Free tool

Schema markup generator.

Give it one URL. It fetches the page as plain HTML, lists the structured data already on it, names the types it thinks are missing, and writes JSON-LD for them using the page’s own content rather than placeholder text.

What it returns is a language model’s reading of one page. That is a heuristic, not a measurement, and it is a different thing from the numbers in an audit. We say so here rather than in a footnote, because a generated block of markup looks equally authoritative whether the model was right or not.

COST · FREESIGN-UP · NONELIMIT · 3 RUNS PER HOUROUTPUT · MODEL-GENERATED

Run itone public URL · plain HTML · up to a minute

How it scores

Three confidence labels, and deliberately no score.

There is no number out of a hundred here, and that is the design. A composite would have to fuse things that get fixed in opposite ways, which is the same mistake a single AI visibility score makes. Instead every recommendation carries one of three labels, assigned by the model from the signal it found in the page it read.

High confidence

The page carries a strong, direct signal for that type: a real question-and-answer block, a physical address, a price, a numbered procedure. The model is echoing something it can point at.

Medium confidence

A reasonable match rather than an obvious one. Usually a type that fits the page's job but not its current structure, which means the markup arrives before the content that would justify it.

Low confidence

It could help and the signal is thin. Treat these as prompts to look at the page, not as work to do.

THE LABEL IS THE MODEL’S OWN JUDGEMENT OF ITS EVIDENCE. IT IS NOT CALIBRATED AGAINST ANY OUTCOME, AND TWO RUNS ON THE SAME PAGE CAN DISAGREE.

What it cannot see

The six things this tool is blind to.

Stated before you run it rather than discovered afterwards. Every generator in this category has these limits. What is rare is publishing them on the same page as the button.

It cannot see anything JavaScript builds.

The page is fetched once as plain HTML, with no browser and no rendering. Content assembled client-side is invisible to it, so a page that reads as thin here may not be thin to a person. That gap is worth knowing about for its own sake: several AI crawlers fetch the same way.

It cannot see the rest of your site.

One URL, read on its own. It cannot tell you that an Organization block already exists on your homepage and that adding a second one here creates a conflict, because it never looked.

It cannot see whether the answer is on the page.

The model gets roughly the first 3,000 characters of visible text, not the whole page. It can generate an FAQ answer that appears nowhere a reader can find it. Checking that every answer exists in the visible copy is your job, and it is the one check that decides whether the markup helps or breaches Google's guideline.

It cannot see whether the markup earns anything.

Structured data is eligibility, not entitlement. No tool can tell you a rich result will appear, and nobody can tell you a citation will follow. Anyone who says otherwise is selling the correlation.

It cannot validate.

Google's Rich Results Test is the validator and this is not it. Markup that parses cleanly here can still fail there on a required property.

It is not a measurement, and it is not stable.

It is one language model, one pass, one page. Run it twice and the recommendations can differ. Nothing it returns belongs in the same sentence as a figure that carries an instrument, a run count and a window.

Method

What actually happens when you press the button.

instrument: A general-purpose language model, reached through the Vercel AI Gateway. Not Caul, and not comparable with anything Caul reports.
what was measured: Nothing is measured. One page is fetched and read, and a model returns its opinion of which schema types fit it, with JSON-LD written from the page's own content.
how: A single plain HTML fetch with a 10 second timeout, no browser and no JavaScript execution. Title, meta, canonical, Open Graph, existing JSON-LD, heading structure, link and content counts and roughly the first 3,000 characters of visible text are extracted and sent to the model in one pass. The reply is parsed and rejected if it does not match the expected shape.
over what window: The moment you press the button. Nothing is stored, nothing is re-run, and there is no second reading to compare against.
what this cannot tell you: This is a heuristic, not a measurement. It reads one page as plain HTML, so anything JavaScript builds is invisible to it. It cannot see the rest of your site, cannot confirm a generated answer appears in your visible copy, does not validate against Google's Rich Results Test, and cannot tell you whether markup will earn a rich result or a citation. Two runs on the same page can return different recommendations.

Markup is the easy half. The measurement is the other one.

This tool tells you what to add to a page. It cannot tell you whether AI answers name you at all, which is the question the free audit answers: a real sweep of your category across the engines, benchmarked against competitors you name, delivered on a call. You keep the report and the underlying data whatever you decide.

Get your free audit All free tools

Free. The ~~$1,500~~ is what this sweep is priced at when it is sold, not a survey of anyone else’s rates.