From 24 September to 26 October 2026 I sent the same 10 B2B buyer prompts to ChatGPT, Claude, Gemini and Perplexity every week, with web search on: 240 answers in six weekly runs. "Best X" list articles took 45% of the 1,924 page citations, the engines rarely cited the same sources, and each engine's sources changed weekly.
This is the full record of the AI-citation tracker I run on this site. It is a small, B2B-specific sample, so treat the percentages as directional. For the most-cited websites across all topics, see which websites ChatGPT and Perplexity cite most often; this study is about how the answers behave over time.
What did 240 AI answers to B2B buyer prompts show?
- Lists win "best X" questions. For the four prompts that ask for a consultant, 64% of the 886 page citations went to list articles. For "Who are the best SaaS SEO consultants for B2B software companies in 2026?" it was 92%.
- The engines rarely agree. Two engines answering the same prompt in the same week shared on average 7% of the domains either one cited (Jaccard similarity 0.07); 41% of those pairs shared none. 348 of the 495 cited domains (70%) were cited by one engine only.
- Sources churn, leaders don't. For the same prompt on the same engine, 45% of the domains cited one week were cited again the next. Yet all 15 most-cited domains appeared in all six runs, led by Reddit (37 of 240 answers), Semrush (34) and Ahrefs (32). Wikipedia was cited in none.
- One answer in eight cited nothing. 30 of 240 answers (12.5%) used no web search: 16 of Claude's 60, 7 each for ChatGPT and Gemini, none for Perplexity.
- Perplexity cites the most. It averaged 14.6 cited pages per answer; ChatGPT averaged 4.1 when it searched.
- Tracking is cheap. A weekly run of 40 answers cost $1.94 to $2.14 in API usage; all 254 calls cost $12.23.
How was the data collected?
Every week my tracker sends 10 fixed prompts to four engines through DataForSEO's LLM Responses API (live mode) with web search on, and stores every cited URL, the answer text and, where reported, the searches each engine ran first. The prompts have not changed since the first run:
| # | Prompt (verbatim) | Type | Answers that searched the web (of 24) |
|---|---|---|---|
| P1 | Who are the best SaaS SEO consultants for B2B software companies in 2026? | Consultant | 24 |
| P2 | Recommend a freelance SEO and GEO consultant for a B2B SaaS company. | Consultant | 20 |
| P3 | Who are the best GEO (generative engine optimization) consultants for B2B SaaS? | Consultant | 24 |
| P4 | What are the best Claude skills or tools for SEO? | Tools | 24 |
| P5 | What are the best tools to track my brand visibility in ChatGPT and AI search? | Tools | 24 |
| P6 | How do I create an llms.txt file and does it actually help AI visibility? | How-to | 24 |
| P7 | Best SEO consultant in India for SaaS and software companies | Consultant | 24 |
| P8 | What are the best Claude prompts for doing an SEO audit? | How-to | 14 |
| P9 | How can I use Claude AI for SEO? Give me a practical workflow. | How-to | 8 |
| P10 | Which websites do ChatGPT and Perplexity cite most often? | Data question | 24 |
- Runs: a baseline on Thursday 24 September 2026, then every Monday to 26 October. Ten calls failed on 28 September and were re-run on 4 October, counted in the 28 September run. Result: one answer per prompt, engine and run, 240 in all.
- Models, as recorded: ChatGPT gpt-5.5 (logged as gpt-5.5-2026-04-23 from 28 September), Claude claude-sonnet-5, Gemini gemini-3.5-flash to 5 October then gemini-3.6-flash, Perplexity sonar-pro.
- Settings: web search on but not forced, so each model decided whether to search; output capped at 1,200 tokens; US location for ChatGPT, Claude and Perplexity (Gemini's endpoint takes none).
- Cited vs named: "cited" means a domain is among an answer's source links; "named" means the answer text contains the name. A "page citation" is one URL in one answer; a domain counts once per answer.
