This is Scale & Strategy, the newsletter equipped with a wide-angle lens (so we can show you the full picture).
Here’s what we’ve got for you today:
How to Create Content Optimized for AI Search
Microsoft: How to Spot a Poisoned AI Model
How to Create Content Optimized for AI Search
Off-page SEO has hit a ceiling. For years, backlinks from “authoritative” sites were treated like gospel.
According to David White, building links purely for rankings now is like showing up to a space race with a flip phone.
Search is no longer a Google-only sport. It’s fragmented across Reddit, LLMs, TikTok, YouTube, and Instagram. Discovery happens everywhere, often without a click ever touching your site.
Even Google admits it: AI-driven sessions last 2–3× longer.
And zero-click isn’t a trend, it’s the default. When an AI Overview appears and you’re not cited, CTR can drop by up to 65%.
That means off-site campaigns can’t just chase links anymore. They need to create visibility inside social feeds and LLM-powered answers to offset the traffic that never arrives.
Relevance has shifted too. It’s no longer just about topical authority. It’s about where the user is in their decision journey.
People aren’t just Googling. They’re:
searching TikTok for reviews
asking LLMs for comparisons
scanning Reddit for unfiltered opinions
If your content isn’t the answer in those moments, you don’t exist.
This is where data matters. Using a traffic index lets you see which topics carry real commercial value across both traditional search and AI prompts.
Layer in social listening and subreddit scraping, and you stop guessing. You see where real conversations are happening and build directly into them.
Format is everything now. A single blog post is not a campaign.
Winning requires modular content:
short-form video for social discovery
prompt-optimized text for AI citation
opinionated takes that travel in communities
Marketers still running the “brainstorm, blog, outreach” playbook are missing the engagement layer entirely.
In 2026, off-site PR isn’t about links. It’s about becoming a source of truth across the entire digital landscape.
If you’re not optimized for AI prompts and social hooks, you’re not ranking. You’re invisible.
Third-party AI models are becoming a quiet enterprise risk. New research from Microsoft suggests it may be easier than expected to detect when a model has been sabotaged.
This week, Microsoft’s security-focused AI Red Team published findings on how to identify models that have been “backdoored”. In these cases, malicious behavior is embedded directly into a model’s weights before or during training, remaining invisible until triggered.
The researchers identified three key signatures that can reveal a poisoned model.
First, attention distortion. When a trigger phrase appears in a prompt, the model stops responding holistically and instead fixates on the trigger, steering the output toward behavior chosen by the attacker.
Second, data leakage. With the right prompting, poisoned models may inadvertently reveal information related to their own backdoor training.
Third, fuzzy activation. Backdoors don’t require exact trigger phrases. Partial or approximate inputs can still activate the malicious behavior, making detection harder and riskier.
Alongside the research, Microsoft released an open-source scanning tool designed to detect these signatures, according to Ram Shankar Siva Kumar, founder of Microsoft’s AI Red Team. The broader scope of the problem remains unclear, largely because there are no consistent standards for auditing AI models.
“The auditability of these models is all over the place,” Kumar said. “I don’t think anybody knows how pervasive the backdoor model problem is. We want to get ahead of it before it becomes unmanageable.”
This matters now more than ever. Developers are increasingly relying on open-source models to cut costs and move faster, often without the security expertise or resources needed to properly vet them.
“Microsoft is in a unique position to invest heavily in AI safety and security,” Kumar noted. “Startups don’t always have large interdisciplinary teams to do this kind of work.”
Security has long been treated as an afterthought in tech. With AI, that habit is dangerous.
As agents take on more autonomous work with less human oversight, poisoned models don’t just fail quietly. They influence decisions. They propagate errors. And they can create cascading effects that are far more damaging than a compromised software library.
In AI, a backdoor isn’t a bug. It’s a hidden decision-maker.
