Skip to content
SilktideHelp

How prompts are chosen

Presence does not pull its from a list of what people ask AI assistants, because no such list exists. OpenAI, Google, Anthropic, Microsoft, and Perplexity do not publish what their users ask. Instead, Silktide writes questions from what it knows about your brand - who you are, who you serve, where, and who you compete with - and aims every one at a moment when someone is choosing between providers like you.

Measurement then shows which of those questions are genuinely contested, and search data shows how much demand sits behind the same need. This page covers both halves: how prompts are written, and how to judge whether one is worth tracking.

What Silktide knows about you

Every topic and prompt is written from the same picture of your brand. Most of it is gathered when you create the workspace, and all of it can be changed afterwards.

InputWhere it comes fromWhere you change it
Background - what the brand doesRead from your website when you create the workspaceSettings → Brand profile
Audiences - the people you want to reachDrafted from your website, confirmed during setupSettings → Brand profile
Country, local area, and languageThe workspace's first regionSettings → Regions
Competitors, and how much each one mattersSuggested during setup, chosen by youSettings → Brands
Topic name and descriptionProposed by Silktide, or written by youSettings → Topics
The prompts a topic already hasYour existing promptsSettings → Topics

Audiences matter most. Prompts are written for the people named there. A payroll company whose audience is "finance teams choosing payroll software" gets the questions a finance lead asks when comparing providers, not what an employee asks about a payslip. If your prompts feel like they are written for the wrong people, fix the audiences first.

Competitors set the scope. Your primary and secondary competitors show the tier and reach at which you actually compete, so questions are pitched where brands like them are the natural answer. Watched brands are left out of this, because they are not competitors.

Search data does not pick prompts. Nothing on this list is a keyword tool or a search volume figure. Prompts are written from what a buyer in your audience would ask, and search demand is looked up afterwards, as evidence (see Is this prompt worth tracking?).

A new region reuses your questions. When you add a region, you choose which existing prompts to copy, usually all of them. Silktide rewrites each one as it would be asked in that language and place, rather than writing a new set, so every region measures the same needs.

Topics come first

A is a front where one of your audiences chooses between providers, and where you could be winning or losing independently of your other fronts. Before Silktide proposes one, it applies a single test:

If you asked an AI assistant five questions a member of this audience would ask here, would the answers name providers, and could you or your competitors plausibly be among them?

That test rules out three common mistakes:

  • Information topics. Regulations, definitions, how a process works. However relevant the subject is to you, the answers name agencies and concepts rather than providers, so nobody is ever measured.
  • The wrong scope. A regional bank whose rivals are other regional banks is not contesting "best banks in America", which assistants answer with the biggest national names every time. It contests banking in its region. Your competitors tell Silktide where your contest is.
  • The wrong person. Topics are written for your audiences, not for a generic member of the public.

Silktide proposes as many topics as your workspace genuinely contains, most central first, and does not pad the list to reach a number. Each suggestion comes with a one-line reason it is worth measuring. When you ask for more later, Silktide knows the topics you already measure, the ones you turned down, and where you stand in each, so it aims at fronts next to where you are trying to grow and stays away from topics you marked Monitor or that show No contest.

Then the questions inside each topic

For each topic, Silktide pictures one specific person from your audience - their role, their situation, the choice they are trying to make - and writes the questions that person would actually type or say. Every question follows the same rules.

  • Aimed at answers that name providers. At least four in five are questions an assistant naturally answers with names: best and top, choosing between options, fit for a particular situation, value and price, alternatives and switching, reputation and trust. The rest are tasks where assistants usually recommend someone while answering. Pure information questions such as "What is WCAG?" are never written, because an answer that names nobody measures nothing.
  • At the scope you compete at. A place is named when a person choosing locally would name it ("in Denver"), and left out when they would not. "Near me" is never used: the assistant does not know where the person is.
  • No brand names. Discovery prompts name nobody: not you, not a competitor, not anyone else. Which brands the answer names is the measurement, so putting a name in the question would spoil it.
  • Honest, not flattering. Questions are aimed where you and your rivals are all plausible answers. They are never built around your product names, slogans, or unique features so that only you could win.
  • Short and natural. Chat-length, usually under 12 words, in your workspace's language and spelling.
  • Varied. Five questions on one template measure one thing five times, so each set mixes its angles.
  • New. Each round avoids repeating or lightly rephrasing the prompts the topic already has.