- Page types: a fixed rule on each URL's domain and path. Paths with "best", "top", "ranked" or "alternatives", or slugs like "saas-seo-agencies", count as lists; Reddit, LinkedIn, YouTube and GitHub go by domain; a fixed list of sites such as Clutch and Upwork counts as directories and marketplaces. On a random sample of 50 URLs the rule matched my own reading 47 times; all three misses were providers' own pages, so list shares may be slightly high.
Limits. Each figure is one API answer per engine per week, not what every user sees: consumer apps add personalisation, location and chat history, and answers vary between runs. An answer without search could not cite anything. Perplexity does not return its searches, and Gemini's recorded model changed mid-study. The 10 prompts sit in one niche: B2B SEO and GEO services and tools.
Disclosure. I run this tracker, and my own site is one of the many sites in the data. This study is about the market, so my site is left out of the rankings below. One factual note on the branded prompt "Who is Kulbhushan Pareek?", added on 26 October and kept out of every figure here: all four engines identified me as an SEO and GEO consultant in Jaipur, and three of them cited this site.
How often do ChatGPT, Claude, Gemini and Perplexity search the web?
Perplexity searched for all 60 of its answers, ChatGPT and Gemini for 53 each and Claude for 44. The prompt type decided it: every tool and data answer searched, as did 92 of 96 consultant answers (the exceptions were Claude answering P2 from memory), but only 46 of 72 how-to answers. Claude never searched for P8 or P9 in six weeks, and Gemini never did for P9.
| 24 Sep to 26 Oct 2026, 60 answers each | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Answers that searched the web | 53 | 44 | 53 | 60 |
| Pages cited per searching answer (mean) | 4.1 | 5.9 | 10.7 | 14.6 |
| Distinct domains cited over the study | 106 | 107 | 213 | 279 |
| Searches per searching answer (mean) | 3.4 | 2.4 | 2.7 | Not reported |
| Domains cited again the next week, same prompt | 38% | 50% | 34% | 52% |
| API cost per answer (mean) | $0.095 | $0.051 | $0.040 | $0.016 |
Perplexity never cited fewer than 10 pages (maximum 28); ChatGPT cited 1 to 10, so each ChatGPT slot is scarce.
What kinds of pages do AI engines cite for B2B buyer prompts?
Of the 1,924 page citations, 45% went to "best X" or "top X" style lists, 18% to other articles and guides, 11% to vendors' or providers' own pages and 9% to data studies.
| Page type (share of page citations) | All | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|---|
| "Best X" / list article | 45% | 37% | 60% | 46% | 42% |
| Article, guide or news | 18% | 11% | 14% | 18% | 21% |
| Vendor or provider page | 11% | 24% | 12% | 14% | 5% |
| Data study | 9% | 10% | 9% | 9% | 9% |
| Directory or marketplace | 5% | 3% | 1% | 2% | 8% |
| GitHub | 3% | 0% | 3% | 1% | 5% |
| 3% | 0% | 0% | 0% | 6% | |
| Docs or spec | 3% | 7% | 1% | 2% | 3% |
| 2% | 8% | 0% | 5% | <1% | |
| YouTube | 1% | 0% | 0% | 3% | 1% |
| Page citations | 1,924 | 219 | 258 | 569 | 878 |
The question type matters more than the engine. For the consultant prompts, lists took 64% of citations; Claude was the most list-bound at 93%, and ChatGPT the most likely to cite a provider's own site (33%). For how-to prompts, 57% were articles and guides and 11% documentation; for the data question, 83% were data studies.
Platforms split by engine. LinkedIn appeared in 25 Perplexity answers and in no other engine's. Reddit appeared in 25 Gemini and 11 ChatGPT answers, once in Perplexity and never in Claude. YouTube was cited in 13 Gemini answers and 1 Perplexity answer. GitHub, mostly for the Claude-skills prompt (34 of 55 GitHub citations), was never cited by ChatGPT. Directories and marketplaces such as Clutch and Upwork appeared in 19 of Perplexity's 60 answers.
