The AI Already Read Your Website. The Question Is Whether It Comes Back.

Retrieval, in plain words

When somebody asks ChatGPT who they should call for a service in their town, the assistant does not answer from memory alone. It goes looking. It checks what it has indexed, picks a short list of pages that might answer the question, and fetches those pages in that moment, seconds before it replies. Retrieval is the name for that fetching step.

The companies describe this themselves. OpenAI documents a visitor called ChatGPT-User, which may visit a web page when a user asks ChatGPT a question, and states plainly that it is not used for crawling the web on its own schedule. Perplexity documents the same arrangement: its Perplexity-User visitor might visit a page to help provide an accurate answer, and include a link to that page in its response.

This is why retrieval matters more than any other number in this story. In a study by Vercel and the ora.ai research lab covering more than a thousand real agent visits, about 82 percent of the answers they could trace were grounded in a page the agent had actually fetched. If your page was not fetched for that question, your business was almost certainly not part of that answer.

OpenAI: Overview of OpenAI crawlers (GPTBot, OAI-SearchBot, ChatGPT-User)

Perplexity: Perplexity crawlers (PerplexityBot and Perplexity-User)

Vercel: Make your site readable by AI agents (study of 1,033 agent runs with ora.ai)

Who comes back, and how often

Here is what the second kind of visit looked like over 30 days on two websites I run, counted October 30, 2026 from the hosting dashboard that now reports this traffic separately.

Luxe Window Works, my window treatment business, received about 5,900 crawler visits and about 3,100 assistant fetches. Read that second number slowly. It means that roughly one hundred times a day, an AI assistant went to that website's pages because somebody, somewhere, was asking a question it might answer. Real Estate With Shirin, the site I built for my wife's real estate business, received about 3,800 crawler visits and about 1,700 assistant fetches, the same pattern at a smaller scale.

The fetching is not done by one gatekeeper. On the window treatment site, the largest share came from OpenAI's search visitor, about 1,400 fetches, followed by ChatGPT's agent at about 900 and Perplexity at about 540, with smaller numbers from DuckDuckGo's assistant, Claude and others. About four of every five assistant visits came from OpenAI's systems. One honest nuance: the hosting platform counts OpenAI's search visitor among the assistants rather than the crawlers, because its visits feed live answers instead of model training. It sits between the two kinds, re-reading pages so answers stay current. The crawlers, meanwhile, are a crowd of their own. The single busiest reader of that site all month was Meta's crawler, at about 2,500 visits, reading for Meta's AI systems.

Vercel: Make your site readable by AI agents

OpenAI: Overview of OpenAI crawlers

One visit is never the whole story

A retrieval rarely reads a single page and stops. In the Vercel and ora.ai study, when agents reached a site's answer pages, they fetched an average of 3.4 pages on that site. And between 86 and 97 percent of those fetches arrived by following links, not by guessing addresses. The homepage was the first step in 92 percent of the visits they traced.

Translated into plain business terms: the assistant usually comes in through your front door and follows your own links to the page that answers the question. A page that nothing links to is a page that mostly does not get read, however good it is. Your site's internal signposts decide what a retrieval can find.

Retrieval also has no memory you can bank. Being fetched for yesterday's question does nothing for today's. Every new question is a fresh decision, made from the index, the corroboration and the pages as they stand right now. That is why a single spot check means so little. Ask once and you have weather. The honest measure is a rate: the same kinds of questions, asked over days, counting how often your business is in the answer.

Vercel: Make your site readable by AI agents (study of 1,033 agent runs with ora.ai)

Why some sites get fetched and others get filed

I run a control case, whether I meant to or not. Starry Van is a buyer's guide site I launched in mid-October, built by the same hands, with the same care and genuinely useful content. In its first weeks it received 375 crawler visits and exactly one assistant fetch. The crawlers filed it. The assistants, answering live questions, essentially never came back.

The difference is not the quality of the site. It is that nothing outside the site vouches for it yet: no business profiles, no directory listings, no reviews, no history. The two established sites are the opposite, the same business facts stated the same way on the website, the Google profile, the directories and the review platforms, all agreeing. Crawling asks whether a machine can read you. Retrieval asks whether it can trust what it read enough to build an answer on it.

Two mechanical failures also end the story before trust even comes up, and the study watched both happen. If the answer on a page is assembled by scripts in the visitor's browser, a fetching assistant that does not run those scripts receives an empty shell. And if the server refuses the fetch, there is nothing to read at all. A page served as finished, readable text can be used by every fetcher that arrives. On the window treatment site, the platform reported no server errors for the month: every page the machines asked for, they got.

Finally, the answer has to actually be on a page. In the study, nearly half of the grounded answers traced back to pages written to answer questions directly. For a local business, the equivalent is the page that answers a buying question in plain words: what it costs, how the options compare, what fits a particular need. Assistants fetch pages that settle questions. Brochures do not settle questions.

One limit, stated plainly: nobody publishes the formula an assistant uses to choose, and you should distrust anyone who claims to sell it. What the companies document is how their visitors behave. What our own numbers show is which sites those visitors return to. The pattern across both is consistent enough to act on.

Vercel: Make your site readable by AI agents (study of 1,033 agent runs with ora.ai)

OpenAI: Overview of OpenAI crawlers

Perplexity: Perplexity crawlers

What to do with this

Start by measuring the split on your own site. If your hosting reports AI traffic, look for the two numbers side by side: crawls and assistant fetches. If fetches are near zero while crawls are healthy, you are being filed, not used, and the gap between those numbers is the work.

Then make the fetching easy. The pages that answer your customers' buying questions should be finished pages any visitor can read, linked from your homepage and from each other, so an assistant arriving at the front door can reach them by following your own signposts.

Then make the answer safe to use. State your business facts once, clearly, and let independent places confirm them: your Google profile, the directories, the review platforms, all telling the same story. That agreement is what turns a read into a return visit.

That is the work Kodecite does. A website built to be read by people and fetched by their assistants, and a business record clear enough that when the assistant goes looking, it comes back to you.