Can AI visibility actually be measured?
Partly, and much less precisely than the category implies. There is no equivalent of Search Console for AI assistants: no platform publishes how often it cited you, to whom, or for which prompts. Every AI visibility number you have seen is therefore built from sampling, not from a source of record.
That does not make it worthless. It makes it a survey, and surveys have to be read as surveys — with sample sizes, dates and error bars, not as counts.
How these numbers are actually produced
Nearly every tool in this category does some version of the same thing: assemble a list of prompts a customer might ask, send them to one or more assistants on a schedule, parse the responses for brand mentions and cited links, and aggregate the result into a score.
That is a legitimate method. It is also a sample of a non-deterministic system, which has consequences most dashboards do not show you.
- Same prompt, different answer — models vary between runs, so a single check is a coin flip rather than a reading.
- Personalisation and memory change results per account, and a tool's sterile session is not your customer's session.
- Answers vary by geography, language and model version, and versions change without notice.
- Prompt selection determines the outcome: the prompt list is the methodology, and a flattering list produces a flattering score.
- Parsing is lossy — being mentioned, being cited with a link, and being recommended are different things that often collapse into one metric.
This is why two tools can report very different share-of-voice for the same brand in the same week and both be reporting honestly. They asked different questions, at different times, to different model versions.
What is genuinely measurable today
Plenty, and it is the less glamorous half.
| Signal | Measurable? | How |
|---|---|---|
| Whether AI crawlers can reach your pages | Yes, reliably | Your own robots policy and server logs |
| Whether a page is structured to be extracted | Yes, reliably | On-page signal analysis |
| Referral traffic from assistants | Yes, partly | Analytics referrers, where the assistant sends one |
| Whether a specific prompt cited you | Yes, for that run | A sampled probe, valid at that moment |
| Your true share of AI answers | No | No platform publishes it; every figure is inferred |
| Why an assistant chose a source | No | Not disclosed by any provider |
What Zylx does, and what it refuses to do
Zylx scores pages for AI answer readiness from real on-page signals and returns the specific blocks that are missing, and it assesses AI crawlability as part of its site crawl. Both are grounded in things that can be inspected.
Zylx does not currently measure whether AI answer engines cite you. There is no citation measurement source wired into the product, and rather than estimate a number, the system records every source as unsupported and says so. That is a deliberate choice: a fabricated visibility score is worse than an absent one, because people plan against it.
If citation measurement matters to you today, use a dedicated tool for it and read the output as a sample. We would rather point you at that than sell you a number we cannot stand behind.
Stated plainly because the alternative is worse: if we shipped an estimated score, every downstream recommendation would inherit its error, and you would have no way to tell.
How to run this sensibly
- Fix eligibility first — crawler access and page structure are deterministic and cheap.
- Treat any share-of-voice figure as a survey: check the prompt list, the sample size and the date.
- Track the direction over time rather than the absolute number, and keep the prompt list fixed so the comparison means something.
- Watch assistant referral traffic in analytics — noisy and incomplete, but it is real behaviour rather than inference.
- Do not set targets against a metric nobody can audit.
Frequently asked questions
Can you measure whether ChatGPT recommends your brand?
You can sample it — send a fixed prompt list on a schedule and record the responses. You cannot measure it in the sense of a count, because no provider publishes how often you were cited or to whom. Any figure presented as a total is inferred from a sample.
Why do AI visibility tools disagree about the same brand?
Because they ask different prompts, at different times, to different model versions, and parse the answers differently. Model responses vary between runs even for identical prompts, so two honest tools can report different results in the same week.
Is there a Search Console for AI?
No. No major assistant publishes per-site citation data, which is the core reason this category relies on sampling rather than reporting.
Does Zylx track my AI visibility?
No. Zylx scores pages for AI answer readiness and assesses AI crawlability, both from inspectable signals. It has no citation-measurement source wired, and records sources as unsupported rather than estimating a score.
What should I measure instead?
The deterministic parts: whether retrieval crawlers can reach your pages, whether individual pages carry self-contained answers, and what referral traffic assistants send you. Those are auditable, and they are also the inputs you can actually change.
Are AI visibility tools worth paying for?
They can be, if you read them as sampled surveys and keep the prompt list fixed so trends mean something. They are not worth it if you plan to set targets against the absolute number, because nobody can audit it.
Start with the part you can actually control
Crawler access and page structure are measurable and fixable. Zylx scores both and hands you the missing blocks.
Which AI crawlers to allowGetting into AI OverviewsDoes llms.txt do anything?What Zylx actually scores
Sources
Zylx product details verified against the live implementation on 2026-08-09.