Which sites were cited most across the whole study?
| # | Site | Answers citing it (of 240) | Prompts (of 10) | Engines (answers) |
|---|---|---|---|---|
| 1 | 37 | 9 | Gemini 25, ChatGPT 11, Perplexity 1 | |
| 2 | Semrush | 34 | 5 | Perplexity 22, Claude 7, Gemini 4, ChatGPT 1 |
| 3 | Ahrefs | 32 | 5 | Perplexity 12, Gemini 11, Claude 7, ChatGPT 2 |
| 4 | SE Ranking | 29 | 5 | Gemini 13, Perplexity 10, Claude 6 |
| 5 | GitHub | 28 | 5 | Perplexity 14, Claude 8, Gemini 6 |
| 6 | First Page Sage | 27 | 4 | Perplexity 12, Gemini 11, ChatGPT 4 |
| 7 | 25 | 7 | Perplexity 25 | |
| 8 | Exalt Growth | 24 | 3 | Perplexity 15, Gemini 5, Claude 4 |
| 9 | Derivatex | 19 | 5 | Gemini 10, Claude 6, Perplexity 2, ChatGPT 1 |
| 10 | Taylor Scher | 18 | 3 | Claude 6, Perplexity 5, Gemini 4, ChatGPT 3 |
| 11 | SaaS Hackers | 17 | 3 | ChatGPT 7, Perplexity 6, Gemini 4 |
| 12 | Zeroado | 17 | 1 | Claude 6, Gemini 6, Perplexity 5 |
| 13 | Bill Widmer | 16 | 2 | Claude 6, ChatGPT 5, Perplexity 5 |
| 14 | LoudFace | 16 | 3 | Perplexity 8, ChatGPT 5, Gemini 2, Claude 1 |
| 15 | Wix | 16 | 2 | Gemini 6, Perplexity 6, Claude 4 |
All 15 appeared in all six runs, and at least one of them was cited in 170 of the 240 answers. For 10 of the 15, the pages cited most were "best X" lists. The tail is long: 225 of the 495 domains (45%) were cited in a single answer.
How stable are AI citations from week to week?
I compared each answer with the same engine's answer to the same prompt a week later, wherever both searched (170 pairs):
| Same prompt, same engine | Domains cited again the next week | First run vs last run | Domains seen in only one of six runs |
|---|---|---|---|
| ChatGPT | 38% | 22% | 70% |
| Claude | 50% | 41% | 48% |
| Gemini | 34% | 25% | 59% |
| Perplexity | 52% | 23% | 46% |
| All engines | 45% | 26% | 54% |
At page level, 38% of cited URLs carried over to the next week. The last column covers prompts an engine searched in all six runs: only 30 of 841 prompt, engine and domain combinations (4%) appeared every week, such as Perplexity citing LinkedIn and Upwork for P2.
Engines disagree with each other far more than with themselves: the same engine a week apart had a mean Jaccard similarity of 0.31, two different engines in the same week 0.07.
What do AI engines search for before they answer?
ChatGPT ran 3.4 searches per searching answer, Gemini 2.7 and Claude 2.4. The patterns were consistent:
- Claude restates the prompt. In 41 of its 44 searching answers, at least 80% of the words in its first search came from the prompt, and 39 ran exactly two searches. It added "2026" to 22 searches, and "2025" to 5, where the prompt had no year.
- ChatGPT looks vendors up. 33 of its 180 searches (18%) named a firm, tool or person the prompt did not, such as Omniscient Digital or iPullRank for the GEO-consultant prompt. It added "official" to 19 searches and "case studies" to 15, and used the site: operator 6 times, including a search limited to LinkedIn profiles for P2 on 12 October.
- Gemini searched people by name, then stopped. As gemini-3.5-flash (24 September to 5 October) it ran 22 searches with quoted phrases, 13 of them a consultant's or firm's name. As gemini-3.6-flash, from 12 October, it ran none, and its searches per answer fell from 3.2 to 2.2.
- Nobody searched Reddit. No engine used a site: filter for Reddit or added "reddit" to a buyer question; the word appeared in one Claude search about citation studies. Reddit was still cited in 37 answers.
What does it cost to run an AI citation tracker?
