You can check your AI visibility yourself, for free. The core check takes about ten minutes, with one step that needs a quick look from your developer.
If you already publish content and still don't show up, the problem may not be how much you publish. It may be that what you publish was built for Google's ten blue links, not for how AI engines read a page. Here is how to check where you stand, and three things to inspect on your own pages when the answer comes back no.
The reader changed, and most content didn't
For years you wrote for one reader: Google's ranking system, deciding which ten links to show. You learned its habits. Keywords, titles, backlinks, one page per topic.
A second reader has arrived, and it reads differently. When someone asks ChatGPT, Perplexity or Google's AI for a recommendation or an answer, the engine does not hand back ten links to browse. It writes an answer and quotes a few sources inside it. What gets chosen now is a passage inside that answer, rather than the whole page it sits on.
That is the shift most content has not caught up to. You can rank perfectly well on Google and still never appear in the AI's answer, because the two readers want different things. Writing for this second reader is what people have started calling answer engine optimization, or AEO. Which is why the first useful move is not to publish more. It is to see what the second reader currently does with you.
How do I check my AI visibility? A step-by-step test
Run it once with a timer. You need a browser, about ten minutes for the core steps, and your developer for one quick check.
No paid tracking tool and no subscription: the free surfaces, one command, and a note of what you find.
- Start a clean, unpersonalised chat. A new conversation is not enough on its own, because saved memory and personalisation shape what you get back. Sign out, or use a temporary or unpersonalised chat, and turn personalisation off where you can. The engine can still guess your location from your connection, so name your town or market inside the question itself.
- Ask the questions your buyers ask, and let it search. Use their words, not your brand name: the problem they have, the category they are shopping, the comparison they are weighing. As a hypothetical, a buyer might type "which project-management tools suit a consulting firm managing several client accounts?", the kind of question asked long before anyone types your name. In ChatGPT, turn on Search so the answer reflects the live web. Note whether you are recommended, mentioned in passing, or not there at all, whether a page of your site is cited as a source, and who shows up instead of you.
- Repeat on each surface, by name. Run the same question in Perplexity, in Google's AI Overview and AI Mode, and in Gemini if you use it. They do not agree with each other, and they often cite largely different sources, so treat each as its own result. If Google shows no AI Overview at all, write down "no AI Overview shown" rather than counting it as absent.
- Check whether your page's text is in the HTML. This step needs your developer for a minute, and it catches a problem the others can't. Many modern sites build their content with JavaScript, and several AI crawlers don't run it. Ask your developer whether your page's main text is present in its initial HTML response (one way: curl -A "GPTBot" your-url), or open the page with JavaScript disabled. This is a technical accessibility check: it shows whether the text is there to begin with, not what any given engine finally retrieved or used. If the text is missing from that initial response, investigate how it reaches the platforms you want to be seen on, before anything else, because a crawler that skips JavaScript has nothing to read.
- Check accuracy separately. In a fresh session, ask "what do you know about [your business name]?" Naming yourself changes the task, so this is not evidence that you get recommended. It tells you something else worth knowing: whether the description is accurate, vague, or confidently wrong. Keep this out of your discovery chats so your own name does not leak into them.
- Write it down. The exact prompt, the date, the surface, and what happened in plain terms: recommended, mentioned, cited, not shown, or feature unavailable. That record is your baseline. Run the same prompts in three months and see what moved.
Three things to inspect when you're absent
If you were absent or thin in the test, you have found a result, not yet a reason. These are the three places worth inspecting first. Each is a possible contributor you can check on your own pages, not a diagnosis the test has made for you. Where an inspection turns something up, you have a candidate to fix. Where it doesn't, the honest note is that the cause is still open.
First, is the answer buried? AI tends to lift a passage, not a page. If the clear answer to the question sits in paragraph six, under three paragraphs of warm-up, there may be nothing clean for the engine to quote. As a rule of thumb, putting the answer near the top, in the first sentence or two, written so it stands on its own, gives the engine something to lift. Check whether your page does that.
Second, can the crawler reach the text? This is the accessibility check from the test. Content that only appears after JavaScript runs is content the JavaScript-skipping crawlers can't read. It is a quiet failure and often an expensive one, because it can undercut everything else on the page. If the test showed your text missing from the initial HTML, this is your first candidate.
Third, are the claims citable? "We help companies grow" gives the engine nothing specific to stand on. What tends to get quoted is specific, checkable and attributable: a real number, a named method, a concrete result. Vague positioning gives an engine little to hold onto, and adding concrete, verifiable facts is a reliable way to give it more. Read your page and ask what, on it, an engine could actually quote.
One myth worth dropping while you are here: an llms.txt file is not a search or citation lever. The major engines don't use it to decide what to cite. If it comes up as a fix, treat it as the new version of the old keywords meta tag.
How to read it, and what it can't tell you
Read your results carefully, and lightly. What you have is a set of answers to a handful of questions, under the conditions you wrote down. That is real, and it is also narrow.
It does not establish your overall standing in AI search, it does not explain with certainty why a particular page was left out, and it does not measure what any of this is worth to you in enquiries or revenue. Answers move from day to day, and the same prompt can read differently an hour later, so one run is a snapshot rather than a verdict. Being named is not the same as being chosen, and being chosen is not the same as a sale.
What the check gives you is the thing most companies are missing: a first, honest look at how the new reader sees you, written down clearly enough to repeat.
What to do next
Run the check before your next content brief or publishing spend. The discovery part takes about ten minutes; the technical check with your developer is a separate, quick step. Write each answer down with the date and the conditions.
If the picture is good, keep the record and run it again next quarter so you can see movement instead of guessing at it. If it is thin, you have a result and a short list of things to inspect, not yet a verdict on the cause.
If you want an outside read of that, our free Signal Check looks at what your check turned up and names one observable problem and one change worth making first. Either way, run your check before you commission the next piece.