How long does it take for AI to start citing a new website?

In Campaign Salience's instrumented launch of wagesight.com, the first time an AI assistant fetched a page to answer a user's question came 9 days 20 hours after deployment. ChatGPT cited that page in an answer six days later: 16 days from deploy to being quoted.

Three clocks

1 - Cloudflare - our server - recorded the moment the site was deployed. 2 - Our web log recorded the moment an AI assistant first came to read a page on behalf of a user; the log keeps every request with its user agent, network origin and path, and sorts them into training crawls, index crawls, live retrievals, scanners and spoofers. 3 - A panel of forty test questions, put to five assistants over nine days in July.

Know your 'bots

In the first three months it existed, wagesight.com recorded 92,116 visits from robots. But most of the visits were "spoofs". The internet is awash with crawlers that impersonate "legitimate" machines. Just over 85,000 of our visitors were machines impersonating legitimate crawl, indexing and retrieval bots. However, the remaining 7,038 robot visits are all verifiable crawls, indexing and user retrieval events, despatched by reputable AI outfits, from Anthropic, to DuckDuck Go and Google Gemini.

What happened

WageSight went live at 11:55 on 5 July 2026: 300+ place pages, each one answering a single question about pay, employment and the cost of living in a single town from latest HMRC payroll figures, with the same numbers available as CSV and JSON and a sitemap submitted the same day.

There was no fanfare, no brand awareness, no launch, pitching or backlinks, just pure data-led comms about a niche and poorly served aspect of the hyper local economy. We built and deployed a suite of website instruments to record robot visits eight days later, on 13 July. And then... we waited.

We didn't wait long. At 07:22:30 on 15 July, 55 hours after the log started, a request arrived from ChatGPT-User, the agent OpenAI sends out when someone's question makes ChatGPT go and read a live page. It asked for /places/rossendale and got it. Somebody, somewhere, had asked ChatGPT about pay in Rossendale, a small borough in the Lancashire hills, and ChatGPT had come to a ten-day-old website for the answer. Nine minutes later it came back, this time looking for data about Barrow-in-Furness.

Six days later we probed Chat GPT's API backend to confirm the situation. On 21 July, ChatGPT answered a question about pay in Rossendale and cited wagesight.com as a source, in the company of ONS, gov.uk and nomis and nothing else.

Why so fast

Because the site was built to be read by machines first and people second. Each page asks one question and answers it. Every figure carries its date. The raw data sits beside the prose rather than behind a PDF. Google had 208 of the site's URLs indexed inside three weeks, with no backlinks and no history, and the assistants were not far behind. Nothing here was promoted. The assistant came because a user asked a question the page answered, and the page was legible to the thing doing the reading.

What we are not saying

That 15 July was the first time. The log wasn't running for the first eight days and we'd be surprised if nothing read the site in that time; we just can't show it. That six days is how long fetch-to-cite takes: it's three observations from one week, and later retrievals converted on other schedules or didn't convert at all.

We are not saying we are the quickest. The consensus in GEO World is "2 to 8 weeks" for a well-matched page to land in an AI answer. And - at the end of the day - if no-one's asking AI questions relevant to the world your organisation communicates about, then it's unlikely you'll ever surface in an AI answer!

Why it matters that we understand AI's mechanics

AI can sometimes feel overwhelming. Like the other big issues of our age, such as climate change, AI just feels too big, too hard to understand and get a grip of. Too often the mechanisms of AI feel like a black box that we cannot get our heads around. But in this one small experiment we were able to trace the trajectory of a fact we placed on a website, its retrieval in response to a live question placed by a human and finally, it's use inside an AI answer. That single fact about pay in Rossendale was retrived on the 15th and cited on the 21st. The same thing happened twice more that week: Barrow, fetched 15 July, cited 21 July; Bath, fetched 18 July, cited 24 July. Six days each time. A fourth page, King's Lynn, was fetched on the 17th and not cited in the window, which is worth saying, because three out of four is a pattern and four out of four would be a fairy tale!

Updated 13 October 2026.