Early results, measured October 2026
New articles, cited as often as a back catalogue that took years to build.
In 2026 I wrote new articles for two B2B SaaS companies. In October I asked ChatGPT the question each article answers and checked whether it cited the article. At both companies the new articles were cited at least as often as the company's older articles, most of them years old.
Company A B2B SaaS
Company B B2B SaaS
Company A, by age
The dot is the share of articles ChatGPT cited. The bar is the range the true share likely sits in, given how few articles each group has. Where bars overlap, the groups can't be told apart.
3–34%
11–39%
6–41%
5–45%
11–28%
10–37%
Two articles published in September were already being cited within weeks. Company A has the larger sample, so its bars are the narrower ones.
Company B, checked three times
Each question went to ChatGPT three times, and an article counts as cited when at least two of the three answers cited it. The result was the same as the first answer alone.
10–57%
4–25%
11–38%
5–16%
Bars past 50% are cut at the edge. Company B's new group is small, so its range is wide.
What decided it: the topic
At both companies, the biggest difference wasn't age or format. It was whether the question sat close to what the company sells.
21–57%
4–20%
12–38%
0–22%
Core and adjacent were labelled before the results were read. Company A counts its new articles; Company B counts new and older together.
- Adjacent questions go to official sources. When the question was a general one, such as what a registry is or how a regulation works, ChatGPT cited the government body or the organisation that runs it. A company blog rarely wins those, however good the article.
- Some keywords are lost before the article is read. A short, broad question can mean something else entirely to ChatGPT, which then answers it from a different industry's pages.
- Format didn't separate winners. FAQs, tables and answer-first openings were about as common in articles that weren't cited as in those that were.
How it was measured
- One question per article. Each article has the question it was written to answer. That exact question went to ChatGPT with web search on.
- One model, one month. Every article was checked on the same OpenAI model (gpt-5.6-luna) in October 2026, so new and older articles compare fairly.
- Cited means linked. An article counts as cited when ChatGPT's answer lists its page among its sources. Every cited page was recorded.
- The older articles. Company A: 30 existing keyword articles picked at random, published between November 2017 and January 2026. Company B: every older article written for a search keyword, published between 2019 and June 2026; news, press coverage, webinars and comparisons naming the company were left out. All were published before my first article at that company.
- Repeated answers. Company A's questions were asked once each, Company B's three times each.
What it shows, and what it doesn't
- These are early results. The newest articles are weeks old and the oldest are four months old. A fair verdict on any one company needs more time, so each will get its own case study once its articles have aged.
- It shows that new articles don't have to wait years to be cited. At two companies in different industries, within weeks to four months of publishing, they reached or passed the rate of articles that had years to collect links and authority.
- It doesn't show that new articles beat old ones. At Company A, 18% and 20% are the same result. At Company B, 27% against 11% leans that way, but 11 new articles are too few to call it a win.
- AI answers change from run to run. Any single article could flip either way on another day. The group shares are steadier than any one result, and at Company B three answers per question gave the same picture as one.
- The model matters. An earlier check on a different OpenAI model cited these articles less often. The comparison holds within one model, not across models.
- Citations aren't rankings or traffic. This measures whether an AI answer names the page as a source, nothing more.
What happens next
I'll ask the same questions about the same articles every month and update this page with the date, adding articles as they go live. Once each company's articles have had time to age, it gets a case study of its own. Whether new articles keep climbing as they age is the question that answers. Neither company is named.