost Half the Time. Marketers Are Publishing Anyway.
NP Digital tested 600 prompts across six major LLMs, ChatGPT, Claude, Gemini, and others. The best model scored 59.7% accuracy. The worst performed far below that. None of these are passing grades. When the best tool in the market gets it wrong 4 out of 10 times, the question isn't which model to pick, it's whether your team has a system to catch what every model misses.
The tasks that fail most are the ones marketers rely on AI for most heavily: HTML and schema creation (46% daily error rate), full content development (43%), and reporting (34%). The higher the precision required, the more AI breaks down. Brainstorming and ideation? Relatively safe. Anything that touches published output or client-facing data? High risk.
The downstream cost is real. 47% of marketers encounter inaccuracies multiple times a week. 72% spend 1-5+ hours weekly just fact-checking AI output. And 57.7% say clients or stakeholders have already questioned the quality of AI-assisted work. Despite this, 36.5% admit AI errors have reached the public, and another 39.8% say they've caught close calls right before publishing.
The implication for marketing leaders: AI literacy isn't a nice-to-have. Teams that invest in verification workflows and prompt training show 3-4x better accuracy, regardless of which model or pricing tier they use. The edge isn't the tool. It's the process around it.
Source: NP Digital AI Hallucinations and Accuracy Report, Feb 2026
Company Spotlight: Firsthand
The people who built programmatic advertising think it's dead.
Firsthand was founded in 2023 by Jonathan Heller, Michael Rubenstein, and Wei Wei. Between them, they built FreeWheel (sold to Comcast for ~$360M), DoubleClick Ad Exchange (a key asset in Google's $3.1B acquisition), and AppNexus (sold to AT&T for ~$2B). These are three of the people who created the infrastructure behind modern digital advertising.
Now they're replacing it.
What Firsthand builds: AI-powered "Brand Agents" that replace static ad units. Instead of a banner, the consumer gets an AI chatbot that represents the brand, answers questions, compares products, recommends, adds to cart. The agent reads surrounding page context and adapts in real-time. A data layer called Lakebed controls exactly what the AI can and can't say, so brands keep control.
Why it matters now: This week's news shows the industry splitting on how AI and advertising coexist. Perplexity is pulling away from ads. OpenAI is testing them at $60 CPM. Anthropic mocked them at the Super Bowl. Firsthand's bet is different: don't put ads in AI, turn AI into the ad.
They've raised $32.6M from Radical Ventures and FirstMark Capital. Their angel investors read like a programmatic hall of fame: Brian O'Kelley (AppNexus founder, Scope3), David Rosenblatt (former DoubleClick CEO), Bob Lord (Horizon Media). The self-reported numbers, 3-5x engagement over traditional digital, double-digit CTRs, are unverified with no named clients yet. But the founders' track record is the credibility. They've called this kind of shift before and been right.
They're not the only ones building this. Toronto-based AdChat has been running chatbot-in-banner ads for years with clients like Audi, Mercedes, and Hershey, reporting 50-second average interactions, 10x the CPG engagement benchmark.
The question for CMOs: if the people who built the last era of ad tech are telling you the next one looks completely different, do you listen before your competitors do?
firsthand.ai, Aperiam Ventures portfolio company
Each week we are reviewing the most interesting companies in marketing and AI. If you are interested in featuring your company in a Marketing Embeddings issue please reach out here.
ONE MORE THING
Marketing Embeddings is read by 20,000+ CMOs, CTOs, and media leaders navigating AI's impact on marketing. Forward this to someone who needs to see it.