The new review form is open: a website selected, a URL pasted, a goal typed into the box labelled "What do you want to learn?", and a line of fine print under the submit button reading "Uses 1 review · 61 remaining". The person filling it in has already started two runs this week, both against the pricing page, both with goals that differ by a handful of words, and both reports are still sitting unopened in the reviews list.
The instinct behind that is a reasonable one: a review looks like a button, and a report that did not quite land seems like something you fix by pressing again with better wording. The counter disagrees, and it disagrees precisely. The allowance check is a strict less-than, so the last review you may start is the one that takes the count up to the limit and the next request is refused outright; and the count is incremented in the same database transaction that creates the review, with the usage row locked while it happens. The charge lands when you press the button, not when a report arrives.
What you are spending is also smaller than it looks from outside. A run is bounded before it begins: the reviewer reads at most eight public page types, plus up to twelve signed-in ones on a paid plan, caps the output of every model call, sends each call at most 100,000 characters of text plus screenshots, and gives up entirely at a four hundred and ten second deadline, extended by forty seconds for each group of four personas you choose. Pressing the button again does not make that envelope larger.
The allowance is therefore not a rate limit placed on an open-ended process. It is a count of small, bounded readings of your own pages, and the unit worth planning around is not the run but the specific question the run is meant to settle.

Count The Allowance Before You Frame The Question
There are only four sets of numbers to know. Focus carries thirty AI reviews and ten thousand visitor submissions, Growth carries one hundred reviews and fifty thousand submissions, Portfolio carries three hundred reviews and two hundred thousand submissions, and the fourteen day trial carries three reviews and five hundred submissions. Thirty reviews on Focus works out at roughly one per working day if you spread them evenly across a month, which is a useful way to feel the size of the allowance before half of it has gone in a single afternoon of reruns.
The common assumption is that a published limit is a soft target the product will quietly stretch and bill for afterwards. It does not stretch. When the step that starts a review finds the locked meter already at the plan's review count, it refuses the request and answers "You've used your AI review allowance. Your existing reports remain available." The billing panel in workspace settings states the same policy in its own words: "Monthly allowance. No surprise overages." Reaching the limit stops new reviews and changes nothing about the reports you already hold.
Because the count goes up when the request is accepted rather than when a report finishes, the number in front of you is a count of starts, not of finished work. The overview prints it as a plain subtraction of usage from the plan allowance, described as the reviews left in your allowance, and the review form repeats the same arithmetic beside the words "Uses 1 review". Read that number before you write the goal rather than after you have already decided you want to start another run.
Treat The Trial Meter As A Total, Not A Monthly Rate
Usage is recorded against a period, and the period is decided by one rule with two outcomes. A paid workspace uses the current year and month in UTC, so its meter is a row named for this month; every other workspace uses a single fixed period named EVALUATION. On the first of the next month a paid workspace begins writing into a row that did not exist before and is created empty, which is the whole of the reset: no scheduled job clears anything; the row name simply changes. That is why the billing panel can promise "Resets on the first of the month, UTC" with no machinery behind it.
A workspace that is not paying never changes rows. Every review and every visitor submission it records lands in that same EVALUATION period for the entire life of the trial, so three reviews means three in total rather than three per month, and five hundred submissions means five hundred in total. The billing panel says exactly this where the monthly wording would otherwise sit, printing "500 total during your trial" beneath the submissions bar and the trial end date beneath the reviews bar. The pricing page states both halves of the rule in a single clause: "Paid allowances reset on the first of each month in UTC; the trial allowances are totals."
So a trial is not a small month and should not be planned as one. You have three runs to spend across fourteen days, after which a new review is refused with "Your evaluation has ended. Your feedback stays here. Choose a plan to run more reviews." Three runs is enough to score one page with a different persona each time, or to read one page before a change and again after it has shipped. It is not enough to absorb a rephrased repeat, and no first of the month is coming to hand the decision back to you.
Spend The Run On A Question The Budget Can Answer
The form gives you two levers and no others: a goal of between ten and fifteen hundred characters, prefilled with "Help a first-time visitor understand the product and confidently take the next step", and optional personas, each shown with its name, category and description, up to your plan's allowance for one review. The goal travels into the call that writes the findings and into each persona's scoring call, every call except the evidence check. It never chooses the pages; the reviewer finds your page types itself, from the sitemap, your links and your app's routes, and always reads the start page you give it. So the goal is not a label printed on the report; it is the instruction that decides which of the captured evidence becomes a finding.
The instinctive goal is a broad one, on the theory that a wider question returns more. It returns less, because the findings call keeps at most five improvements across every page it read, and a broad goal spreads those five thinly across up to twenty page types instead of settling the one decision you care about. Each page contributes at most its first six thousand characters, less when many pages share the text budget, cut into excerpts of at most four hundred and fifty characters, beside up to twelve measured facts and its first screen at both widths, and a finding that does not name one of those pieces of evidence is discarded by the server before you ever see it.
