Free tool

Meta description generator.

It reads one public page, counts the description already on it, and returns three drafts written to a stated rubric.

The drafts are written by a language model, which makes them a starting point rather than a finding. The character counts are counted, in the browser, from the text the model returned. The page marks which is which on every card, because the two are not the same kind of thing.

COST · FREESIGN-UP · NONESTORED · NOTHINGUPDATED · 2026-08-15

How it runsone page, one model call, nothing kept

Give it a URL

STEP 1

Any public page. Nothing is stored, and there is no account and no email gate.

The page is fetched and read

STEP 2

Title, headings, existing description and a sample of the visible text go to the model. Nothing behind a login is reachable.

Three drafts come back

STEP 3

One benefit first, one question opening, one number first, each with its character count recounted here and its rubric result stated.

What it checks

Four things, and it names which of them it counted.

One of these is arithmetic and three are judgements about copy. A tool that presents them as one kind of output is telling you less than it knows, so they are labelled separately on every result.

The description already on the page

Whether one exists at all, what it says, and how long it is. A page without one hands the choice of sentence to Google.

Length against the result layout

Character counts for the existing description and for all three drafts, counted from the strings rather than taken from the model's own arithmetic.

Whether the page's own subject survives into the draft

The model reads a primary keyword off the title, headings and body, then reports whether each draft contains it. Read off the page, not from keyword data.

Call to action and emotional trigger

Whether each draft asks the reader to do something, and which lever it pulls if any. Both are model judgements about copy, stated as such on every card.

How it scoresone length band · four presence checks

There is no score out of a hundred.

The rubric is a length band plus four presence checks, and it stays two separate readings rather than becoming one number. Blending a count with four judgements would produce a figure that cannot be acted on: you would not know whether to cut fifteen characters or rewrite the sentence, and those are opposite jobs.

The length band. Under 70 characters is too short. 70 to 155 is optimal. 156 to 160 is acceptable, and may truncate on some result layouts. Over 160 is too long. The count comes from the length of the string itself, recomputed in your browser rather than read from the model's own claim about it, because a count that can be checked should be.

The four presence checks. Is the primary keyword in the draft, is there a call to action, is there an emotional trigger, and which of the three styles did the draft use. All four are read by the model, so all four are judgements. Every card says so under the note.

The thresholds are the conventional ones for the SERP snippet, not a Caldrin measurement. Google truncates on pixel width rather than character count, so treat 155 as a working ceiling rather than a line in the platform's documentation.

What it cannot see

Stated before you run it, not discovered afterwards.

A free tool that overstates itself costs more credibility than it earns. Six limits apply to every run, and none of them are fixable by running it again.

Whether Google will use the description at all.

Google rewrites descriptions on a large share of results, using body copy it picks itself. No tool can tell you in advance which of your pages that happens to. A better description improves the odds it keeps yours, and that is the whole of the claim.

Your click-through rate, before or after.

Nothing here touches Search Console, a SERP or a log file. The notes on each draft explain why a description might earn a click. That is an argument about copy, not a prediction, and it has no measured CTR behind it.

The rest of your site.

It reads the one URL you give it. It cannot tell you whether this description duplicates another page's, whether the page is indexed, whether it is canonicalised elsewhere, or whether it competes with a page of your own.

Search demand behind the keyword it names.

The primary keyword is inferred from the page by a language model. There is no volume, no difficulty and no competitor data in the run. It is what the page appears to be about, which is not the same as what people search for.

Anything a fetch cannot reach.

One server-side request, ten second timeout, no JavaScript execution and no login. Content that only appears after render, or behind a paywall or an account, is invisible to it, and so are pages that block automated requests.

Whether AI answers name you.

A meta description is a search result problem. Whether ChatGPT, Gemini or AI Overviews name your company in your category is a different question, needs a different instrument, and is not something this page can answer.

Method

The instrument here is a language model, and it is not Caul.

Caul is the measurement instrument, and it produces citation figures with a sample size and a capture window attached. This tool produces drafts and an opinion about them. Nothing on this page is a Caul figure, and no output of this tool should be quoted as one.

instrument: A general-purpose language model, called through the Vercel AI Gateway. Not Caul.
what was measured: Two counted quantities: the length of the description already on the page, and the length of each returned draft. Everything else on the result screen is model-generated.
how: The URL is fetched once, server side, with a ten second timeout. Title, headings, existing description and a sample of the visible text are sent to the model with a fixed rubric. Character counts are recomputed in the browser from the returned strings rather than taken from the model.
over what window: One read, at the moment you press the button. Nothing is stored, nothing is re-run, and there is no second reading to compare it against.
what this cannot tell you: A model assessing a page is a heuristic, not a measurement. The drafts are copy suggestions and the keyword, trigger and call-to-action verdicts are judgements about text, none of which is evidence about rankings, clicks or citations. Reading the same page twice can return different drafts. Character counts are exact; nothing else on the screen is.

If the real question is whether AI answers name you.

That one needs a sweep rather than a page read: a fixed prompt set run across the engines, benchmarked against competitors you name, delivered on a call so the findings get explained. It is free, and you keep the report and the underlying data whatever you decide afterwards.

What is in the report All free tools