ChatGPT meta description generator

This route is a generator, which has a precise consequence: it produces content from a brief you supply, not an analysis of a page it inspected. There is no crawl in this codebase, no URL fetch and no reading of your existing tags. The model sees exactly the words you type and nothing else about your site.

That is worth stating first because it determines where the quality comes from. A description written from a four-line brief describing a page is a description of that brief. If the brief omits the one thing that distinguishes your page, the output cannot contain it, and no amount of regeneration will recover information that was never supplied.

Mechanically: 5 runs per rolling 24 hours, 1 to 12,000 characters, an optional 300-character instruction, text sent to our AI server, and output returned with no preamble, commentary or code fence.

Google does not truncate on character count, it fills a horizontal space, so the real constraint is rendered width and the character guideline is a proxy for it. The proxy is loose in a way that the round number 155 conceals. One published guide puts the variance concretely: a description heavy in wide letters can be cut around 140 characters, while one written with narrow letters survives past 165.

The width figures themselves are less settled than the category implies. Two 2026 guides fetched for this page give the desktop slot as about 920 pixels and about 990 pixels respectively, a 70-pixel disagreement about the same rendering, and neither publishes how it measured. On mobile they converge at roughly 680 pixels, around 120 characters.

The correct response to that spread is not to pick a favourite and quote it as a specification. It is to write to the tighter of the two numbers, front-load the meaning so truncation costs you nothing important, and stop treating the last five characters as an engineering problem. Precision beyond the precision of the underlying evidence is theatre.

With the measurement step included, because this route does not perform it for you.

Step 1

Supply a brief carrying the page type, the reader, the query it targets and the one differentiating claim. Those four determine the output almost entirely.

Step 2

Constrain in the instruction field: a character ceiling, a variant count, and keyword placement in the first clause. Nothing is constrained by default.

Step 3

Measure the output externally. The panel's counter is bound to the input, so it counts your brief. Read the first 120 characters in isolation as a mobile check.

Search engines discard the supplied description frequently. One analysis of page-one results is reported at 71 percent of mobile results and 68 percent of desktop; another at roughly 62 percent. Read across sources, the claim that most descriptions are replaced is well supported even though no single figure is.

What is less supported is any particular percentage, and the difference matters if you are going to make decisions with it. These are observational studies of an opaque ranking system, sampling different keyword sets at different times, with methods that are not published alongside the numbers. They agree on direction and disagree on magnitude by nearly ten points, which is about what you would expect from good-faith estimates of something nobody outside Google can measure directly.

The operational consequence is unaffected by which figure is right. You are writing a candidate that is used when it is the best available answer to the query, which makes the description a bid rather than a specification, and makes specificity the only lever you hold.

The character counter above the input is bound to the input field. It measures your brief. It does not measure the generated description, and on a route where length is the entire subject that is a distinction worth making explicitly rather than allowing a reader to assume the convenient version.

There is also no SERP preview, no pixel measurement, no truncation simulation and no comparison against your existing tag. The tool generates text. Every measurement in this workflow happens somewhere else, and pretending otherwise would be the easiest false claim available on this page.

This is also why the instruction field matters more here than on most routes. A ceiling stated up front, under 140 characters, is cheaper than generating and measuring and regenerating, and with five runs per 24 hours the cost of iteration is not zero.

No. The description is not a ranking input and its length is not one either, which the fetched guides state directly and which follows from how the field is used. Its function is entirely at the click: it is the one line arguing that your result answers the query.

Accepting that resolves most of the arguments people have about these fields. There is no keyword density to optimise, because density does not feed a ranking model and, on the evidence of the rewrite triggers, actively increases the chance your text is discarded. There is no penalty to avoid. There is a reader scanning a list.

No promise is made on this page about rankings, click-through rate, or whether any search engine will use what the tool produced. We generate a candidate from a brief, and the systems that decide its fate belong to someone else.

The three rewrite triggers named across the guides are a query mismatch, generic wording and keyword stuffing, and they are not equally expensive to fix. The mismatch is the valuable one, because it is diagnosable from data you already have: find the query a page actually ranks for, and check whether the description addresses it. If it does not, that is a real defect with a real fix.

Generic wording is the failure the brief causes, and it is where a generator is most likely to disappoint someone who fed it three words. The model can only redistribute the specificity you gave it.

Keyword stuffing is the cheapest to fix and the least valuable, because nobody writing carefully does it anyway. Effort spent counting characters sits below all three of these, which is the ranking most pages in this category get backwards.

No crawling, no URL input, no reading of existing tags, no sitemap processing, no bulk generation, no API and no CMS integration. One brief per run, five runs per rolling 24 hours, and the text sent to our AI server.

No measurement of the output, no preview and no truncation simulation. No fixed number of variants, since the count is whatever the instruction asks for. No promise about rankings, click-through or whether a description will be used.

And no affiliation with Google or OpenAI. Where a figure from a published guide or study appears on this page, it is attributed to that source, including the ones that disagree with each other, because the disagreement is part of what the reader should know.