Knowing the rest of the envelope explains why a second attempt at the same broad question comes back just as thin. Without personas a run makes at most two model calls: one writing the report, and one on a second pass that acts as an independent skeptical evidence editor and runs only when a draft finding exists, each with its own output cap, and an answer that runs past its cap fails the review rather than being trimmed. Each persona adds one scoring call after the report is written. The report request asks for up to five directly supported improvements and states plainly that zero findings is a valid answer. That is what one unit of your allowance buys.
Know Which Failures Return The Credit
A review that stops on our side before the first model call always returns its credit and never counts toward the limit that follows. Once the model has started, a model error, an unusable answer, the time limit or a restart returns the credit up to five times each calendar month (UTC) for each workspace, or five times across a trial. The credit goes to the period the review was charged to rather than to whichever period is current when the failure is recorded, and the five are counted in that same period. Past the fifth, the review stays failed, its credit stays spent, and the review page says so.
The first model call still matters in two cases. A review you cancel returns its credit only before that call: cancel while it is waiting in the queue and the credit comes back; cancel after the model has started and it stays counted. A review whose pages hold no text to quote works the same way. A starting page that cannot be reached always returns the credit, because the review stops before the model is ever called.
Three guards sit in front of the meter and are easy to mistake for the allowance itself. A workspace may have only two reviews in flight at once, and a third attempt is answered with "Two reviews are already running. Wait for one to finish before starting another." Review starts are rate limited per workspace at twelve per hour and sixty per day, whichever plan you are on. And every start carries an identifier created with the form, so a resubmission returns the review already saved under that identifier: an impatient second click on "Start the review" costs one review, not two.
Reviews that get stuck do not keep one of your two places. A review still in the queue after six hundred seconds with nothing left to run it is marked failed, and one that passes its allowed running time is closed with "This review timed out. You can start a new review." The queued review always returns its credit; the timed-out review returns it under the same limit of five if the model had started, and always if it had not. That is the reason to look at a slow review's actual state before starting it again: a review still running has already used its credit, and a replacement takes a second one from your allowance while the first one may still finish and deliver its report.
What This Does Not Tell You
A spent review buys an interpretation of your pages and nothing more. The limitations stored on every completed report say so exactly, closing with "Interactions beyond signing in, page speed and full accessibility compliance were not tested." The report page adds, beside the findings: "These are AI interpretations, not accounts from real visitors. Use them to form better questions and decide what to test." No way of allocating the allowance changes that; thirty runs on Focus are thirty readings of pages, and not one of them observed a visitor, and apart from the sign-in step where you set that up, none clicked a control or submitted a form on your site.
The meter is equally silent on whether a run was worth starting. A review that completes with nothing surviving the evidence check is counted in full, and it stores the summary "This limited review found no sufficiently supported change to recommend. That does not mean the site is flawless. Try a more detailed public page or a different review goal." The usage the product reports for a workspace is two integers, reviews and submissions for the current period. It does not record which findings you accepted, which you rejected, or whether anyone opened the report; tracking a finding does create an AI-sourced item in the inbox, so that decision leaves a trace, but nothing joins the trace back to the review allowance it was spent from.
First Steps
Three actions turn the allowance from a number you notice at the limit into a constraint you plan around, and each of them uses a surface the product already has.
- Open workspace settings and read both usage bars and the line below: a monthly UTC reset when paid, an end date on a trial. Note the reviews left and the date.
- Replace the prefilled goal with one question for one audience and one decision, pick personas only for what you want scored, and start at the URL it concerns.
- Track the findings you accept and note why you reject the rest. Start the next review once a change ships or a new question appears, not after a weak report.
Let The Next Question Earn Its Review
The two halves fit together tightly. The per-run limits are deliberately small, at no more than twenty page types, a fixed output cap on every call and four hundred and ten seconds, plus forty for each group of up to four personas, because a narrow reading carrying verified evidence is worth more than a broad one without it; and the per-period allowance is a count of exactly those small readings. A workspace that presses the button whenever a report disappoints is no longer settling a question but repeating one in the hope of a different answer, and the meter records every attempt either way.
So let the question set the pace. On a trial that means three questions and no rewrites, because the EVALUATION period never rolls over and nothing about the calendar will rescue a wasted run. On Focus, Growth or Portfolio it means thirty, one hundred or three hundred questions in a UTC month, refreshed because the name of the period changed rather than because anyone promised it would. The counter under the button is not an obstacle to the work; it is the closest thing the product keeps to a record of how many things you and your team decided were worth finding out.
