Happy Sunday, Everyone!

I trust you’re all enjoying the last few weeks before the Holiday rush begins. It’s hard to believe we’re less than 10 business days from BFCM and that 2026 dates are appearing on the extended forecast.

As I’ve had some time to reflect on conference season, there’s been one topic that has dominated almost every conversation: how to “win” AI search. That’s the question marketing execs everywhere - legal, senior living, B2B SaaS, eComm, even CPG - are trying to solve.

I get it. ChatGPT processes between 1-2B messages a day (though SEMRush found that there are ~9 messages per “search” and only 30% to 40% of all ChatGPT messages have search-like intent - even still, that’s 66.7M - 88.9M searches every day). On Google, AI overviews now appear for approximately 53% of all queries (that’s about 4.5B AIO overviews generated every single day).

That’s a lot of numbers to come to one conclusion: AI-augmented search is a big deal. Generative AI has fundamentally changed how billions of people find information. Put simply, that means that GenAI is changing search.

Those of you who have subscribed to this newsletter or heard me speak probably are familiar with my test for media buying: the risk/reward for ad platforms gets interesting when drop shippers start advertising. The corollary to that is my thesis for marketing: when niche SaaS “solutions” start appearing, the underlying thing is likely more than a fad. Well, in the past month, I’ve received dozens of decks from companies focused on GEO - everything from visibility detectors to content optimization tools to full “AHRefs for AEO” platforms.

I think it’s safe to say we’re past the “fad” stage, but we’re far from a finished product. The entire space is evolving at a truly mind-boggling rate.

That’s both the challenge and the opportunity.

This is a space I’ve been watching for years – and over the last 12-18 months, we’ve spent a great deal of time trying to understand the mechanics behind AI search. We’ve analyzed thousands of AI Overviews and GenAI outputs. We’ve run dozens of different tests across just about every industry we could think of. We’ve read a truly remarkable number of Google patents and attended more webinars, informational sessions and conference presentations than I care to count – all in an effort to better understand three big questions:

1. How do AI platforms determine what content to feature in AI outputs?
2. What can we do in order to maximize the chances of a given brand/piece of content being included in AI outputs?
3. Which types of AI outputs are most valuable to the business/brand?

Before I share what we’ve found so far, I want to include two disclaimers: (1) this is a rapidly evolving space. Just because something worked yesterday does not mean it will work tomorrow. Past precedent is not indicative of future performance. (2) Correlation is not causation. AI systems are - by their very nature - partially non-deterministic. We analyze the relationships between actions (changes, inputs, etc.) and outputs (what appears in AI Overviews, ChatGPT answers, etc.). There may be confounding variables of which we are unaware.

With those out of the way, let’s get to what we’ve found and what we think we know.

Earlier this year, I conducted an analysis of 125 AI Overviews (“AIO”) across 18 topics - everything from law + legal services to senior care, online shopping, product recommendations and service-specific questions. For each AIO, we documented how many third-party references (links) were included, then we reviewed each included link in its entirety, both manually AND with AI tools. In spaces where we did not have deep expertise, we also reviewed other page-1 articles on the same topic, from the same SERP, to better understand the quality and depth of content in the space. We then clustered the results to find patterns, and applied those learnings to a few sites (including my own) to see if this resulted in greater visibility in AI search.

What we found is that there are (at least) four factors that tend to be common among articles/content that is included in AI Overviews:

Factor #1: Exceptional Content

This has been an undying refrain in the world of SEO for a decade, but here it genuinely appears to be true. The articles that tend to be linked in AI Overviews tend to include authoritative, well-researched, properly-cited, comprehensive and differentiated content relative to the other articles available in the specific niche/space. There’s a lot here, so let’s break down what I mean by each term:

Authoritative Content:

The content is written by a bona fide person whose profile/publication history/biography is indicative of someone who is a true expert in the space and qualified to write/speak on the topic at hand. An example of this is Ben Evans being featured in an AIO asking, “What is meant by the phrase ‘software is eating the world’?” - Ben Evans was a major at Andreessen Horowitz for years and has given multiple oft-cited presentations on the topic. He is - by all accounts - a genuine expert in this particular area.