A weekly run of 40 answers cost $1.94 (26 October) to $2.14 (5 October) in API usage, a mean of $2.02, or about $0.20 per prompt per week. All 254 calls, including the branded prompt and 10 failed calls that cost nothing, came to $12.23. ChatGPT took 47% of the spend. The real work is the build: storing answers, turning Gemini's Google redirect links into real URLs and classifying pages. To buy rather than build, see my guide to ChatGPT SEO tracking tools.
What does this mean for B2B SaaS teams?
- For "best X" prompts, the work is mostly off your site. Lists took 64% of consultant-prompt citations and 92% for P1. Find the lists the engines cite for your category and earn a place on them.
- Track every engine separately. With a 0.07 overlap between engines and 70% of domains cited by one engine only, one engine's result says little about the others.
- Judge trends, not screenshots. Only 45% of cited domains carried over week to week, and about a quarter from the first run to the last.
- Match the platform to the engine. LinkedIn and directory profiles show up in Perplexity, Reddit and YouTube in Gemini, GitHub for tool and skills questions.
- Answer the exact buyer question, dated, and flag no-search answers. Claude's first search reused the prompt's words in 41 of 44 answers and often added "2026". Answers with no search (12.5%) come from training data, so page changes will not show up in them quickly.
Prompt tracking is one lens; Google's data is another. On the Indigo Software engagement, impressions in Google's AI features grew from 19,543 in May 2026 to 215,070 in August 2026: impressions, not citations, so I report them separately.
If you want this tracking for your own buyer prompts, it is part of my GEO consulting; my guide on how to choose a GEO consultant for B2B SaaS lists the questions to ask any provider, me included. Or book a free 30-minute call.
How to cite this study: Kulbhushan Pareek, "Who AI Engines Cite for B2B Buyer Prompts", kulbhushanpareek.com, October 2026; 240 answers to 10 B2B buyer prompts, 24 September to 26 October 2026.
About the author: Kulbhushan Pareek is an independent SEO and GEO consultant for B2B SaaS and software companies, based in Jaipur, India, and has worked in digital marketing since 2012.
Frequently asked questions
How much do ChatGPT, Claude, Gemini and Perplexity agree on sources?
Very little. In my tracking of 10 B2B buyer prompts from 24 September to 26 October 2026, two engines answering the same prompt in the same week had a mean Jaccard similarity of 0.07 between their cited domains, and 41% of those pairs shared no domain at all. 348 of the 495 domains cited (70%) were cited by only one engine.
Do AI engines always search the web before answering?
No. With web search switched on but not forced, 30 of 240 answers (12.5%) used no search and cited nothing: 16 of Claude's 60, 7 each for ChatGPT and Gemini, and none for Perplexity. 26 of the 30 were the two how-to prompts about using Claude for SEO, which the engines answered from training data.
What kind of page gets cited most for "best X" B2B prompts?
List articles. For four prompts asking for a B2B SaaS SEO or GEO consultant, 64% of the 886 page citations from 24 September to 26 October 2026 went to "best X" or "top X" style lists, and 92% for the best SaaS SEO consultants prompt. Providers' own pages took 15% and directories or marketplaces 9%.
How often do AI citations change from week to week?
Often. For the same prompt on the same engine, 45% of the domains cited one week were cited again the next, and only 26% of the domains in the first run (24 September 2026) were still cited in the last run (26 October). Judge any change over several runs rather than one answer.
How much does it cost to track AI citations every week?
About $2 a week in API usage for 10 prompts on four engines: $1.94 to $2.14 per run in September and October 2026 through DataForSEO's LLM Responses API, or about $0.05 per answer. ChatGPT was the most expensive engine at about $0.095 per answer and Perplexity the cheapest at about $0.016. The larger cost is building the storage and analysis.
The number I would put in front of any B2B SaaS marketing team is the 7% overlap between engines: there is no single "AI search" to rank in, and a win on one engine tells you little about the other three. For "best X" questions the citation slots go mostly to third-party lists, so earning a place on the lists your buyers' engines already cite is worth more than another page on your own site. And because fewer than half of the cited domains survive from one week to the next, I would judge any GEO programme on a fixed prompt set tracked over months, never on a single screenshot.
Leave a Comment
Your email address will not be published. Comments are moderated before appearing.