Prompts

Prompts are the whole measurement. Every figure in Recomma is counted from chats for these questions, so the set you keep decides what the numbers mean.

The table

One row per tracked question, each with its own reading:

This is where an Overview figure becomes actionable. A visibility of 73% across the project is a number; the six questions at 0% are a list of things to do.

The default sort is worst-first, because the question you open this page with is “where am I losing”.

Questions/mo

Nobody publishes how often people ask an AI assistant about something. The figure here is an estimate, from DataForSEO, modelled from the questions Google shows in People also ask — not counted from AI conversations. The column treats it as one: a ~ on the number, and Google’s own monthly searches beside it in the tooltip where there are any.

It counts every question that contains all of a phrase’s words, in any order. Each word added can only lower it, so a full question reads close to nothing, and each row is read by the short phrase behind it:

1. the keyword it was drafted from — the biggest, where there are several;
2. otherwise the question Google showed people asking;
3. otherwise, for a question somebody typed, its topic’s name.

Hover a figure to see which phrase it is. Two kinds of reading are drawn apart from the rest:

• Greyed — the topic’s figure, for a prompt with no keyword of its own. Every prompt falling back to the same topic shows the same number, and the tooltip says how many share it.
• broad — a phrase that is also an ordinary word, catching questions about something else. When the estimate runs above 30% of Google’s volume for the same phrase — ten times the usual ratio — the column says broad instead of a number that would top it for the wrong reason, and the row sorts with the unread ones.

Use it to compare prompts, not as a count of people asking. In our test of 308 phrases it ranked phrases almost exactly as Google does, at around 3% of Google’s volume; month to month, its figures moved together by a third across unrelated phrases while Google’s stayed flat, so the column shows the latest month and no trend. The study, with its data, is on the blog. A dash means the phrase has not been read yet; a plain 0 means it was read and no question contains it. Figures are refreshed monthly.

Names you

The second band of controls carries Names you: only the questions whose answers named your brand in the window, or only those whose answers never did. It reads the same count the Visibility column is built from, so a question at 0% is a question the filter puts on the “doesn’t name you” side. The same control sits on Chats and Sources.

In your words

A question can quote your own site. A search-backed engine matches those phrases straight back to your pages, so the answer finds you because the words came from you — which is the site being found, not the brand being recommended. Those two are not the same measurement and should not be read as one.

The toggle carries the count, and pressing it keeps only those questions. Read them separately: a set with several of them scores higher than the same brand would on questions a buyer wrote, and the gap between the two readings is worth more than either.

Tags

Tags are yours — a label on a prompt that means whatever your team needs it to mean: a campaign, a launch, a region, an owner. A prompt can carry several.

Select rows and assign one to all of them at once. Assigning adds; it does not replace, because the rows in a selection rarely carry the same tags and quietly stripping them would destroy work nobody asked to touch.

The tag menu in the filter bar narrows the table to one of them, with the number of prompts carrying each.

The topic rail

Topics on the left, prompts on the right. A prompt carries exactly one topic, and the rail lists whatever topics the prompts carry — there is no separate list to maintain, and a topic with no prompts under it does not exist.

Adding a topic and writing its first questions is therefore one action, not two.

Tabs

• Active — tracked and sampling.
• Suggested — proposed but not yet tracked. These cost nothing while they sit here. Tracking one starts consuming credits on the next cycle.
• Inactive — kept, not sampling. Their history is preserved.

A prompt can move to Inactive without anybody pressing anything. When ten or more answers in a row name no tracked brand and no other company, the question is not measuring a market — it is being answered generically — and sampling it further spends credits on nothing. It is paused, and the row says so. Resume it if you want it sampled anyway; it will not be paused a second time.

Adding prompts

Four ways in, all landing in the same place:

• Add prompt — write one yourself.
• Import keywords — paste a list, typically from Search Console. Recomma turns them into questions; a bare keyword measures nothing, because models answer questions.
• Discover prompts — Recomma reads what you already track and proposes questions you are missing. Suggestions cost nothing until accepted.
• Suggest prompts from uncovered demand — on Keywords, the keywords your category is contested on that no tracked topic reaches. Same writer, same destination; the list comes from the market rather than from you.

Adding a topic proposes its first questions in the same step, ticked by default, with the stage each one was assigned. What you untick is never stored.

Writing a good prompt

A prompt should be a question a buyer would actually type, specific enough that a useful chat names brands:

Cover the whole decision. A set that is all awareness questions tells you about category presence and nothing about whether you win the comparison — and the comparison is where deals are lost.

Deleting a prompt

Deleting removes its history. The readings you have built are per prompt, so a prompt deleted after a month takes that month with it. Prefer Inactive if you might want the question back.

Related