We observed a similar phenomenon in senior living + financial queries. The first link for the search “What factors to consider when choosing a senior living community” was this article in US News by award-winning health journalist Paul Wynn. And for “How to plan for retirement in your 30s”, this Investopedia article by highly-regarded, oft-cited personal finance author Lucy Lazarony was #1.

In cases where included articles did NOT have authorship information, it was often (70%+) the case that the article was written by a highly credible organization or group, such as this article by Fidelity (in response to the retirement in your 30s query above) or this “Ultimate Guide” by WGK Law in Baltimore that was #2 for “How to pick a personal injury lawyer after a car accident.” In both cases, the article did NOT have authorship information, but was published by a credible organization within the specific space. A similar pattern repeated itself across virtually every one of the other 15 topics we explored.

The Takeaway: including well-written, comprehensive author profiles tends to correlate with greater inclusion in AI Overviews. This is particularly true if those author profiles are robust and include links to other articles written on the topic, awards won from 3rd party organizations, and/or citations in media.

Well-Researched Content:

If “authoritative” refers to who is behind the content, “well-researched” refers to the substance that person is bringing to the table. In every topic examined - across senior living, personal finance, legal queries, health, and professional services - AI Overviews consistently favored content that displayed clear signs of deep research. Put another way: the articles that were cited weren’t burn-and-churn blog posts created with surface-level understanding; to the contrary, they were pieces that demonstrated a deep understanding of the topic, inclusive of its nuances and grounded in factual, verifiable material.

A defining characteristic of the well-researched articles was the presence of multiple, diverse, highly credible sources woven directly into the piece. You can clearly see this in action in the US News article on senior living - there are six (!!) different recognized experts cited throughout the article. These are not decorative references - they are structural elements that strengthen and support the article’s core thesis. Authors routinely drew from governmental bodies, regulatory agencies, academic research, landmark studies, proprietary analyses, legal statutes or long-standing industry frameworks. The quality of the content in these articles tended to mimic professional-grade explainers or practitioner guides vs. more traditional content created for SEO purposes.

Another factor we observed was that, while the volume of sources mattered, so too did their quality and relevance. It appears that AI systems are looking not just as citations, but at relevant citations. As an example, in the AIO for “Differences between roth 401k and roth ira” - the #1 article included multiple links to the relevant IRS guidelines, along with multiple, credible sector-specific links like Investopedia. The same pattern held true in senior living, with multiple included articles linking to CMS or state health departments and/or the NIH, CDC or credible medical organizations like Johns Hopkins, the Mayo Clinic or Cleveland Clinic when explaining health conditions.

These choices signaled not just diligence but comprehension - an understanding of which sources actually matter in a given domain and why those sources should be trusted.

Another thing we observed: well-researched content often used citations as a form of logical scaffolding. Claims were built sequentially, with data and evidence supporting each step. Rather than making unsupported assertions, these articles unpacked the reasoning, showed the data behind the assumption then tied it back to the core point of the article. That structure made the content inherently more reliable, and therefore more compatible with AI Overviews.

The research depth also showed up in how articles included in AIOs managed nuance/complexity in each topic. Instead of reducing complicated topics into simplistic terms, well-researched articles tended to explain contextual variation. This included explaining how rules differ by state, how tax treatment differs based on income level, filing status and jurisdiction, why different replacement window options might be recommended in different home types + geographies, how product recommendations may shift based on a user’s needs/use case, how senior living costs vary by care level, or how legal outcomes shift based on jurisdiction. These nuances matter because they mirror the variability of real-world conditions, reducing the need for the model to interpolate or “fill in the gaps” on its own. Put differently: research-driven nuance makes the model’s job easier.

Finally, well-researched articles demonstrated completeness. They answered not only the main query but also addressed the constellation of sub-questions an informed reader might logically ask next. That thoroughness was evidence of a research process that considered the topic holistically. This last part was interesting, because it reinforced what we read in several recent Google patents, including: US 20240338396A1 “Contextual Estimation of Link Information Gain” (2024), US 20230185814A1 “Pipeline for Document Scoring” (2023) and WO 2024086134A1 “Document Scoring System”.

So, while we can’t make a definitive claim about completeness being a proxy for credibility, it does stand to reason that Google patent filings show that completeness (and related signals such as information gain, semantic completeness) are factors in document scoring/ranking.

