Diagnose
Run one question against the assistants, then read a single brief that says why the answer looks the way it does and what would change it.
Diagnose takes one question a buyer might ask, samples the assistants that can answer it, pulls in every signal Trakkr already holds about that question, and writes one brief explaining why the answer looks the way it does. It is the single-question read: Prompts watches a chosen list every day, Research sweeps the whole category once to find questions worth watching, and Diagnose is where you go when one of those questions deserves a real explanation with the evidence attached.
Open it when a position surprises you, when someone asks why you do not show up for something, before you commission a page you would rather justify first, or a few weeks after you shipped a fix. It is deliberate and capped, so it is not a page you run against every prompt you track.
Running a diagnosis
Type a question of three characters or more and press Diagnose. The counter beside the button shows what you have used against your allowance this month. Under it, Low visibility queries offers up to six chips drawn from your own past diagnoses that scored below 50, weakest first, and a table lists everything you have run. Picking a chip starts a fresh run of that question. While a run is in flight the page shows each assistant as it answers, so you see the finding land rather than a spinner.
Two things happen at once. The assistants are sampled, and every other signal Trakkr holds on this question is gathered into evidence packs.
The sampling asks the question the way a person would, several times per assistant, and only in the modes each one can honestly be asked in. Trakkr calls Claude and Gemini with no search tool, so they can only answer from training. Perplexity always searches. ChatGPT can do both.
| Assistant | Sampled with search | Sampled from memory |
|---|---|---|
| ChatGPT | Yes | Yes |
| Perplexity | Yes | No |
| Claude | No | Yes |
| Gemini | No | Yes |
The packs gathered alongside it are your canonical visibility score for this query, your ranking trajectory over recent weeks, phrasing variants of the question, the pages cited while the assistants answered, your own pages on the topic and the schema they carry, brand perception themes, Reddit threads that shape answers in your category, and any earlier diagnosis of the same question.
All of it goes into a single Gemini Flash synthesis call written against a strict schema. Nothing on the page is free-form: every claim carries a [r1] reference chip you can click through to the exact report, citation, page, variant, model, or competitor behind it.
Reading the brief
The completed page is one argument, top to bottom: a verdict with the tracked series beside it, then four bands of evidence ending in the plan.
The verdict
A one-sentence headline with the supporting explanation beside it, then up to three consequences that follow from it. Under those, How the answers were assembled draws the supply chain of the answer: Asked, Searched, Read, Listed, Chosen, with the bottleneck stage marked. That is usually the fastest read on the page. If the drop is between Read and Listed, the pages the assistants read do not include you. If it is between Listed and Chosen, they do and you still are not picked.
Tracked visibility
The rail on the right carries the canonical score for this query, its change since the last run, a sparkline, and a sentence on the shape of the movement (rising, declining, stable, or volatile). Every past run of this question is listed underneath, so you can open an older brief and compare. If the query is not tracked there is no score here, and the rail says so: tracking it is what starts the series.
The pages that write this answer
Every page cited while the assistants answered, with how many answers cited it, your slot on that page, and which brand the page leads with. Open a row for the receipt: what the page actually says, read from the page. Minor sources collapse behind a toggle.
Below the table, Read, then not used lists pages an assistant fetched and then chose not to cite. Your own pages appearing there is the sharpest signal on the band. A source you never want counted can be ignored from its row menu, which applies to this brief and every later one, and ignored domains are listed at the bottom with a Restore button.
Who wins this answer
Every brand the assistants named, with a position per model and an aggregate score. Your row is highlighted, and the header shows a model-agreement percentage: high means a settled consensus, low means the assistants tell different stories and one number will mislead you. Underneath, when phrasing variants were measured, the band names your strongest phrasing and your blind spot. A wide swing between them is a phrasing problem, not a content gap.
Who names you
One row per assistant, split into With search and From memory, each showing how many runs named you out of how many. A dash means that assistant cannot be asked that way, not that it never named you. The sentence beside each row is the reading, and the four cases need different work.
| What the runs show | What it means |
|---|---|
| Named with search and from memory | The assistant has learned you and finds you. Nothing to fix here. |
| Only with search | Your pages carry it, the training does not. Keep the pages strong and cited. |
| Only from memory | It knows you but does not surface you when it looks. A retrieval and citation problem. |
| Named in neither | No page edit fixes this on its own. |
Open the answers drawer from the row menu, or from Read the N answers in the page menu, to read the raw text of every sampled answer.
