Intro
What is your AI Agent buying?
Marketers are fumbling in the dark when trying to optimize for LLM referral traffic. Referral click rates are plummeting internet-wide as consumers rely on "zero-click" AI summaries. To achieve any type of measurement, most marketers are relegated to asking the LLM itself how and why they're being ranked (see the "Synthetic Audiences", section) — the more advanced are analyzing agentic web traffic directly to see.
The good news is there are emerging academic studies helping us actually understand the what and why behind LLM recommendations — and thankfully, it's familiar.
The studies show LLMs have human-like biases, where familiar optimization techniques work:
- LLM selection can be influenced by an average of 5% based on a single word in the description: specific categories show this can be up to 25%.
- ChatGPT (specifically) is 3X more likely to choose a product in the top row.
- LLMs in general show a negative bias towards higher prices and sponsored posts, and a positive bias towards social proof and strong reviews.
While measurement of LLMs is currently more art than science, these emerging optimization techniques show promise marketers can continue to affect marketing efficacy.
Walled Gardens
Can a Gemini powered-Siri take on ChatGPT?
After toying with the idea of partnering with Anthropic, Apple is reportedly nearing an agreement with Google to use Gemini AI models to power the new Siri, according to Bloomberg.
The idea of a Google-backed Apple AI is intriguing. The more context an LLM has on your life (vis-à-vis your personal data), the more effective, personalized, and (hopefully) relevant the output will be. Google already has a huge swath of personal history on users — search, email, maps, etc. — and when combined with the data on your phone, this could lead to a dramatic leg up on data to feed an LLM.
But it's a race: today ChatGPT's daily active users are roughly five times higher than Gemini's, and every interaction further entrenches the dominant model by feeding it richer behavioral data. Opening default access to Gemini on 1.5 billion iPhones could dramatically shift the balance — potentially challenging ChatGPT's lead. This is likely why Google is willing to let Apple pay a relative bargain — just $1 billion — for access. While that amount may seem an astronomical amount for us mere mortals, Google earns 1 billion dollars a day.
AI
Synthetic audiences
New technologies often come with intimidating names. Synthetic audiences may sound complex, but the concept is simple: using AI — typically an LLM — to generate an audience persona you can test your marketing or advertising against. In fact, most LLM users have already created them without realizing it. Any time you ask a chatbot to play a role — “How would you respond if you were my boss?” or “What would you say if you were my customer?” — you’re creating a synthetic audience.
As with most things in AI, your output quality depends on your inputs. “My customer” provides little context compared to: “You’re a SaaS customer of a leading martech firm, on the bleeding edge of technology, and highly demanding of your vendors.” The data you use to build a synthetic audience can range from anonymized customer examples to high-level descriptions or definitions.
Synthetic audiences are increasingly being used to measure the opaque ranking and referral logic behind LLMs. One example is RivalSee, a company that uses an LLM to review your domain, generate a list of ICPs, and then query other LLMs on how your brand ranks for relevant topics; a clever way to get a third-party perspective on your marketing copy.
The immediate benefit of synthetic audiences is perspective. Like talking with an expert from another field, they force marketers to view their industry or product through a broader lens. A synthetic audience might overlook a key competitor — or surface one you’ve ignored. They’re also fast: a single marketer can generate insights in hours that once required months of customer surveys.
But, as with all LLM interactions, the results are inherently derivative. Synthetic audiences can provide directional guidance, but lack actionability. They may reveal patterns and perceptions but interpretation and deciding what to do next, still requires human judgment.
Amanda's Take
Apple’s services growth is increasingly an advertising story
Apple’s Services business is on track to clear $100B in annual revenue. While we tend to think of Services as App Store fees and iCloud, roughly 40% of this revenue now comes from advertising, including Apple Search Ads and the enormous Google default search deal, estimated at $20B per year.
Apple’s multi-year privacy push did more than convince the market that iPhones are secure. It also changed how advertisers allocate their budgets. With less ability to track users across apps, more spend has shifted away from third-party networks and toward Apple’s own monetization surfaces. Apple is now one of the largest advertising companies in the world — without calling itself one.
At the same time, mounting legal pressure around App Store commission rules threaten the highest-margin portion of that Services stack. If even 20% of App Store transactions shift to external payments, Apple could lose ~$11.6B annually. That dynamic only increases the strategic importance of Apple Ads and on-device monetization going forward.
Podcast
Building mission-driven products in regulated industries with Patrick Wesonga
In our latest episode of How I Grew This, Patrick Wesonga joins Amanda and Adam to share how he empowers teams to stay focused, lead with empathy, and use data as a source of influence rather than intimidation. From countering the “hippo effect” to unlocking curiosity and clarity, Patrick offers timeless lessons for anyone leading in a fast-moving product environment. [video] [podcast]