How long does it take to get cited by an AI assistant?
On our two sites, both new domains with no backlinks (links from other websites), the first answers served (an assistant fetching a page to reply to someone) came inside a fortnight; on one, most arrived in the four days after ChatGPT's search crawler took 35 of its pages (server logs, 29 September to 13 October 2026). A firm with three years of history was sent far more visitors by the assistants. That is what we saw, not a promise.
QuotableBoth of our new sites were fetched by an AI assistant to answer a real question inside a fortnight of going live (server logs, September and October 2026).
What we saw on two new sites
An "answer served" is one line in a server access log: a person asked an assistant something, and it fetched our page to write its reply. ChatGPT does this as ChatGPT-User, the user agent it sends when it fetches a page while answering someone. Nobody has to click. The fetch is the citation.
The first of our two sites is a UK address-tracing service we launched on 29 September 2026. In the fifteen days to 13 October it served 27 answers: ChatGPT 24, DuckDuckGo AI 2, Perplexity 1 (server logs of our address-tracing site). The shape matters more than the total. Six answers came in the first eleven days, three of them from ChatGPT. On 9 October OAI-SearchBot, the crawler that builds ChatGPT's own search index, took 35 pages. The next four days brought 21 ChatGPT answers: five, five, nine, and two by midday on the 13th.
Investigation Bureau, a UK private investigation site, went live on 5 October 2026. On 8 October, three days in, ChatGPT fetched its pages six times while answering. Two more answers came on 11 October: eight in eight days, across twelve pages, one answer drawing on five of them at once (investigationbureau.uk server logs, 13 October 2026).
The order of events: crawled, indexed, fetched
Three things happen in order; each leaves a different line in the log.
- Crawled. An index builder fetches your pages so the assistant has something to search. The table below names them. Googlebot, for instance, had fetched 58 pages of our address-tracing site by 13 October 2026 (server logs).
- Indexed. The page sits in an index the assistant searches at answer time. Crawled is not indexed, and indexed is not ranked: Google had our address-tracing site at an average position of 34 across 543 impressions in the 28 days to 8 October 2026 (Search Console). You cannot see inside OpenAI's index, only the crawler that feeds it.
- Fetched while answering. A person asks, the assistant searches, picks a page and fetches it live. That hit is the citation. A click-through, where the person then visits your site, is rarer again: none on our address-tracing site in fifteen days, one on Investigation Bureau in eight (server logs, 13 October 2026). The assistant answered in place; the person took away the name.
| Index builder (user agent) | What it feeds |
|---|---|
| OAI-SearchBot | ChatGPT search |
| Claude-SearchBot | Claude search |
| PerplexityBot | Perplexity |
| Bingbot | Bing, Copilot and DuckDuckGo |
| Googlebot | Google, AI Overviews and Gemini |
The first two stages are where a site stalls, and the log shows it: none of the index builders in the table for weeks on end, or one that comes once and does not return. A site in that state cannot be fetched while answering, because there is nothing for the assistant to search, so every question it would have answered goes to a page that was indexed. The third stage depends on what your page says in its first paragraph and on who else answers the same question. On a plan the Audit names the stage your site is stuck at and writes the exact changes for your developer; which AI crawlers to allow in robots.txt lists the user agents to look for.
Why Bing matters
Each assistant answers from a different index. Copilot reads Bing. DuckDuckGo AI reads Bing. ChatGPT searches its own index, built by OAI-SearchBot, and fetches pages live as ChatGPT-User. Google's AI Overviews and Gemini read Google. Perplexity runs its own crawler. A site that is in Google but thin in Bing is missing from two assistants outright, and each of the others has to find it with its own crawler.
Bing is also the index you can see into. Bing Webmaster Tools, Microsoft's free console for site owners, shows what it has indexed, so a site that is thin there can be caught early; Google comes when it comes. Our two sites show the gap. By 13 October 2026 Bingbot had fetched 105 pages of our address-tracing site but only four Investigation Bureau pages. OAI-SearchBot had taken 35 pages from our address-tracing site and only Investigation Bureau's home page (server logs). The younger site had served fewer answers, eight against 27. Our reading, and only a reading, is that this gap is part of what holds it back. A site in that position is losing answers it never sees; on a plan the Audit sets out what is holding Bing back and writes the changes for your developer.