The fix plan
Each row is one root cause, ordered by how many sampled answers it could move (or by severity when the run had no sample to count). A row carries what to do, the insight behind it, a You and Leaders evidence pair, and a functional owner and timeframe.
| Severity | Read it as |
|---|---|
| critical | The main reason you are not where you want to be here. Start with it. |
| significant | A real contributing factor. Next. |
| moderate | A smaller drag, worth knowing, lower priority. |
Owners are functions, not people: Content, Marketing, Dev, Product, or Partnerships. Each fix also carries an action type, so the Actions queue knows what shape of work it is: create content for a gap, create FAQ content, add schema markup, fix meta descriptions, add llms.txt, unblock AI crawlers, get listed on a source, pitch to a publication, or update a cited page.
Acting on a diagnosis
Hand the fixes to Actions. Every gap already exists as a pending action, created when the diagnosis completed. Tick the ones you want and Add fixes to Actions in the bar at the bottom of the page moves them into your Actions queue as in-progress work with the diagnosis evidence attached. Rows already in Actions show a link straight to them.
Track the query. Track query in the header adds this exact question to your active Prompts list, so it joins the tracked run and the Tracked visibility rail starts filling. If it is already tracked the button reads Tracked and does nothing. Tracking does not duplicate the fixes, which use the normal Actions path.
Re-run after you ship. Re-run starts a fresh diagnosis of the same question and spends a diagnosis from your allowance. Once there is more than one run, What changed since the last run in the page menu opens the difference. A later run can show that the answer moved. It does not claim your work caused the move, which is what Results is for.
The page menu also carries Export: print or save as PDF, download the source ledger as CSV, copy the link, or copy the verdict as text.
Limits
Diagnoses are counted per brand, per calendar month, and the count resets on the 1st. The counter next to the search box shows where you are, and the box locks when you reach the cap.
| Plan | Diagnoses per month, per brand |
|---|---|
| Growth | 25 |
| Scale | 200 |
| Trial | 10 |
Enterprise follows the Scale allowance. A Diagnose link from elsewhere in Trakkr can land on an existing brief rather than spending a new run: if the same question was diagnosed for the same market in the last 30 days, the page opens that one. Typing a query into the box yourself always starts a new run, as does Re-run.
Common questions
Why does Diagnose use four assistants when Prompts uses eight?
Diagnose samples the four chat assistants where most buyer questions land, several times each, so it can measure the spread rather than take one answer per model. The daily Prompts pipeline fans out across all eight so the visibility score reflects the whole market. The brief uses your Prompts score as the authoritative number, so the two pages never disagree about how visible you are.
Why does part of the brief say there is not enough data?
Each band is backed by a specific evidence pack. If a pack is thin or absent (a new brand, no Perception runs, no audited pages, an untracked query), the brief says so rather than inventing detail. The sparse parts fill in as you use more of Trakkr, and a re-run picks up whatever has arrived since.
Can I diagnose a competitor instead of myself?
Not directly. A diagnosis is scoped to your active brand, and the verdict is always about your position. To read a competitor, run the question for yourself and use Who wins this answer, which lists every brand the assistants named with a position per model.
Does running a diagnosis change my visibility score or my tracked prompts?
No. It runs on its own pipeline and writes to its own tables. Your daily Prompts run is untouched and a diagnosis does not consume prompt slots. The only thing it spends is one diagnosis from that brand's monthly allowance.
What happens when I press Track query?
The question is added to your brand's active prompts and joins the next tracked run. If a prompt with that exact text already exists, nothing changes and the button reads Tracked. From then on the Tracked visibility rail has a real score to plot.
Why is a recommendation owned by "Content" or "Dev" rather than a person?
The brief assigns each fix to a function so the right team knows it is theirs. Trakkr does not try to guess your org chart, and your team triages from there.