The Takeaways:

• Anchor major claims in primary or highly reputable secondary sources
• Prioritize depth of research over volume of content
• Use evidence as structural support, not decorative reference
• Incorporate nuance, variability, and edge cases as part of the explanation
• Build the article as if synthesizing research, not assembling keywords

Properly Cited Content:

Originally, I was going to include “properly cited” as part of well-researched, but after reviewing about 500 articles, I came to the conclusion this was something different. Well-researched refers to the structure and content of the document; properly-cited refers to how the content is presented, attributed and validated. There’s a clear relationship between the two, but fundamentally, these seem to be different things.

AI Overviews consistently feature articles that not only use credible sources but make the sourcing legible to machines and humans. This was one of the clearest differentiators between content that appeared in AIOs and content that didn’t.

This is consistent with how Google has described its systems in both patents and public statements. Google’s 2022 patent, “Contextual Estimation of Link Information Gain” (US11354342B2) is the most obvious example. In that patent, Google outlines a mechanism for scoring documents based on the additional, verifiable information they contribute beyond what the system has already encountered. The core idea is straightforward: documents that add validated information, rather than repeat or remix existing material, should be prioritized as they create more value to the user and the system. This is something that we’ve heard in the SEO world for years - “Write content that is unique” - but with a slight twist: proper citation/structure seems to be one mechanism for making the additional information included in a well-researched article / piece of content discoverable and verifiable to Google/ChatGPT.

And again, this comes through in our analysis of those AIOs: virtually every AIO-included article we reviewed had at least one external link to a qualified, relevant, reputable source.

Another relevant disclosure appears in Google’s broader family of patent filings around document evaluation, such as WO2024086134A1, which references the concept of “knowledge utility” as part of an overall document quality score. While you (probably) don’t want to spend the next few days falling asleep while reading it, the gist is quite clear: content must be demonstrably rooted in reputable, external evidence to maximize its perceived utility. Citations - clear, explicit, verifiable - are one of the simplest ways to communicate this connection to search engines + AI systems.

Google has also publicly reinforced this direction. In its Search Quality Rater Guidelines (and in multiple Search Liaison statements over the past few years), Google explicitly emphasizes the importance of citing authoritative sources, especially for topics governed by expertise, regulation or high-stakes decision-making (YMYL queries in SEO parlance). While these statements are not algorithmic disclosures, they are directionally aligned: transparency, verifiability and source clarity are important factors in the content Google prioritizes, especially in response to high-leverage queries.

When we examined the articles that were cited by AI Overviews and ChatGPT, these principles show up everywhere. Most articles didn’t hide references or make vague claims to “studies” or “research” – to the contrary, they were directly linked to relevant government agencies, regulatory bodies, academic research, industry standards organizations and expert-authored publications.

In addition to links/references, the second thing we noticed was that markup mattered. Pages that used structured citation formats - inline references, footnotes, dedicated reference sections - were included more than those that did not.

My working theory is that this structure made the lineage of each claim easier for models to trace. We know that GenAI systems are prediction machines - they predict the next-most-logical word based on the surrounding context. It therefore stands to reason that including highly relevant links adjacent to claims would help the system map assertions back to reputable sources. This, in turn, is likely to reduce ambiguity and ultimately lowers the system’s hallucination risk.

All of this leads me to one, inescapable conclusion: properly cited content has an outsized impact in GenAI outputs because it does two things: (1) it reads as more credible to humans AND (2) it is more computationally interpretable to AI systems.

The Takeaways:

• Include citations to primary or highly credible secondary sources
• Make citations explicit, structured, and unambiguous
• Use reference sections or consistent inline attribution patterns
• Prioritize sources Google already trusts: government, academic, regulatory and/or credible institutional outlets

Comprehensive Content:

Comprehensiveness is one of the strongest and most consistent signals in content selected for AI Overviews and in other AI outputs (e.g., the ChatGPT responses we tested). What surfaced in AIOs were articles that demonstrated a complete, end-to-end understanding of the topic: definitions, context, nuance, exceptions, comparisons, methodologies and the subsequent follow-on/adjacent/related queries a well-informed reader might ask.

Translated: this is not not “comprehensive” in the way the SEO industry has historically butchered the term (aka 3,000 words of loosely related filler). This is true comprehensiveness - a piece of content that actually helps the reader understand the topic at hand and what to do next.