Why a firm with three years of history was sent far more visitors
We also read the analytics of a London private investigation firm, as aggregates only; we do not name it. An analytics tool only sees people who click, so its figures are visitors sent by an assistant, a stricter event than an answer served.
Between 24 July and 7 October 2026 ChatGPT sent it 2 visitors in July (from the 24th), 14 in August, 43 in September and 5 in the first seven days of October; Copilot sent 4. That is 68 assistant-referred visitors. Eight tapped the phone number, three sent the form, one became a £3,300 instruction (the firm's own analytics, 7 October 2026).
Two reasons it did better, in our reading:
- History. Three years of trading almost certainly means the firm's name is in the models' training data, the text the models learned from before anyone asked them anything. The assistant probably knows it before it searches. A domain registered last month can only be found by search.
- Trust signals. Named senior staff, fixed quotes, ICO registration, 23 reviews, a London address. Research agrees: brands with 80 or more answered reviews appeared in 75% of AI answers, against 1% with none (Seer Interactive for Trustpilot, March 2026, vendor-commissioned).
Its history did not do all the work. An OSINT service page (open-source intelligence: research from public records and the open web) published on 6 September 2026 became the second most cited page on the site within a month, behind only the home page, with 10 assistant-referred landings (its analytics, 7 October 2026). September's jump came in the month that page, a technology page and llms.txt (a plain-text file at the root of a site that tells assistants who you are and which page answers what) were published, and after ChatGPT's search crawler took the site in. Fresh pages still counted on an old site.
History is a head start, not a wall. The full account is in the case file.
What to expect, month by month
The table sets out what we have seen on our own sites and on the established firm, and what dated research says. It is not a forecast for your site, and not a promise of a citation, a position or an enquiry. The right-hand column is what each stage looks like when it has not happened; on a plan the Audit reads your log against it and writes the exact changes for your developer.
| Period | What we saw, or what research found | If it is not happening |
|---|---|---|
| Week 1 | A first answer served can come early. Our address-tracing site served two on its first day, 29 September 2026; Investigation Bureau's first six came three days after going live, on 8 October 2026 (server logs). | No index builder in the log by the end of the week. Nothing can be fetched while answering until something has been crawled, so the clock has not started. |
| Weeks 2 to 4 | On our address-tracing site, ChatGPT's search crawler came on day eleven, 9 October 2026, and the answers followed it, as above (server logs). Google impressions had started, far down the results (Search Console, 28 days to 8 October 2026). | Read the access log for ChatGPT-User. Index builders present but no answers served, or the same one page taken every time: the assistants can find you and are choosing someone else, question by question. |
| Months 2 to 3 | No data from our own sites yet; the older is fifteen days old. The only multi-month record we hold is the established firm's, above, and its three years of history make it a poor guide to a new site's second month. | A log still quiet in month two is a site the indexes hold but the assistants do not use. Each week it stays that way is a week of questions answered with another firm's name. |
| Beyond | 84% of links cited by assistants point at third-party media (Muck Rack, 25 million cited links, May 2026). Brand mentions correlate with AI visibility about three times more strongly than backlinks (Ahrefs, 75,000 brands, December 2025). | A site that nobody else writes about is one the assistants rarely name, however good its own pages. The cost is the enquiry that went to the firm they did name. |
Two things do not shorten the wait. Schema markup (structured data: code on the page that labels your organisation, articles and questions for machines) and llms.txt do not move citations on their own in controlled tests (Ahrefs, May 2026; Google, June 2026). And a page that only exists once JavaScript runs never starts the clock: no AI crawler executes JavaScript (Vercel and MERJ crawler study, 2025; our own logs, 2026). See do AI assistants read JavaScript websites.
How to check your own site
Run the free check: it fetches your robots.txt, llms.txt and sitemap.xml, and your home page as a browser and as ChatGPT-User, GPTBot (the crawler OpenAI uses to collect training text), ClaudeBot and PerplexityBot, so a block at your firewall shows.
Then ask your host or developer for the server access log and search it for these user agents. ChatGPT-User, Perplexity-User, Claude-User, Meta-ExternalFetcher, DuckAssistBot and MistralAI-User are answers served. OAI-SearchBot, Claude-SearchBot, PerplexityBot and Bingbot are index builders. Which of them appear tells you how far you have got: no index builders yet, index builders but no answers, or answers served. How do I know if ChatGPT has read my website shows what one line looks like. If you would rather have the counting done each Monday, that is how we work.