Lessons from tracking 56 banking prompts across four AI systems that apply to any industry operating in the age of generative search.
Based on: Banking segment LLM analysis, Feb–Mar 2026
Models: ChatGPT · Gemini · Google AI Overviews · Google AI Mode
Tool: SEMrush LLM Visibility
For any content strategist or SEO practitioner, this data provides a rare empirical window into how LLMs evaluate and surface content. The patterns are clear, the implications are urgent, and most of them apply directly to your industry.
Before we discuss what to do, it is worth naming what the data actually shows. These are not interpretations, they are measurable shifts in how AI models responded to the same prompts one month apart.
Average brands per response dropped from 14.1 to 9.5. AI systems are becoming more curated, not broader. The circle of cited brands is shrinking
Of all matched prompts, nearly a third changed the recommended content format between February and March.
Landing pages as a recommended format dropped from 8 to 5 entries. Plain blog posts rose from 10 to 13. Educational content is being favoured over transactional pages.
Five responses from March refer to 2026. AI updates its answers and rewards more recent articles.
Each of these shifts applies across other sectors. Whether you are in healthcare, B2B SaaS, e-commerce, legal services, or manufacturing, the underlying mechanics are the same: AI models are trained on patterns of what authoritative, useful content looks like, and those patterns are changing.
In traditional SEO, it was common to optimise a landing page for a query that was fundamentally informational. The logic was: rank for the query, then convert. AI systems have broken this logic. When a user asks “what are the best tools for automating KYC verification?”, Google AI Overview, ChatGPT, and Gemini all recommend blog posts and listicles as the appropriate content format, not landing pages. The AI has already classified the intent as informational and is routing users to content that matches that intent.
The practical consequence: If your product landing page is the primary content you have optimised for a query that AI classifies as informational, you are invisible in AI-generated results. The AI will find the blog post that explains the topic, not the page that sells the product. You need both, and the blog post needs to exist and be authoritative before the landing page becomes relevant.
The 32.6% drop in average brands cited per AI response is perhaps the most important finding in this entire dataset. It tells us that AI models are not trying to be comprehensive directories of relevant brands. They are trying to cite the most credible, most authoritative, most clearly positioned sources on a topic. As the citation pool shrinks, the stakes of being in it rise sharply.
This is not about mentioning a keyword more often. It is about owning a topic so thoroughly, through depth of coverage, internal linking, external citations, author credibility, and content freshness, that the AI has no better option than to cite you.
One of the most interesting findings from this dataset is the spontaneous appearance of “2026” in AI responses, not as a result of any user asking for current information, but as the AI’s own temporal framing of its answer. When Gemini opens with “In 2026, the landscape for KYC automation has shifted from basic document scanning to Agentic AI…“, it is signalling something important: it believes its answer is current, and it wants the user to know that.
This behaviour has a direct consequence for content. If an AI system is actively looking for content that it can represent as current-year relevant, then content that clearly signals recency, through publication date, updated data, or explicit “state of [topic] in 2025/2026” framing, has a structural advantage over evergreen content that carries no temporal signal.
How to build recency into content so AI systems recognise it:
AI crawlers read schema markup. Use Article schema with datePublished and dateModified. Update the modified date every time you refresh the content, not just cosmetically, but substantively.
“The best tools for KYC automation in 2026” will outperform “The best tools for KYC automation” in AI visibility, not because of keyword stuffing but because the AI can infer temporal relevance.
Content citing a 2022 report alongside a 2024 report creates a mixed recency signal. Prioritise the most recent available sources and remove outdated ones.
Summarise what changed and why. This gives the AI a direct recency signal even for articles with older original publication dates.
Treat your most important pillar content like annual reports, scheduled for refresh every 12 months with measurable data updates.
The appearance of YouTube, Reddit, Medium, Instagram, and Facebook in recommended content formats, even in a heavily regulated, traditionally text-centric sector like banking, is not a coincidence. It reflects a broader shift in how AI systems define authoritative content.
