Set it up by hand
The manual alternative to agent-led onboarding: write your own tracked prompts, set the cadence, and register the brands yourself.
Updated October 20, 2026
Everything the agent does during onboarding you can do yourself, in the app, in about twenty minutes. Some people prefer it, and it is the right path when you already know exactly which questions you want measured.
Write like a buyer, not like a search box
"crm software" is a keyword. "What CRM should a 12 person agency use?" is a prompt. Assistants answer the second kind, and the answer names a handful of brands, which is the thing being measured.
Four patterns cover most of a market:
- Category choice. "Best project management tool for a small design studio"
- Comparison. "Linear or Jira for a team of fifteen"
- Problem first. "How do I stop my team missing client deadlines"
- Named. "Is Linear worth it", and the same question about each rival
The named ones feel strange to track and are the most revealing: they show how assistants describe you when the brand is already in the question.
Add them
1. Register the brands first
Open Name matching under Presence and add your own brand with its alternate spellings, then each competitor with theirs. Doing this before the first run means the first run is already being scored against the right set, rather than being re-read later.
2. Track a prompt
Under Presence, open Prompts and track a new one. The Prompt field takes the full sentence, including the qualifier that makes it a real question: the team size, the industry, the country, the budget. A qualifier is what makes an answer specific enough to contain brand names at all.
3. Set "Runs on"
Weekly is the right default for a starting set. Daily belongs on the prompts whose pages you are actively working.
4. Set samples per run
An assistant answers the same question differently each time, so one probe is close to a coin flip. More samples per run narrow the band on the result. Multiple samples per run are an Agency and Enterprise feature; below that a run is one probe and a weekly cadence accumulates the evidence over time instead.
5. Mark one prompt as a control
Set Prompt type to control on one prompt you will deliberately never optimize. In week six, when everything moves at once, the control is the only thing that tells you whether you did it or the model drifted. See control prompts.
6. Give each prompt a topic
A topic is the cluster the prompt belongs to, and the product reports visibility per topic. Tags are free-form for anything else. Twelve prompts do not need either; eighty do, and retro-fitting them is tedious, so set the topic as you add each prompt.
Write all your prompts, then read them as a list and delete any two that would get the same answer. Near-duplicate prompts cost money twice and tell you one thing.
Then leave it alone
Set the schedule and let five runs land before reading a trend. Spend the wait on reading your first results, so you know which numbers to trust when they arrive.
Questions about set it up by hand
- How many prompts should I start with?
- Ten to twenty. Enough to cover the buying journey, few enough that you can read every answer in the first week.
- Are prompts capped?
- Tracked prompts are unlimited on every plan. You pay your model provider for the runs, so there is nothing for us to meter.
- Can I import a keyword list?
- You can, and it is usually the wrong list. Keywords are what people type into a search box; prompts are what they type into an assistant.