If that sounds familiar, it should. Google (and Microsoft) have made multiple big public deals about content quality and task accomplishment. This seems to be a direct extension of that. Write content that helps the primary audience understand the entire topic and what to do about it.

As with everything else in this section, there seems to be a great deal of alignment between what we saw in the actual AIOs and what we found in Google’s patents + public statements/documents on this topic. Several Google patents and public-facing documents describe completeness - semantic, contextual and/or informational - as a desired property within the systems that evaluate content.

One example: Google’s patent “Contextual Estimation of Link Information Gain” (US11354342B2), although centered on “additional information,” implicitly relies on a document’s ability to offer more than expected, with the expectation set based on other public documents in the space. A document cannot “add” information unless it contains a broader or deeper set of ideas than what is already known or indexed. The takeaway is clear: content that meaningfully expands on or adds new concepts to a given area is more valuable.

If that wasn’t enough to convince you to create richer, more comprehensive documents, check out this patent: US20210124801A1. If you don’t want to wade through ~15 pages of technical jargon, here’s the tl;dr: systems prefer documents that cover the conceptual domain thoroughly enough to serve as a reliable reference source. While this specific filing is scoped to question-and-answer systems rather than ranking, the architectural logic overlaps directly with how AI Overviews work: when synthesizing answers, the model gravitates toward sources with the most complete, domain-aligned information because they minimize the need for the system to infer missing details.

Publicly, Google has echoed similar themes. In multiple Search Quality Rater Guidelines updates, as well as statements from Google Search Liaison, Big G repeatedly emphasizes the importance of “high-quality, in-depth information” and warns against “thin coverage” of complex topics. Again, these statements - by themselves - are vague and should not be taken as algorithmic disclosures. But, when you combine the patents, the public statements AND the real-world data on what’s being included in AI Overviews, the picture becomes much clearer: Google wants sources that thoroughly satisfy user intent on topics where accuracy, completeness and/or sector-specific expertise are highly relevant or required.

And that brings me to what we found.

Of the ~700 articles included in those AI Overviews, the vast majority answered both the immediate query AND multiple logical follow-on questions. For the query, “How to choose replacement windows for an older home”, the #1 article for us (a replacement windows buying guide from consumer reports) was positively comprehensive - including specifics on what to look for, explainers on the different types of replacement windows, how variation in materials and construction impacts efficacy and durability, what questions to ask when selecting a manufacturer and installer and much more. Similarly, this USN article on choosing a senior living community didn’t stop at features and amenities; it mentioned licensing, care models, staffing questions/concerns, pricing structures (buy in vs. rental), safety + wandering protocols, dining questions, what to ask about activities, plus had multiple links to other quality articles on each sub-topic.

One thing to note: not all of these articles were excessively long - most were between 800 and 1,500 words. Some were as long as 3,500 words. But we (and the AI systems) characterized most as “complete.”

The articles anticipated what the audience would want to know next and addressed it proactively, either in a subsequent section, in a linked article or as part of a FAQ area. Many articles (especially in YMYL areas) also included details on contextual variance: how answers change depending on the state, industry, demographic group, health condition or regulatory environment. That granularity matters because it reduces the cognitive distance between the query and the correct answer for the reader/user - something AI Overviews appear strongly optimized to favor.

My take on this is that comprehensiveness functions as both an “easy button” and a reliability signal. When an article captures the full scope of the challenge/concern/request, the model doesn’t need to stitch together partial answers from multiple sources or invent connective logic. It already has everything it needs in one place. The patents suggest Google has been architecting systems to value exactly that property for more than a decade. The data we found suggests that this is exactly what’s happening in AIOs: the content that answers the question best AND makes Google’s life easiest is tending to win increased visibility in AI Overviews.

The Takeaways:

• Cover the entire query space, including the adjacent questions a knowledgeable reader would expect
• Aim for completeness rather than length. There’s no magic number of words, so instead, focus on covering the topic, not just including more words.
• Include context, nuance and variability across jurisdictions, vehicles, scenarios and/or regulations
• Treat the article as a comprehensive reference that reduces the model’s need to infer missing details

Differentiated Content:

Differentiation may be the most misunderstood of all the quality signals, but in the context of AI Overviews, it’s also one of the most strategically important. Differentiated content isn’t about rewriting the same material with different adjectives—it’s about contributing something substantively new: unique insight, original analysis, practitioner context, proprietary data, or expert perspective that doesn't already exist in the indexed corpus.

