It is Thursday morning and the inbox is open on one item from last month: a visitor wrote that the annual price was not clear until the second screen of the pricing page. The item carries a green tag reading Closed in TopDo, because an engineer moved that task to a closed state two days ago. Under the tag sits a sentence the product put there on purpose: “A closed task is ready for a fresh review; it is not yet a verified fix.”
That sentence accurately describes what the workspace just did. Someone pressed Refresh status, GreatFeedback read one node from TopDo with a read-only agent key, checked that the node belonged to the connected workspace, and wrote three things onto the item: the state word, the version TopDo returned, and the time of the check.
Notice what was not read. The pricing page the visitor wrote about was not fetched, and no text on the website was compared against the complaint. The whole exchange happened between two pieces of internal software, GreatFeedback and TopDo.
GreatFeedback keeps those two facts in separate places and never derives one from the other. The tracker state lives in a TopDo mapping holding a node identifier, a link, a state, a version and a checked timestamp. The item’s own status is a different field that only a person sets, and no code path in the backend moves it for you.

Two Facts That Look Like One
The instinct is reasonable and almost universal: the tracker is the system of record, so a closed task is the end of the story. One queue, one status word, one answer for everybody. The alternative, keeping a second register of what has actually been confirmed on the page, feels like duplicate bookkeeping for the same work.
It misreads what a closed node is evidence of. A closed node shows that a person in TopDo moved it to a closed state; it is not evidence that the deployment reached production, that the change landed on the page the visitor reported, or that it addressed what the visitor struggled with rather than what the ticket title said. The stored state is one of three, open, blocked or closed, labelled with your workspace’s own status name, and the card prints “Status not checked” when no refresh has run.
The grounded alternative is to read the tracker fact as a scheduling fact: it tells you when it is worth going to look at the live page, and nothing more. The documentation says so in the product’s own words: “A closed task is not automatically a verified fix. Run a new review after the change ships to gather fresh evidence.”
Read The Status Refresh For What It Actually Is
The status refresh is a pull, and it happens only when a person asks for it. No channel exists by which TopDo announces a closure to GreatFeedback; the endpoint runs when someone presses Refresh status, and it refuses with “Send this to TopDo first.” when the item was never delivered. Until somebody presses that button, a task closed three weeks ago still shows the state it held at the last check.
What the refresh does check, it checks with care. The returned node must come from the connected TopDo workspace or the request fails with “TopDo returned a node from a different workspace.” It must carry a finite numeric version or the request fails with “TopDo returned an invalid status version.” A status version lower than the one already saved is discarded and the stored mapping is returned unchanged, so a late or reordered response cannot move an item's recorded state backwards.
Hold those observations together and the boundary is plain. Every safeguard on that path protects a claim about a node in a task manager, and none of them says anything about a public web page, because the refresh never opens one. When the card turns green you have earned a reliable statement that somebody closed the work.
Keep The Original Report As The Standard
Verification needs a fixed target, and the feedback item is built to be one. The record keeps the visitor’s message, the page address with query string and fragment stripped, the element they selected, and, for a tracked AI finding, the identifiers that link the detail view back to the original report. The patch endpoint accepts only three fields: status, priority and title. The message and the quotation cannot be edited away.
The comment thread is where the standard gets written down. A note is between 1 and 5000 characters, stored with the author’s name and a creation time, returned in creation order, and recorded in the audit log as feedback.commented. A status change writes no audit entry at all. The durable, attributable trail on an item is the notes your team wrote, which makes the note the right place for the check you intend to run later.
Write that check as a sentence anyone could run without asking what you meant. “Verified when the pricing page states the annual charge in dollars before any click” is a check. “Improve pricing clarity” is a wish nobody can settle six weeks later.
Run The Same Page And The Same Goal Again
A review is three inputs: a public URL whose origin is registered for that website, a goal between 10 and 1500 characters, and optional personas, up to the plan's allowance for one review. All three are stored on the review, and the Review again control on the report page brings back the address, the goal and the same personas, so a comparable second run repeats no choice you made the first time; a persona deleted since is left out, which makes its scores incomparable between the two runs.
The second run is a fresh sample, not a cached verdict. The reviewer opens the pages again, in a real browser wherever they render, takes at most the first 6000 characters of each, splits them into excerpts of at most 450 characters with identifiers such as p1_e3. Where a page renders, it also measures its structure and takes screenshots at desktop and mobile widths. Every finding must name one piece of that evidence; the server, not the model, attaches the exact quotation or screenshot and URL. A second, skeptical pass may only keep identifiers from the draft list, and the report prints how many suggestions were removed. The run costs one review from your allowance: 30 per calendar month on Focus, 100 on Growth, 300 on Portfolio, 3 in total on the trial.
Compare the quotations themselves, not the summaries. The old report holds the exact page text that supported the original finding, printed under the label OBSERVED ON THE PAGE, or OBSERVED SIGNED IN AS for a signed-in page; the new report holds whatever that page says today. A finding missing from the new summary is weaker evidence than a changed quotation, because a single run reads only one page of each type, at most eight public types spread across the site, so a page other than the start page may simply not have been read this time. The system prompt puts it in stronger terms: an omission is not proof something is absent from the site.
What This Does Not Tell You
A fresh review verifies only what it can see: the text, measured structure and screenshots of the pages it opened. The scope note on every report from the current reviewer ends with the same sentence: “Interactions beyond signing in, page speed and full accessibility compliance were not tested.” If the original report was about a form that rejected a valid email address, a message that appears only after a visitor clicks, or a page that loads slowly, no second review can settle it, and treating a completed run as confirmation is worse than not checking at all.
The findings themselves remain interpretations. The report states it plainly: “These are AI interpretations, not accounts from real visitors. Use them to form better questions and decide what to test.” A clean second review means the reviewer found no supported objection in the pages it read; it does not mean the visitor who wrote to you in March would now succeed where they had struggled before.
The resolved status carries no machinery behind it either. It appears in the status list, the detail view’s select and the inbox filter, and nothing else in the backend reads it: resolving an item sends no message to the visitor who left an email address, writes nothing back to TopDo, and closes no task. Resolved is a statement your team makes to your team, which is why the evidence behind it has to sit in the thread beside it.
First Steps
The loop closes with three actions, and the first happens before any code is written. Each maps onto a control that already exists on the feedback item.
- Add a team note naming the page and the exact sentence or behaviour a later check must find, then send the item to TopDo with its evidence and source link.
- When the card shows Closed in TopDo, press Refresh status to record the state and time, and keep the item’s own status until you have looked at the page.
- Rerun the review with the same URL, goal and personas, compare the new quotes with the original ones, then mark the item Resolved, noting what you compared.
Spend One Review On Being Sure
A Focus workspace has 30 reviews in a calendar month. Spending one to check a shipped change costs about three percent of the month’s allowance and returns a dated report with quoted text and, where they could be captured and saved, screenshots that either support your claim or do not. Set against a resolved item nobody can defend when the complaint returns, that share is well spent on the only instrument in the product that reads the page rather than the tracker.
The separation is already built into the record: a TopDo mapping that knows a node, a state and a version; an item that knows what a visitor wrote and on which page; a thread that knows what your team decided and who said it. Keep writing into all three, and the status word on the item stops being a guess about the outside world.
