You're optimizing for one type of AI Search. There are at least three.

The same prompt can surface completely different brands depending on how the model responds. The goal is to show up across all of those outcomes.

We spend most of our week reading GEO research so you don’t have to. This week, a few different studies we stumbled across changed how we think about AI Search.

For two years, we've treated AI Search visibility like one ranking to optimize for. In reality, there isn't just one AI Search. There are several, running side by side, and they don’t always agree on who or what to recommend.

Here’s the map of the problem:

The diagram clarifies this argument in one picture: there’s one buyer and one question, but three different sets of conditions and three different answers about which solutions to trust.

‘Fast’ and ‘deep’ are two different search engines

Semrush and Kevin Indig (growth advisor and author of the Growth Memo newsletter) ran the same prompts through ChatGPT twice: once in its quick-answer mode and once in its high-reasoning “Thinking” mode. Only about a quarter of the cited sources overlapped between the two. In short: The type of content that surfaced shifted depending on how hard the model was thinking.

In good news for content teams, recommendations tend to change in a predictable direction. Fast answers lean on user-generated content and community posts; deep reasoning leans on official documentation, government pages, and reference-grade material.

Share of citations by source type: minimal vs. high reasoning in ChatGPT. Source: Semrush, 2026.

The chart above illustrates this cleanly: User-generated and Reddit-style sources carry fast-mode answers, while official docs and government sources climb in deep mode. A brand that optimizes only for one of these has, without knowing it, optimized for one setting of a dial the buyer controls.

The ‘reasoning’ dial isn’t the only one that affects an LLM’s answers

The mode is one axis, but there are others. Separate analyses from Chris Green (Technical Director at Torque Partnership) and Suganthan Mohanadasan (co-founder of Snippet Digital) found that ChatGPT’s cited sources also change depending on which hidden retrieval pipeline it routes a query through. Semrush’s larger index of 126 million prompts adds a third axis: Across 22 industries, the brands winning AI answers change from one vertical to the next.

That means there are three separate axes in play — reasoning mode, retrieval pipeline, and category — and any one of them can alter how AI answers a user query. The “AI visibility” number a vendor sells you is an average across conditions that behave like separate channels, and the average hides exactly the gap you need to see. It is a bit like grading a restaurant on the mean of its breakfast and dinner service and calling it one score.

One strategy shows up whichever way the dials turn

Stack mode, pipeline, and category together, and AI Search starts to look like a moving target you can't aim at. But underneath the fragmentation sits a key, consistent finding that marketing teams can use to their advantage.

Across three editions of Muck Rack’s study, earned media (meaning third-party editorial coverage rather than your own site) accounts for 82 to 89% of everything AI cites. Paid and advertorial content sits at 0.3%. The pattern holds across ChatGPT, Claude, and Gemini.

An independent University of Toronto study (Chen et al., 2025) found the same overwhelming bias toward earned, authoritative sources over brand-owned and social content.

That is one facet of a durable GEO strategy: Getting coverage in the outlets that AI already reads helps get you cited in both fast mode and deep mode, across pipelines, and in your category.

There is a catch worth noting: Muck Rack's December 2025 report also found only about a 2% overlap between the journalists PR teams actually pitch and the ones AI cites. In other words, most brands are doing earned media, just aimed at the wrong people.

What to take away from all this

Here’s my take on how to turn a messy diagnosis into a fairly short to-do list:

1. Stop tracking a single AI visibility number. Check how you appear across engines (ChatGPT, Gemini, Perplexity), and within ChatGPT, across both fast and deep reasoning.

2. Audit for your category, not the average. AI-answer winners differ by vertical, which means your SEO ranking tells you less about your AI standing than you'd think.

3. Find the outlets and writers AI actually cites for your category, and aim your earned-media effort there. The 2% pitch gap is the easiest win on this list.

4. Make your owned content legible enough to be quoted verbatim: clear claims, real data, and named experts are what survive when the model switches between fast and deep reasoning.

The brands that keep treating AI Search as one contest will keep optimizing for one setting of a dial their buyers are already turning without them. The ones that build real, cited authority stop worrying about the dial at all, because they show up whichever way it turns. That’s the version I’d bet on.

---

---

AI moved fast this week, with headlines pointing in very different directions. Here’s what stood out.

• OpenAI says ChatGPT ad dismissals dropped by half. The company reports users are waving off in-chat ads 50% less than at launch, crediting better relevance and usefulness. The number is self-reported, so read it as a direction of travel rather than a verdict.

• Google shipped a spam update and new AI labels in the same week. The June core spam update brought ranking volatility, while new “strongest match” and “strong match” labels in Google Ads flag how closely a sponsored result matched the query. Worth a Search Console check against your top pages.

• Amazon bought its first ChatGPT ads, and limited what OpenAI learns from them. The retailer wants the reach of ChatGPT’s audience without handing OpenAI a clear view of which shoppers convert. Advertising on a rival’s assistant while denying it your data is about the most Amazon way to enter a channel.

• Consumers aren’t fans of AI-made ads. New Harris Poll data finds large majorities are more skeptical of advertising they believe was generated by AI. File under: reach is cheap, credibility isn’t.

• Oracle tied 21,000 job cuts to AI in a regulatory filing. The disclosure is the largest to date that names AI as the direct cause of layoffs, about 13% of the workforce.

---

• Head of Growth Marketing @ Abridge (San Francisco, CA; $250,000–$275,000 USD annual salary)

• Senior Director, Digital & Growth Marketing @ Fastly (New York, NY; $209,560–$288,000)

• Director of Organic Growth @ Scale Media (California, USA; $200,000–$300,000)

---