Google’s own technical documentation, including the Contextual Estimation of Link Information Gain patent, underscores this principle. Although the patent language is abstract, the underlying logic is simple: content that brings additional, non-redundant information is more valuable to the system than content that merely restates what is already known. Put differently, differentiation isn’t a nice-to-have; it’s an input the model can actually work with.

But here’s the deeper truth—and the opportunity: almost every organization has something unique to offer, but almost none of them actually use it.

Most companies default to creating surface-level, commoditized content. They summarize, paraphrase, or lightly remix what’s already ranking. They treat content as an SEO checkbox instead of a strategic asset. Meanwhile, sitting inside the organization is a reservoir of differentiation that models would immediately recognize and value:

Proprietary processes

• A law firm with a unique intake methodology
• A financial advisory with a structured framework for retirement optimization
• A senior living operator with a specialized staffing model or care philosophy.

These aren’t marketing fluff; they’re operational processes/values that are unique to your organization. Don’t be afraid to showcase them to the extent possible/reasonable.

Original research or internal data

• Benchmarks reports + studies
• Survey results
• Authoritative guides + original research
• Outcomes data
• Industry-specific research or data

Each one of these is the kind of information that, when published, becomes an authoritative reference point for the entire category.

Team expertise

• A 30-year operator explaining the nuances of assisted living licensing
• A CFP breaking down tax implications the average writer wouldn’t know to look for
• A former regulator explaining what actually triggers investigations
• A roofing contractor with 25 years of field experience explaining why certain types of roof installations are predisposed to failure

Each one of these integrates authority and original thinking/experience in a way the system is unlikely to find elsewhere - which makes the content both unique and uniquely valuable to the reader. Quite simply, most of the articles we found from brands tended to include some of this content.

The irony is that organizations spend years developing these capabilities internally….only to publish content that ignores all of it. One thing that was abundantly clear to me (other than that I genuinely don’t enjoy reading patents) was that brands - whether intentionally or not - choose to compete on sameness rather than unique value/differentiation. Brands produce content that blends in rather than stands out. But, as of now, it strongly seems that AI systems reward content with incremental, verifiable, domain-specific value, which makes this particular brand of “playing it safe” a competitive liability.

What I observed across AI Overview citations was the opposite. The articles being cited didn’t just answer the question; they contributed something meaningful to the space. They included information we didn’t find in other articles included in the 10 blue links on P1. Most of them contained several examples of unique value or differentiation - whether that’s a differentiated approach, a particular care system (i.e. VITALStrong from FutureCare), a specific methodology (We asked “what makes eComm advertising agencies unique” and the response extensively Common Thread Collective’s “Prophit System”).

This is why differentiated content wins. Redundant content gives the model next-to-no incremental value, which makes including it a net-negative to the system (i.e. why would the system give you something if your content provided nothing?). Differentiated content reduces user + model uncertainty, fills gaps and provides the system with higher-fidelity inputs, which tend to produce higher-quality outputs.

Factor #2: Written For Your Actual Audience

One of the most consistent patterns I observed in analyzing AI Overviews was that the content being selected wasn’t just authoritative, well-researched, properly cited, or comprehensive - it was written for a particular audience using language/terminology those people actually use and understand.

This may sound obvious. It isn’t. Most brands do not do this.

A shocking amount of brand content is written for an imaginary persona conceived in a brainstorming session or an audience-of-everyone that ends up resonating with no one. It’s bland, generic and blah.

But, when I reviewed AI overviews, the opposite was true. The content that was included was rooted in real audience insight. It was written with a clear understanding of what the reader knows, what they don’t, what they fear, what they hope for, what they’re trying to accomplish and what stands in their way. Put another way, the overwhelming majority of the articles featured were written by someone who understands the emotional and informational context surrounding the query.

This aligns perfectly with something I’ve been saying since 2016: you cannot persuade someone you do not understand. Every successful marketing strategy begins with clarity about the audience’s goals, their anxieties, their constraints, the tradeoffs they’re forced to make and the result they’re trying to achieve.

That is the kind of insight + content that comes from only one thing: legitimate, well-done audience research.

