AI Visits Your Website Every Day. Only One Kind of Visit Can Bring You a Customer.

Two kinds of AI visitors

The first kind arrives ahead of time, on its own schedule. These are crawlers, programs that read your pages and file them away. OpenAI documents two of its own. One, GPTBot, gathers content that may be used to train its models. Another, OAI-SearchBot, builds the index that ChatGPT draws on when it searches, and OpenAI says plainly that a site which blocks that crawler will not be shown in ChatGPT search answers.

The second kind arrives in the moment, because a person asked a question. When someone asks ChatGPT or Perplexity about businesses like yours, the assistant may visit a web page on the spot to answer accurately. OpenAI calls that visitor ChatGPT-User, and is explicit that it is not crawling: it acts because a user asked. Perplexity documents the same arrangement with its own pair of visitors, one that indexes and one that fetches at question time and may link your page in the answer.

Both kinds will appear in your traffic. Only the second kind is connected to a customer who is deciding something right now.

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

Perplexity: Perplexity crawlers (PerplexityBot and Perplexity-User)

What the numbers look like on real business websites

Hosting platforms have started reporting these visits separately. On Vercel, where our sites live, the dashboard now splits AI traffic into crawlers and assistants, and names each one. Here is what the last 30 days looked like on two sites I run, counted on October 30, 2026.

Luxe Window Works, my window treatment business, received about 5,900 crawler visits and about 3,100 assistant fetches. Real Estate With Shirin, a real estate website I built for my wife's business, received about 3,800 crawler visits and about 1,700 assistant fetches. On the window treatment site, the platform reported no errors for the month: every page the machines asked for, they got.

Read those pairs of numbers as two different facts. The first number says the machines can read you. The second says that when people asked questions in your market, the machines came to your pages for the answers. The first is table stakes. The second is the one connected to revenue.

Vercel: Make your site readable by AI agents

Being read is not the same as being used

It is possible to be read constantly and used almost never. I know because I run the experiment. Starry Van is a buyer's guide website I launched in mid-October, built with the same care as our client sites: fast, clean, full of genuinely useful content. By the end of the month its crawlers had visited 375 times. Its assistant fetches, the visits tied to a live question, stood at one.

The difference is not quality of content. The difference is that nothing outside the site vouches for it yet. It has no business profiles, no directory listings, no reviews, no history. It is a new voice in the room that nobody has corroborated. The established sites above 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 with each other.

Independent research points the same way. 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. When a page's answer was hidden behind JavaScript, or the server refused the request, the agent could not use the page at all. Being fetchable at the moment of the question is the whole game.

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

What makes an assistant fetch your pages

Nobody publishes the exact formula, and you should distrust anyone who claims to have it. The companies document how their visitors work, not how their choices are ranked. What follows is what we can verify, and what our own numbers support.

First, the page has to be readable by a machine that does not browse like a person. Many AI fetchers do not run the scripts that assemble modern web pages. A page served as finished text can be read by all of them. A page that assembles itself in the browser can hand a fetcher an empty shell, and the Vercel study above watched exactly that failure happen.

Second, the business has to be one consistent, checkable fact wherever it appears. Assistants answer with more confidence, and fetch more readily, when the business they are describing is the same business on every surface: same name, same services, same areas, same story, confirmed in independent places. We call the work of building that agreement Clarity, Verification and Trust. Clarity: one clear account of the business everywhere. Verification: claims backed in places the business does not control. Trust: the result, an answer a person and a machine can rely on.

That is why we build websites the way we do at Kodecite. The website is the source of truth, and every profile, listing and article is a spoke connected to it. A wheel rolls because the spokes connect to the hub. Facts scattered across the internet with nothing tying them together are just spokes on the ground.

OpenAI: Overview of OpenAI crawlers

Perplexity: Perplexity crawlers

You cannot buy the visit that matters

OpenAI now sells advertising in ChatGPT, so it is fair to ask whether you can simply pay for the second kind of visit. You cannot. What you can buy is a sponsored card placed below the answer, shown to people on the free tiers, clearly labeled as advertising.

OpenAI states its rule in its own advertising principles, under the heading Answer independence: ads do not influence the answers ChatGPT gives, and ads are always separate and clearly labeled. The answer itself, the part that names businesses and explains the choice, is not for sale at any price.

So there are two ways to appear when a customer asks an assistant about your kind of business. You can rent the card underneath someone else's recommendation, or you can be the recommendation. Only one of those is an asset you keep.

OpenAI: Our approach to advertising and expanding access to ChatGPT

What to do with this

Start by looking. Ask whoever hosts your website for your AI traffic numbers, or pull them from your server logs if you know how. Look for the split: how much is crawling, and how much is assistants fetching at question time. If the second number is near zero, you are being read and not used, and that is fixable.

Then check the basics. Pages that error when a machine asks for them cannot be used as answers. Facts that contradict each other across your website, your Google profile and your listings give an assistant reasons to pick someone easier to verify. Pages that answer real buying questions in plain words give it reasons to pick you.

That is the work Kodecite does: one clear record of your business, corroborated everywhere it appears, on a website built to be read by people and their assistants alike. If you would like to know which kind of AI visits your business is getting, that is a conversation we are happy to start with your own numbers.