Your AI visibility score has dropped. Does your website need work? Here’s what I’d check before asking the team to change anything.
A report lands on your desk saying your AI visibility is down and your website content needs work. It sounds urgent, and it sounds measurable. But before you assign that work, it's worth asking what the number is actually measuring.
An AI visibility score counts how often your organization is mentioned or linked in AI-generated answers to a set of questions. A lower score means you showed up less often in those answers. It doesn't tell you why, and it doesn't tell you whether your content is the problem.
Citation tracking tools generally work the same way. They pick a list of questions, run them through ChatGPT, Gemini, Google's AI features, or similar services, and count how often you appear. That sounds reasonable until you look at the conditions.
Those questions are typically asked by a system with no history. There's no memory of earlier conversations, no location, no previous follow-up questions, and no sense of what the person already knows. Real people don't search like that. A prospective student who's spent twenty minutes asking about nursing programs, costs, and campus housing is going to get a very different answer to "which schools should I consider?" than a blank session asking the same thing cold.
That's the core problem. AI answers are increasingly shaped by the person asking. The tools measuring your visibility mostly strip that context out, so what you're getting is a reading of how you appear to nobody in particular.
The questions themselves matter too. If the tool chose them, they may not reflect what your audience asks. And AI services answer the same question differently from one attempt to the next, so a small sample can swing without anything on your side changing.
Suppose you appeared in 40 of 100 answers last month and 30 of 100 this month. With a consistent test, that's ten fewer appearances in those answers. It's not evidence that fewer people found you, visited your site, or applied. So before you react, ask the provider a few direct questions:
If the answers are fuzzy, treat the score as a weak signal. It's not a brief.
You can't control how an AI service ranks sources, and you can't see the context each person brings. What you can control is whether your content gives an AI tool something it can confidently use.
That comes down to a short list of things. Is the answer actually on your site, stated plainly? Is it current? Does it match what your other pages say? Is the page structured and accessible enough that a machine can tell what's a heading, what's a fact, and what's a caveat? If any of those fail, an AI tool has a reason to lean on someone else's page, or worse, to get your answer wrong.
This is also why I'd rank wrong answers above missing mentions. Suppose an AI answer says a grant is only available to homeowners when renters can apply too. Someone could rule themselves out without ever reaching your website. That matters more than slipping a few points on a dashboard.
When you find a wrong answer, check the pages it links to. What you find decides the next step:
A rising score deserves the same scrutiny. More mentions aren't a win if the answers keep getting your deadlines wrong.
Visibility recommendations often come with technical fixes attached. Schema markup and llms.txt files are the two I see most, but there’s plenty of evidence that neither of those matter as much as great content.
Schema markup labels a page so software can understand it. Ahrefs initially found that AI-cited pages were almost three times more likely to use it, but better content or better site maintenance could explain that. In a follow-up comparing pages that added schema with similar pages that didn't, they found no clear citation gain in Google AI Mode or ChatGPT, and slightly worse results in AI Overviews for reasons that weren't resolved. Those pages were already frequently cited, so the results may not carry over to less visible sites.
An llms.txt file lists the pages you want AI tools to read. A study of May 2026 traffic found 97% of roughly 38,000 valid files got no requests that month. And a request doesn't mean the content was used or cited.
None of this rules out other benefits. Schema has legitimate uses. But I wouldn't approve either one just to chase citations. If there's another reason for the work, be clear about what it should achieve and how you'll know.
You'll get a far more useful read by testing real questions than by watching a score. Start with what people already ask in site search, service inquiries, or audience research: eligibility, deadlines, costs, availability. "Can renters apply for this grant?" "Is the office open on weekends?" "What does this program cost?"
If you're already using Conversational Search in Squiz Funnelback Search, the questions people type there are a strong place to start. Include questions that don't name your organization, too.
Then run the check:
Read past the mention. Are the details right? Do the linked pages back them up? Could someone find what they need next? This won't replicate every personalized conversation your audience has, and it's not meant to. It shows you where your content holds up and where it doesn't.
Turn what you find into specific work. "Improve AI visibility" isn't a task anyone can act on. "The grant page still says 'homeowners only.' Confirm with the grants team that renters are eligible, then correct it" is.
That's the part we've focused on at Squiz. Squiz Content Intelligence doesn't try to guess what an AI service told a stranger last Tuesday. It maps the topics on your site, generates the questions people are likely to ask about each one, and tests whether your content can actually answer them. Then it shows you where answers are strong, weak, missing, or conflicting, alongside the accessibility and structural issues that make content harder for people and machines to read. Your team reviews the recommendations and decides what to change.
Check again monthly and after fixes, and compare the answers you saved, not just the score. If answers improve, that's a good sign, though changes to the AI services themselves can play a part.
Your next content update deserves a better brief than "the score went down." Start with whether your content deserves to be the answer.
If your review points to a content problem, see how Squiz Content Intelligence can help your team investigate and prioritize fixes.