The content that surfaced in AI Overviews reflected that audience understanding implicitly. As an example, the #2 cited article for “How to audit a eComm Meta Ads account” was this one from Marin. What was noteworthy about it was that it was written for someone who was familiar with Meta Ads - using common acronyms and basic concepts without additional context, while explaining concepts (like the importance of exclusions) that might not be fully understood by the audience. As you read the article, you’ll notice how It anticipated the obvious follow-up questions because the author understood the reader’s knowledge gaps.

The same thing was true in Senior Living, personal finance, diet/cooking, exercise/fitness, etc. – the articles that ranked consistently were written for a specific audience, and they were written in such a way that the target audience would “get it”.

To quote something from Aaron Orendorff, the article gave a damn about the person reading it. It respected the reader’s time and knowledge, while providing real, genuine value.

Audience-specific writing also creates structural clarity. When you deeply understand the audience, the content becomes more organized, more intuitive, more readable, and more aligned with how people naturally think. Models pick up on that. Not because the models care about human emotion, but because clear, contextual content leaves fewer gaps for them to fill. It reduces ambiguity, which reduces hallucination risk. In a very real sense, writing for people becomes a reliability signal in AI systems.

We saw this again and again in the best-performing AIO-included pages. A senior living guide that acknowledges the emotional burden - specifically guilt - that falls on the most common “influencer” - the oldest adult daughter. A legal explainer written as if the reader has never hired an attorney before. A financial guide that defines every acronym because the author understands that most readers don’t naturally know what a 403(b) is. A home improvement guide that uses plain language because the audience isn’t made up of general contractors.

All of this is the direct result of brands that take audience research seriously, then actually use that research to create content worth reading/sharing. Overwhelmingly, the content that showed up in AIOs + ChatGPT responses wasn’t the generic, middle-of-the-bell-curve bullshit - it was legitimately good, well-researched, “wow-that’s-actually-interesting” content.

For years, I’ve said that the best marketing is simply the best understanding of the audience, translated into words + actions. That same principle holds true here.

Exceptional content - and the kind of content that AI Overviews tend to feature - is written for someone specific, using language that person can understand, from someone they are inclined to trust.

Factor #3: Created In The Formats Your Audience Is Likely To Consume

The next pattern I found was both ingenious and deceptively simple: content that appeared in AI Overviews was often content that had been created in formats the target audience already uses and prefers. It wasn’t just long-form articles. Most AIOs included 1-2 video links (usually from YouTube). About 60% included a third-party social platform like Quora or Reddit. 1-in-5 AIOs included at least one other media type (audio content, infographics, etc.).

Across hundreds of AIO-included pages, we saw the same trend:

• Short-form explainer videos embedded directly in the article
• Infographics breaking down decision criteria or workflows
• Step-by-step image carousels for DIY topics or recipes
• Podcast excerpts embedded to explain nuance
• Original research or studies linked or included
• Conference decks/presentations integrated into the article
• Photos of slides or charts
• Diagrams, charts, and call-out boxes to provide structure and clarity

We’ve all heard for some time that search is becoming multi-modal. But as we reviewed these AIOs, that phrase took on a new meaning. We actually saw hundreds of YouTube video links - like “How to audit a Google Ads Account” included FIVE YouTube videos among the 9 cited links. “How to choose the right replacement windows for a house in Maryland” included 3 YouTube videos from 3 different channels. The AIO for “How to select a senior living community for my mother with dementia” included two YouTube videos, plus two in-depth articles and a helpful guide from the NIH.

While there are many theories for why this is the case, I have my own two-pronged theory: first, multimodal formats communicate information more efficiently. A 90-second video clip often explains a concept better than 900 words. While video has lower information density, it often has higher intelligibility. The same thing is true for images + tables - well-designed flowchart can replace an entire page of text. When content includes these elements, it becomes more digestible, more clear and more likely to be understood.

Second, models can parse multimedia structure even when they cannot “see” the images. Most AI systems can “see” images, and they’ve gotten exponentially better at parsing the content in videos via transcripts, context clues, alt text, captions, headings, file names and contextual elements (like preceding and following paragraph text). Basically: rich formats make the content easier for people to consume and easier for systems to evaluate.

