Gemini Press release polisher standards

One constrained generation over a document you supply. Input of 1 to 12,000 characters plus an optional 300-character instruction goes to our AI server, and a model returns the rewritten text under a prompt requiring it to preserve the original meaning and factual claims and to emit no preamble, commentary or fence.

The size relationship is unusually favourable. Two independent sources in this result set put a press release at 300 to 500 words, roughly 1,800 to 3,200 characters. The entire document fits inside a single request, so the model evaluates the headline, the lead, the body and the boilerplate as one object rather than as fragments.

The dominant product in this market generates a release from a short brief. The failure modes of the two approaches are not symmetrical, and on a document distributed to press the asymmetry matters.

Generation manufactures attributed speech

A release requires a quote. Given a brief without one, a generator supplies it and attaches a named person's title to it. That is a fabricated statement in a document whose purpose is to be quoted verbatim.

Generation manufactures specifics

Figures, dates and positioning language arrive as plausible filler. The reader cannot tell from the output which numbers came from the brief and which were completed by the model.

Polishing is bounded by the input

This route is instructed to preserve the factual claims supplied. Gaps in the source tend to remain gaps rather than becoming confident assertions, which is the behaviour you want when the document carries legal exposure.

Neither removes the review

QuillBot's own generator page instructs users to proofread and fact-check the output. That advice applies to a polish with equal force, because nothing in either system verifies its own result.

The instruction is the only control on the operation, and scoping it to one objective per run produces measurably more usable output than bundling several.

Step 1. Submit the complete document

Headline through boilerplate in one request. Submitting the body alone denies the model the relationship between the lead and the quote, which is where most structural weakness in a release actually sits.

Step 2. State one objective and its constraint

Reduce to 400 words while retaining both quotes. Or restructure the opening paragraph to answer who, what, where, when and why. A request carrying three objectives typically satisfies each of them partially.

Step 3. Verify against the conventions

Run the checks in the next two sections manually. Nothing on this route validates AP style, and the model's compliance is a property of its training rather than of any rule engine here.

PR Newswire, the largest distributor in the category, publishes these explicitly. Every AI generator page fetched for this route omits them entirely, which is a useful signal about what those products optimise for.

Numerals

One through nine spelled out, 10 and above as numerals, with numerals used unconditionally for ages, percentages, temperatures, times, dates, monetary amounts and sports scores.

Series comma

AP omits it. The published example is keynote speeches, panel discussions and networking opportunities, with nothing before the conjunction.

Title position

Formal titles capitalise before a name and lowercase after it or standing alone. This is positional rather than semantic, which is why it is so frequently wrong.

Reference sequence

Full name and title on first mention, surname alone thereafter. Periods and commas sit inside the quotation marks.

Lengths

Body 300 to 500 words, headline under 100 characters where possible, boilerplate 100 words or less and positioned before the media contact rather than after it.

It is the first line an editor evaluates and the most mechanically specified element of the document. eReleases sets it out in detail.

Structure

City, state where required, then the date in parentheses, then an em dash before the body begins. Worked examples include NEW YORK, N.Y. (Oct. 15, 2025, 10 a.m. ET) and LOS ANGELES (Mar. 1, 2025, 6 a.m. PT).

Stand-alone cities

Thirty major United States cities require no state, and forty-seven international cities require no country. Adding one anyway marks the writer as unfamiliar with the form.

Never abbreviated

Eight states are always written in full: Alaska, Hawaii, Idaho, Iowa, Maine, Ohio, Texas and Utah. AP abbreviations are also not postal codes, which is the more common error.

Date form

No ordinal suffixes, so June 3 rather than June 3rd. Four-digit years. No relative references such as yesterday, because the document may be read days later.

Every item below is a property of the system rather than a positioning choice, and each is a reason the manual review stage is not optional.

No style validation

There is no AP rule engine, no linter and no compliance report. The model applies conventions it has learned when asked, and nothing downstream confirms it did.

No formatting or layout

The engine emits plain text. Logo placement, header alignment, the release-timing line and the three centred hash signs at the close all remain in your template.

No distribution

No wire submission, no media list, no scheduling and no embargo handling. The document is returned to you and stops there.

No fact verification

Preservation of factual claims is an instruction constraining generation, not a check on truth. An incorrect figure is polished and remains incorrect.

No file input

Pasted text only, with no file upload and no document parsing, unlike several generators in this market.

No claim is made here, because the causal chain does not support one. Pickup depends on the newsworthiness of the announcement, the fit with a given outlet, the relationship behind the send and the competing news on the day. None of those is a property of the prose.

The defensible claim is negative and worth stating precisely: editing removes disqualifiers. A buried lead, a malformed dateline, a headline that describes nothing and a boilerplate longer than the news all give a reader reasons to stop. Eliminating them stops avoidable loss of interest rather than creating interest.

Where the underlying announcement is not newsworthy, polishing produces a well-formed document about nothing, and that is a content problem no editing stage resolves.

Three cases. A release with no quote, where the correct action is to obtain one rather than smooth around its absence. A release whose facts are unconfirmed, where polishing raises the confidence of the language without raising the confidence of the claim, which is the worst direction to move.

And a release under embargo or legal review, where the wording has been agreed by parties who did not agree to its being rewritten. A polish that improves the prose of a negotiated sentence has changed a negotiated sentence.

Outside those, on an ordinary announcement with real facts and real quotes that simply reads too long and too loosely, this is exactly the right instrument, and a full release runs through it in a single pass.