For most of the SEO era, authority was defined by links from high-domain-authority sites. That definition is expanding. AI systems are now weighting:
YouTube is the most significant of these. It appeared as the anchor of the social format cluster in March, covering both long-form explainer content and short-form (YouTube Shorts). The fact that a query as dry as “bank bots” generated a YouTube format recommendation tells us that even highly technical, B2B-oriented topics are being evaluated for video suitability by AI systems.
YouTube has several properties that make it uniquely powerful as an AI visibility platform. It is owned by Google, making its content natively indexed by the same system that powers AI Overviews. It has automatic transcription, meaning every video creates a searchable text corpus. It has structured metadata (titles, descriptions, chapters) that AI systems can parse for topic relevance. And it has watch time and engagement data at scale, the kind of behavioural signal that AI systems weight heavily when evaluating content quality.
The YouTube strategy implication: You do not need a massive production budget or a media team. What you need is a library of structured, topic-specific videos that answer the same questions your blog content answers, with proper titles, descriptions, and chapters. A 10-minute explainer video on “how KYC automation works in 2026” can appear in AI recommendations alongside your written content, doubling your surface area for the same query.
Check our YouTube SEO Case study
Reddit’s appearance in the recommended formats for KYC automation via ChatGPT reflects something specific: AI systems have learned that users trust peer validation. A Reddit thread where practitioners discuss which tools they actually use carries a different kind of authority than a brand blog post, and AI models are starting to weight that distinction. Medium functions similarly, it is a platform associated with thoughtful, cited, practitioner-written content rather than corporate marketing.
You cannot directly control what appears on Reddit. But you can influence it through community participation, by becoming the source that Reddit threads cite when they need authoritative information, and by ensuring your content addresses the real-world concerns that practitioners discuss in forums rather than the idealised questions that appear in brand-led content.
A 30.6% format change rate in a single month is genuinely alarming from a traditional content strategy perspective. The instinct might be to chase each change, pivoting budgets from landing pages to blogs, then adding video, then building Reddit presence. That approach is reactive and expensive. The more durable response is to build a content architecture that is format-agnostic at its core and multi-format in its execution.
The central insight is this: every piece of research, data, or expert knowledge your brand produces should exist simultaneously in multiple formats. Not because you need to be everywhere, but because different AI systems weight different formats for different query types, and that weighting is shifting monthly.
The modular content production workflow:
This is the foundational document. It contains original data, cited sources, expert quotes, and comprehensive coverage of the topic. It should be 2,000–4,000 words for pillar topics. This is what gets cited by AI systems in informational query responses.
Not a reading of the blog post, a structured explanation of the same topic in video-friendly format. Upload to YouTube with full chapters, a keyword-rich description, and a transcript. This is what gets cited when AI systems recommend video content for the same query.
These seed LinkedIn, Twitter/X, and create the social signal layer that AI systems are beginning to weight.
Medium rewards first-person, cited, practitioner-voiced content. It is a different audience and tone from your brand blog, but it contributes a peer-authority signal that AI systems are tracking.
Not as promotion, as genuinely useful participation. Answer questions using data from your core asset. The Reddit thread that cites your research is more valuable for AI visibility than a paid link from a directory.
The declining brand count in AI responses is not a temporary fluctuation. It is the natural consequence of AI systems maturing and becoming more discriminating. The response is not to get more links or produce more volume, it is to own fewer topics more completely.
Topical authority in the AI visibility context means having comprehensive, interconnected, frequently updated content coverage of a topic cluster, not just one good blog post. It means having the pillar article, the supporting cluster posts, the video content, the glossary pages, and the case studies all pointing at each other and all signalling freshness simultaneously. That is what keeps you in the citation pool as it shrinks.
1. Map your topic clusters and define your prompt set. Before you can track AI visibility, you need to know which prompts matter. Group your target queries into 3–6 thematic clusters, as done in this banking analysis. Aim for 40–60 prompts that represent the real questions your buyers are asking AI systems.
1 Comment
Carolyn Holzman
1 weekInsightful study. We’re also seeing that domains with a strong central topical signal and strategic inner linking are getting considered to be cited in the overlay in search and other AI. In B2B enterprises, we’re encouraging customers to organize their blog content in a hub and spoke format as well. Thanks for sharing these findings.