It’s not surprising that some of the best AIO-included content resembled a weird fusion of a Wikipedia article, a buyer’s guide, a short course and a media kit. It was like an Ogre or an onion - it had layers (I have a 4 year old and a 2 year old. We watch a LOT of Shrek). There was something for everyone - a video that provided an in-depth review for a mobile user, a well-written article for an in-depth reader, and a quick summary infographic for the browser who just needed the jist.

This exercise was the clearest indication to me yet that the future of search is multimodal. The surprising thing is that the content winning in AI Overviews already reflects that.

The Takeaway: Create content in the formats your audience actually consumes - video, imagery, visual guides, short-form and long-form text and layered multimedia that helps the reader understand faster and better. If you’re not sure what content your audience actually wants/needs, either (a) ask them, (b) observe them using heatmaps/scrollmaps/analytics and/or (c) research their content consumption preferences using a tool like SparkToro. If the audience over-indexes on video-centric platforms (YouTube, IG, TikTok), then create more videos; if they over-index on text-focused platforms (X, Reddit, Quora, LinkedIn), then tailor your content to those preferences.

Factor #4: Distributed Where The Audience Is Likely To Be

Finally - and this one was a genuine surprise - the content that appears in AI Overviews tends to be the same kind of content that naturally travels within the platforms the target audience already uses and trusts (in many cases, we inferred the audience using our existing client/audience research). We validated this using both AI systems and - where possible - Sparktoro. We also found these articles tended to have a significant number of links (thanks, Moz)/references, which we took as an added signal that the content was likely to be shared/distributed.

Put another way - most AI systems aren’t surfacing obscure, unread, un-shared articles simply because they are optimized. To the contrary, these systems are increasingly featuring content that people like the searcher actually engage with, in the places they naturally spend time.

When we traced the backlinks, mentions, embeds, citations, social references, newsletters and cross-links of the content included in AI Overviews, we found a striking pattern: in many cases, the content already had traction among the target audience.

Examples include:

• Senior living content heavily shared in caregiver Reddit threads
• PI content that was shared on social platforms like IG and X
• Personal finance content that was shared on X, LinkedIn and YouTube
• Product comparison/buying guides that were shared on niche forums
• B2B SaaS content that was shared in industry groups (I’m in a few, so I actually saw them), in conference decks and in industry-specific YouTube channels.

Obviously, this pattern is more difficult to definitively state, simply because I don’t know every place to look – but from what we could see, in our limited sample, the upshot was clear. The content that resonates with actual people ends up in the channels those people trust. Models pick up on these signals. The exact process here is opaque, but my inference is that there are multiple modalities: explicitly through links/references and implicitly through engagement patterns on the actual AIO.

When a given piece of content is already circulating in the right ecosystems, the model has more evidence that the content is useful, credible and worthy of inclusion.

This aligns with Google’s long-standing preference for “signals of trust,” but it also reflects something deeper about the nature of AI search: AI systems are heavily influenced by behavioral signals embedded across the web. Content that people trust enough to share, bookmark, reference, quote, embed and/or link to is the same content that AI systems tend to consider reliable.

Candidly, that makes intuitive sense. If your target audience doesn’t engage with your content in the wild, why would an AI model feature it? How does that facilitate task accomplishment or information gain? How does it improve the experience for the user (which, ultimately, is Google’s/ChatGPT’s goal – be useful so that you are used more).

The Takeaway: Ensure your content is distributed in the channels your audience already uses and trusts. Spend the time to get to know the platforms where your primary target audience shares, discusses and discovers information. Proactively distribute your content in these places - either via seeding, direct outreach, ongoing engagement, whatever.

Factor #5: Technically Optimized (At Least, To A Point)

The last pattern worth mentioning is technical optimization…though not in the way most SEOs might think.

Our preliminary findings were surprisingly unremarkable. Most of the articles included in AI Overviews were technically solid: reasonably fast loading, mobile-friendly, cleanly structured, and supported by at least some schema or structured data markup. But very few were hyper-optimized or pristine by SEO-industry standards. Not all of them passed Core Web Vitals. Not all of them had flawless scores in PageSpeed Insights. And many had minor technical imperfections that would send an SEO audit into a tailspin.

In other words: technical excellence appears to help, but it is not the deciding factor. Good is good enough. Exceptional content - not perfect page speed/CWV scores - seems to be doing more of the heavy lifting.

