If you have spent any time navigating Generative Engine Optimization (GEO) or Answer Engine Optimization (AEO) professional circles over the past year, you have likely heard the same strategy repeated on an endless loop: Be on Reddit. Be on YouTube. Meet your buyers where active online communities already place their trust.
While that advice holds merit for a prospective buyer who is still wrapping their head around an emerging category or trying to understand a complex problem space, new industry data reveals it completely misses the mark once that same buyer shifts phases. When a buyer stops browsing broad topics and starts comparing named vendor options, the digital footprints that AI models rely on change dramatically.
A recent empirical study conducted by marketing agency Ten Speed and analyzed by Nelson Brassell challenges conventional wisdom about where brands should invest their limited content budgets. However, looking past the initial press release headlines reveals a nuanced picture—one uncovered only after subjecting the underlying methodology to rigorous, pointed questioning.
Main Facts: The Shift from Community Forums to Brand-Controlled Assets
The core finding of the Ten Speed investigation is striking: when B2B buyers reach the active evaluation and vendor-comparison stage—using generative AI tools like ChatGPT, Perplexity, Claude, and Gemini—they are rarely being directed to community-driven discussion boards.
Instead, AI models lean heavily on brand-controllable web properties. Product pages, company-managed blogs, direct comparisons, and structured directory profiles dominate the citation landscape.
Key high-level takeaways from the research include:
- Brand-Controllable Dominance: Web pages directly written, owned, and managed by marketing teams account for a staggering 88.3% of all citations recorded during the evaluation phase.
- Product Pages Lead: Product pages alone capture 24.1% of all AI citations, making them the single largest category for bottom-of-funnel queries.
- The Low Share of Forums: Reddit, YouTube, and independent forums combined account for just 4.2% of total citations, with Reddit carrying the vast majority of that small share and YouTube barely registering at roughly 1%.
Rather than relying on peer-to-peer discussions when making a final shortlist, AI search engines systematically prefer authoritative, structured, and first-party vendor resources.
Chronology: How the Study Was Built and Scrutinized
To understand how these conclusions were reached, it is necessary to examine the operational timeline and methodology behind the research.
Step 1: Defining the Prompt Architecture
Ten Speed designed 170 unique prompts explicitly engineered to mirror how a B2B buyer speaks to an AI assistant during advanced research stages. These prompts bypassed top-of-funnel definitions (such as "What is a CRM?") and targeted deep evaluation scenarios, such as "Pipedrive vs. HubSpot for sales-led companies" or "How does Software X handle SOC 2 compliance reporting?"
Step 2: Tracking Citations via Peec AI
The prompts were deployed across Peec AI, a specialized tracking platform that monitors how major generative AI models (ChatGPT, Perplexity, Claude, and Gemini) cite specific URLs in their responses. To prevent the data from merely reflecting a single agency’s bias, the prompt set incorporated Ten Speed’s diverse client base—spanning fintech, physical security, hospitality, and IT automation—alongside look-alike industry competitors. This process yielded a total dataset of 7,387 citation appearances.
Step 3: Critical External Auditing
Following the publication of the initial report, external analysts subjected the findings to intensive scrutiny, sending the research team a series of pointed methodological questions to verify the integrity of the data, uncover potential inconsistencies, and determine the exact statistical boundaries of the conclusions.
Supporting Data: Breaking Down the 7,387 Citations
A granular look at the 7,387 citation appearances provides a roadmap of where AI models actually pull their information during the vendor selection phase.
AI Citation Distribution by Content Type:
==================================================
Product Pages: [████████████████████████] 24.1%
Articles / Blog / PR: [█████████████████ ] 17.4%
Comparison Pages: [█████████████ ] 13.0%
Listicles: [█████████████ ] 13.0%
How-To Guides: [█████████ ] 8.9%
Homepages: [████████ ] 7.8%
Directory Profiles: [████████ ] 7.2%
Reddit / YouTube / Misc:[████ ] 4.2%
==================================================1. Product Pages (24.1%)
Product pages anchor the bottom of the funnel. When an AI model needs definitive answers regarding features, integrations, and core functionality, it indexes direct product documentation above all else.
2. Articles, Blog Posts, and PR (17.4%)
Earned media, company press releases, and informative blog content continue to hold substantial weight, bridging the gap between high-level education and deep technical evaluation.
3. Comparison Pages & Listicles (~13% Each)
Comparison formats punch significantly above their weight class. While comparison prompts made up roughly 20% of the total prompt set, they generated close to 27% of all citations—yielding a 1.33x return on investment. Many B2B content teams miss this opportunity by treating "X vs. Y" pages as defensive afterthoughts rather than aggressive capture surfaces.
