September 15, 2026 · Maksim Shchegolev

AI SEO: What It Is, What Actually Decides Citations, and How to Check Your Own Site

AI SEO is the work of making a site readable, identifiable and quotable for the systems that answer questions instead of listing links. It is decided by five things a machine can check in seconds: whether your robots.txt lets the AI fetchers in, whether there is a map for them at /llms.txt, whether your markup says what you actually do, whether any paragraph is short and factual enough to quote, and whether your pages can be dated and sourced.

Want this checked on your own site? Run the free AI visibility check, or order the full audit.

Ask an assistant which company to use, and it does not hand you a list of ten links. It fetches a few pages while it is writing, names two or three sources in the answer, and stops. Everyone else in the market is not ranked lower. They are absent.

AI SEO is the work that decides whether you are one of the two or three. It has almost nothing to do with keyword density and almost everything to do with whether a machine can open your pages, tell what you are, and find a sentence worth quoting.

In one sentence

AI SEO is making a site readable, identifiable and quotable for systems that answer instead of listing, and it is decided by five things a machine checks in seconds rather than by anything a copywriter argues about.

Why it is a different job from SEO

Search ranks pages in advance and shows a list. A person scans the list and clicks. The work is to be higher in that list.

An answer engine assembles the answer at the moment of the question. It sends a fetcher to a handful of URLs, reads what comes back, and writes a paragraph with two or three attributions. Three consequences follow, and they are the whole difference:

1. The fetch happens live, so access is decided live. If your robots.txt says no to the fetcher, nothing else about your site matters. There is no cached version of you to fall back on.
2. The unit that wins is a passage, not a page. The model needs a short, self-contained, checkable statement. A page that opens with the founding story has nothing to give it, however well that page ranks.
3. Being mentioned is not being visited. Your name inside an answer earns nothing on its own. The click follows the citation only when the citation is attached to a source the reader wants to open.

The five things that decide it

Every check that matters falls into one of five areas. This is the same list our free check measures, and each one can be verified by hand in a browser, which is the point: nothing here needs a tool to believe it.

Five areas, and none of them requires a subscription to see. That is deliberate: a score you cannot reproduce is a score you cannot trust, including ours.

The order to do it in

The cheapest fixes are also the ones that decide the most, which is unusual and worth exploiting.

1. Open the door. Unblock the answer-engine fetchers in robots.txt. Minutes of work, and while it is shut everything else is invisible.
2. Say who you are. One block of Organization markup on the home page with your real activity, address and profiles. An hour, and it is the difference between a described business and an unnamed website.
3. Give the engine one thing to quote per page. Rewrite the first paragraph of your five most important pages so it answers the page’s question in 20 to 90 words with a figure in it. An afternoon.
4. Date everything and name your sources. A visible date, the two fields in markup, and a named source beside every number. A machine treats an undated claim as a claim it cannot check.
5. Publish a map. Generate llms.txt from the live pages rather than writing it once. Cheap, and it is the only place where you say in your own words which pages matter.
6. Then look at the harder things. Tables, question-and-answer blocks with matching markup, and pages that are currently empty without JavaScript.

How to tell whether it worked

Three measurements, in the order of how much you can trust them.

Access is a fact. Read robots.txt again, fetch a page as a plain request and see whether the text is in the HTML. It is either open or it is not, and the answer does not vary between runs.

Arrivals are a fact. Visits that come from chatgpt.com, perplexity.ai, claude.ai and the rest are real traffic, and almost nobody counts them, because analytics either buries them among referrals or drops them into direct. Counting them on your own site is free: our visit counter is one line of code and shows the assistants apart by name.

Citations in answers vary, and honest reporting says so. You can ask a model the questions your buyers ask and note whether you were named. What you cannot do is treat one run as a measurement: the same question asked twice, an hour apart, produces different sources. Anyone showing you a single percentage without a method and a repeat count is showing you a number they cannot defend.

What our tool does, plainly

Our free check reads your robots.txt against the 14 AI crawlers we track, fetches llms.txt and the sitemap, and reads up to five pages the way a fetcher reads them, including what only appears when scripts run. It returns a score out of 100 made of the five areas above, so you can add them up by hand and get the same number, and the one fix that moves it most, written as three lines: what is there now, what to change, how to check it worked. No signup, no card, nothing to install.

Two facts about the score that matter more than its size. What we could not read is not counted against you: a refused robots.txt is marked “not measured” and drops out of the denominator, because an unread file is not evidence of a problem. And we publish our own number: this site scores 100 of 100 and the Russian one 99, measured on 2026-09-15 by the same code at build time, so the number falls the day we let our own site slip.

If you want the rest of the findings rather than the first one, that is what the paid tiers are for, and the prices are on those pages. Nothing on this page needs them.

What AI SEO will not do for you

It will not invent demand. If nobody asks the questions your business answers, being quotable changes nothing about your pipeline.

It will not survive a bad product page. An assistant that names you sends a person to a page that still has to do its job.

And it will not hold still. Fetcher names change, llms.txt may be adopted or ignored, and answer formats move under everyone. The parts of this that are safe to build on are the parts that were true before AI search existed: an open site, a stated identity, a checkable claim, a dated page.

Where to start

Open yoursite.com/robots.txt right now and look for the fetcher names. That check costs nothing and answers the only question that comes before all the others. If you would rather have all five areas measured at once, run the free check and read the first fix it hands you.