Obviously, your content must be crawlable in order for it to be eligible to be included in any type of search results (including the ones used to generate AI overviews). Your site should be secure. Your pages should load on mobile and desktop. These are all “expected” functionalities, and candidly, every article had these in a passable sense.

There’s also a meaningful nuance here. While using Schema doesn’t appear to be a strict inclusion factor for AI Overviews, it does show up often enough that it’s hard to ignore. From both a first-principles perspective and an observed-data perspective, it makes sense: structured data helps information retrieval systems understand what a page is, who wrote it, what it claims and how it fits into the broader context of the topic. It makes sense that an information retrieval system (like the one Google uses to select which articles to then consider for inclusion in an AIO) would recognize and potentially score a page with appropriate structured data higher than one without.

But, I also didn’t see a ton of evidence that using Schema is a proverbial Golden ticket to inclusion (as some people on SEO Twitter have suggested). The reality is probably somewhere in the middle: the right schema almost certainly reduces ambiguity, improves interpretability and increases the likelihood that content is included in AIOs. Put simply: schema probably won’t get you into an AI Overview on its own, but it can help, and it rarely hurts. If the goal is to appear in more AI-mediated search experiences, adding relevant structured data is a low-cost, low-friction step with meaningful upside and virtually no downside. Don’t abuse it, and don’t spam it, your mileage may vary, etc.

The big takeaway here is that the content winning in AI Overviews is not doing so because it is technically perfect. It’s winning because it is useful, credible, differentiated and aligned with user intent. The technical foundations of the article is a necessary, but not sufficient, condition for inclusion. To use an analogy, the foundation of a home is almost never the reason it wins an award from “Better Homes & Gardens”. All the technical stuff does is make it easier for systems to understand and integrate the exceptional content already there.

There’s a lot here. Much of this is still evolving. None of this is definitive. But the patterns we’ve found from testing are quite clear: AI search isn’t rewarding hacks, shortcuts or gimmicks.

It’s rewarding the brands doing the real work.

The content showing up in AI Overviews is rarely accidental. It isn’t luck. And it certainly isn’t the result of clever prompt engineering or another round of “SEO best practices.” It’s the predictable output of something far more fundamental: brands that deeply understand their audience, create genuinely helpful content rooted in expertise and evidence, differentiate themselves with something meaningful and deliver that content in the formats and channels their audience actually uses.

In a way, AI hasn’t changed what “great content” means - it has finally revealed it. The models are simply reflecting back what actual people (not SEOs) have been screaming for years: Give me something real. Something useful. Something that respects my time and increases my understanding. Something that actually helps me get the thing I need to do, done.

If AI search has a north star, it seems to be this: reward the content that makes the world a little clearer.

As a marketer, that is liberating. Everything I’ve outlined here - authority, research, citations, completeness, differentiation, audience alignment, multimodality, natural distribution - comes down to one deceptively simple mandate: make something worth finding, then distribute it where it can be found. The brands that embrace this are going to win disproportionate visibility in AI search. Not because they gamed the system, but because they built something the system is designed to elevate.

We are early in this shift. Things will continue to change. I have no doubt the specifics will evolve, but I think the foundational principles outlined here will stay. The brands that take this moment seriously - the ones willing to rethink their content from first principles, invest in expertise, embed audience research into every decision, and publish work that actually earns attention - will be the ones we’re all pointing to 12–18 months from now, wondering how they moved so fast.

If I learned anything from reading a few thousand pages for a single conference slide, it’s this: AI isn’t killing search. It’s clarifying it. And for marketers willing to do the work, that’s the biggest opportunity we’ve had in a decade.

As for the final question - whether AI overviews can drive legitimate business results - consider this chart of leads received:

This shows the number of leads from AI sources (Gemini, ChatGPT) in dark blue, with the percentage of total leads (the full bar) in yellow. Over the last 7 months, the percentage of leads coming from AI sources has increased nearly 10x - from just over 2% to just under 20%.

We’re still early - but the results are clear: AI Overviews are here to stay - and we’re just starting to see the impact they can have.

That’s all for this week!

Cheers,

Sam

Loving The Digital Download?
Share this Newsletter with a friend by visiting my public feed.

​

​

Follow Me on my Socials

​​​

​

1700 South Road, Baltimore, MD 21209 | 410-367-2700
​Unsubscribe | Manage Preferences​