4. Homepages (7.8%)
Homepages emerged as a surprising citation driver. AI models frequently pull macro-level positioning and overarching brand summaries directly from root domains when evaluating a company’s general suitability.
5. Directory Profiles (7.2%)
Platforms like G2, Capterra, and Software Advice act as structured databases for AI search engines. When users request software options within a specific category, models rely on these profiles’ structured data, category tags, and integration lists.
6. Forums and Video (4.2%)
Reddit and YouTube—frequently hyped as the primary battlegrounds for AI visibility—combined for a meager slice of the pie. Reddit performed the bulk of this work, while YouTube barely registered at around 1% for these evaluation-stage queries.
Official Responses and Methodological Transparency
A study is only as reliable as its authors’ willingness to address its limitations. When pressed on statistical discrepancies, data boundaries, and missing variables, the researchers provided transparent—if sobering—clarifications.
Resolving Data Inconsistencies
Initial versions of the published prompt breakdown table cited 220 total prompts, creating a contradiction with the 170-prompt denominator used across every other chart in the piece. Ten Speed confirmed that 170 is the correct figure and committed to correcting the mislabeled visual. Similarly, an inconsistency regarding how the "comparison" category was tagged across different charts was successfully reconciled to a consistent 34-prompt count.
Sample Size and Client Confidentiality
When asked to reveal the exact number of distinct client brands and verticals sitting behind the 7,387 citations, the research team declined to provide a precise range. Their rationale: combining a specific client count with named verticals risked re-identifying a small, confidential client base. While this boundary is understandable from a client-privacy standpoint, it introduces a limitation for independent researchers wishing to evaluate the breadth of the underlying sample.
Platform-Level Gaps and Statistical Testing
Perhaps the most crucial admissions involved data gaps:
- No Platform-Level Breakdown: ChatGPT, Perplexity, Claude, and Gemini undoubtedly exhibit different citation behaviors at the bottom of the funnel. However, the study did not isolate platform-level data during this particular extraction.
- Descriptive vs. Statistically Tested Splits: The eye-catching 88.3% versus 4.2% content-controllable split was reported as a descriptive pattern, not the result of a rigorous statistical test. While the study utilizes nonparametric tests elsewhere for citation rates across page types, the headline-grabbing split is a snapshot rather than an established law of AI behavior.
Rather than dismissing these critiques, the researchers leaned into transparency, admitting their data gaps rather than glossing over them. This honesty elevates the credibility of the findings while simultaneously warning marketers against treating a single point-in-time snapshot as an immutable industry benchmark.
Implications: What This Means for Your B2B Content Strategy
The insights extracted from Ten Speed’s research do not render community-building advice obsolete; rather, they demand a more sophisticated, funnel-aware allocation of content resources.
Marketers must adjust their strategies to match how modern buyers—and the AI models assisting them—behave during the decision-making process.
1. Optimize Product Pages and Homepages for AI Readers
Treat your product pages and homepage as though an AI model is reading them completely cold. Because product pages and homepages account for nearly a third of all evaluation-stage citations, they must clearly and plainly answer foundational questions:
- What does this software actually do?
- Exactly who is it built for?
- What native integrations does it support?
Clever, highly metaphorical brand positioning gives an AI model far less substance to work with than flat, highly specific, literal descriptions.
2. Treat Comparison Content as a Front-Line Growth Strategy
Generating a 1.33x return on comparison-format prompts proves that "X vs. Y" pages are vastly underutilized. B2B content teams should stop treating comparison pages merely as defensive battlecards for the sales team. Instead, proactively build comprehensive, objective comparison pages against multiple competitors—not just your two most obvious market rivals.
3. Actively Manage Directory Profiles as Content Assets
At 7.2% of citations, directory profiles on G2, Capterra, and similar software review sites function as structured data repositories for AI search engines. A stale description, an outdated feature list, or a miscategorized profile will directly misinform an AI model precisely when it is curating a shortlist for a ready-to-buy customer.
4. Adopt a Skeptical Framework for AI Research
Before incorporating any new AI visibility statistic into a corporate slide deck, marketing strategy document, or budget pitch, digital leaders should subject the data to rigorous scrutiny:
- What is the true denominator behind the percentages?
- Does an aggregate average across multiple platforms hide significant outlier behavior?
- Is the headline finding a statistically tested correlation, or merely a descriptive observation?
Conclusion
Navigating the evolving landscape of Generative Engine Optimization requires resisting the temptation to build multi-million dollar strategies around a single study or a viral marketing headline. The research from Ten Speed proves that when buyers are ready to make a purchase decision, AI models bypass community chatter in favor of structured, authoritative, brand-owned assets. By combining this operational insight with a healthy dose of methodological skepticism, modern marketers can align their content budgets with the reality of how enterprise decisions are made today.
