Half of today's executives start their research on AI platforms. We wanted to know what they were seeing. What we found instead was that it depends on when they ask.
Across 31,500 runs, that held whether the question was asked again seconds later or weeks later. On ChatGPT, Claude and Gemini, an answer a month later differed from the original by as much as an answer asked again in the same session.
How often are you running with the first thing AI shares? Is AI omitting the right answer to your question? In our attempts to streamline research and buying, what could we be losing?
See for yourself. Select a C‑Suite persona below to see the same prompt run twice, and what changes between the two responses.
What metrics should I prioritize to prove the ROI of AI-generated content?
Our research found that 50.4% of identical prompts from C‑Suite personas produced a substantively different response when asked again in the same session, and 57% did when asked again over four weeks. If you’re a CTO evaluating SaaS platforms, the answer you get is close to a coin flip away from changing, whether you ask again in a few seconds or a few weeks. But you’re probably not repeating prompts. So how much of your research is dictated by chance? And are you comfortable with that?
Measure AI visibility as a rate. A screenshot of one answer is one draw from a range of possible answers. On ChatGPT, Claude and Gemini, that range held steady for a month, so tracking how often your brand appears across many asks tells you more than any single check.
Across the C‑Suite, there’s about a 54% chance the answer you’re reading would change if you asked again. If you’re a CHRO, that drops to 39%. If you’re a Chief Medical Officer, it climbs to 70%.
Your C-Suite audience isn't a monolith; their experiences with AI platforms vary widely.
average variance across all nine personas and four platforms, meaning more than half of identical prompts returned a meaningfully different answer when asked again.
Its answers changed about 10% of the time, compared with 65–73% on Claude, Gemini and Perplexity.
If a team can focus on only one platform, ChatGPT is where its work holds longest. That cuts both ways, since a misperception about a brand is just as likely to stick there.
Longest answers of the four, averaging 7,100 characters, nearly 3× Claude.
Most consistent opening of the four. Avg 2,871 chars per response.
The most heavily formatted platform, with bullets and bold in nearly every response.
Structurally distinct: 100% bold, 98% bullet lists, inline numbered citations. Shortest avg response at 2,362 chars.
While AI answers are unpredictable, the way platforms organize their answers is rigid, and it held across four weeks of re‑runs even as the answers inside it changed. Gemini and Claude default to strict section headers, ChatGPT writes long-form essays, and Perplexity delivers rapid-fire bullet points with footnotes. Structure tells you what each platform looks for online.
AI engines don’t read the web like humans; they break articles into small, bite-sized fragments to build structured summaries. If your key claims and data are buried inside long, winding paragraphs, the engine’s extraction system chops them up and may miss the point. But if you organize your content into clean, self-contained sections with clear subheads and direct facts, you make it far easier for any AI model to lift, synthesize, and cite your brand.
Peer-reviewed research in Generative Engine Optimization (GEO) confirms why structural engineering directly impacts brand presence. In the foundational study published at KDD 2024 by researchers from Princeton University, Georgia Tech, the Allen Institute for AI, and IIT Delhi (Aggarwal et al., GEO: Generative Engine Optimization), researchers demonstrated that deliberate content-optimization strategies can boost visibility inside generative search engines by up to 40%. The highest-performing levers were not traditional keyword density, but architectural extractability: fluency optimization (+28% visibility lift) and the inclusion of standalone statistics and authoritative quotation additions (+30% to +40% visibility lift).
In modern Retrieval-Augmented Generation (RAG) pipelines, automated chunking algorithms fragment narrative documents into isolated semantic windows (typically 300–800 tokens). When a brand’s claims are dispersed across narrative prose, that semantic context is severed during retrieval. When content is built as modular, answer-first units with descriptive H2/H3 headers, self-contained metrics, and verified attribution, it survives chunking intact and slots into the comparative matrices and cited footnotes these platforms construct for executive buyers.
The 154,944 citations we collected came from 16,503 different domains, and it took 786 of them to account for half of all citations. Even LinkedIn, the top source, made up only 5%. No single placement or channel owns the answer.
On AI that searches the live web for every query, more than half of cited sources were replaced within a month. Government sources, professional associations, analysts and trade press kept roughly two-thirds of their spots after four weeks.
Medium and Reddit, by contrast, were cited about as often as NIH at the start. NIH kept 76% of its spots, while they kept only 21–27%.
On platforms that search the live web, the range of possible answers moves as sources turn over. Being cited once doesn’t mean you’ll be cited next month. Staying cited takes a steady supply of new material from sources AI already treats as credible, like earned coverage in trade press and analyst commentary, so live search keeps finding your brand when it looks.
LinkedIn appeared 8,243 times across 154,944 citations for C-Suite persona queries. It was the top source for 8 of 9 personas, every one except our Chief Medical/Clinical Officer, where PubMed leads. For PR and content teams, this is immediately actionable:
LinkedIn posts are direct inputs to what AI tells B2B buyers when they research a category.
LinkedIn appearances across 154,944 total citations for C‑Suite persona queries across the full study
| Persona | Primary source | Secondary source |
|---|---|---|
| CMOMicah | Gartner | |
| CISOSteve | Microsoft Learn | |
| CTO / CIOIda | CIO.com | |
| CFOPetra | Deloitte | |
| CDODenise | Medium | |
| Chief MedicalHenry | PubMed | AMA |
| CHROCandace | SHRM | |
| Global CommsCaroline | Forbes | |
| CROPeter | Gartner |
Press releases barely register, at under 2% of the citations we categorized, while LinkedIn and trade coverage show up throughout. So, if you want LLMs to notice that shiny new product or flashy acquisition, amplify it on social and in the trades where and when you can.
Subscribe to PAN’s newsletter for AI citation data and analysis, delivered monthly.
SubscribePAN’s Brand to Demand approach connects citation presence to pipeline. See what that looks like in practice.
See the approachPAN’s new brand-to-demand intelligence layer, dropping soon.
Comprehensive findings. Full data visualization. Get it in your inbox the day it publishes.
Join the waitlistSources are the five most-cited in responses to each persona’s prompts across the full study, drawn from 154,944 citations and grouped by organization. Percentages are each source’s share of that persona’s citations. Variance rates and badges are calculated from 31,500 runs (9 personas × 125 prompts × 4 platforms, asked again within the same session and at four points over four weeks) using Dice word-set similarity at a 0.50 threshold.
The original C-Suite Signals report asked ChatGPT where executives get their information. The AI Citation Index extends that question to four platforms and measures how consistently they answer. We asked each prompt three times in a row on each platform, then asked the same 1,125 prompts again at four points over four weeks, for 31,500 runs in total. Timing made little difference. On three of the four platforms, an answer weeks later varied about as much as one seconds later. The platform that searches the live web for every query changed more as its sources turned over.
Each of 125 prompts per persona, across nine personas, was submitted three times in a row on ChatGPT, Claude, Gemini and Perplexity, then again at four weekly collection points between late August and late September. Every repeat answer was compared with the first answer to the same prompt on the same platform. Variance is measured with Dice word‑set similarity at a 0.50 threshold, and the rates on this page combine every comparison.
Citation data comes from Perplexity, the only platform in the study that cites live web sources in every answer; the other three were not prompted to return sources to mimic average user prompt behavior. The 154,944 citations from both analyses were source-coded by domain and persona. Source types cover the 400 most-cited domains plus all .gov and .edu sites, which together account for 44% of citations. Source retention is measured per prompt, by domain, checking whether a domain cited for a question in week one was cited for the same question in later weeks.
15,000+ ChatGPT-cited links across nine personas. It’s the source hierarchy the four‑platform Citation Index builds on.