Every question is written twice, once for assistants and once for search engines, expressing the same intent. Assistants are asked the conversational version; Google and Bing are sent the keywords.

Asked of assistantsSent to search engines
Which payroll provider is the best value for a small team?best value payroll software small business
Which hospitals in Denver are best for knee replacement?best knee replacement hospital denver

When you write a prompt yourself, Silktide writes its search phrasing for you. It never overwrites a search phrasing you chose.

Reputation questions work differently

A brand topic asks about one brand by name, so its questions follow a fixed pattern instead: what is it, is it legit, what do people think of it, what are the alternatives, and is it worth it. There are at most seven, and they use the brand's main name only, because whether assistants connect your other names to you is part of what is measured.

These prompts are marked branded. Brand topics are excluded from headline visibility and position, so naming yourself in an answer to "What is Acme?" cannot inflate your competitive standing.

You have the final say

Generated prompts are a starting point. The questions your sales and support teams hear every week are better evidence of what your customers ask than anything Silktide can infer, so add them. Edit wording that does not sound like your customers, and delete anything too narrow or duplicated.

Changing your brand profile affects prompts written from then on. It does not rewrite the prompts you already have.

Is this prompt worth tracking?

This is the question customers ask most often, and an honest answer starts with what nobody can tell you.

Nobody knows how often a question is asked

The companies behind the assistants do not publish how often any question is put to them. Figures sold as "AI prompt volume" are estimates, modelled from panels of users or from search data, and a panel cannot say reliably how often one exact wording is asked.

Exact wording is the wrong unit anyway. People phrase the same need in countless ways, often mid-conversation. A prompt in Presence is not a claim that those precise words are typed thousands of times a day. It is a sample of an intent: one representative way of asking a question your audience does ask, measured repeatedly so the result is reliable. Several prompts in a topic sample the same arena from different angles, and the topic's result is the thing to act on.

What you can show instead

Four kinds of evidence, all from your own workspace, answer "is this worth it?" better than an invented volume.

  1. Search demand for the same need. Each prompt's search phrasing is matched to Google's monthly search volume for the country and language it is measured in. It appears as Monthly searches on Search rankings (under What you rank for) and on Search gaps. It counts typed Google searches rather than AI conversations, but it is the best public evidence that people want this thing, and the same people are increasingly asking assistants instead.
  2. The answers name providers. On AI gaps, a question where your competitors are named reliably is a live contest: assistants treat it as a recommendation, and somebody is winning it. A question where no tracked brand is named is not earning its place, and a whole topic like that shows No contest on the Topics screen.
  3. Assistants go shopping. The Searches screen shows what an assistant looked up before it answered. When its first search names a shortlist of vendors, it has treated your prompt as a buying decision, which is exactly the moment worth being present for.
  4. Visits arrive. Analytics shows the visits AI platforms and search engines actually send to your sites. Visits are counted per site, not per prompt, so they will not single out one question. They do show whether AI answers send you customers at all, which is the business case for tracking the questions behind them.
Reading Monthly searches
  • The figure is Google's average monthly searches over the last twelve months, for the prompt's country and language. It is refreshed at most once a month, because that is how often Google updates it.
  • Monthly searches appear once the workspace measures a search engine. See Settings → Platforms.
  • Unknown means the phrase has not been looked up yet; it fills in after the next measurement. No data means Google had no figure for that phrase, which is common for long, specific phrasings and does not mean nobody asks. 0 means Google reports that nobody searches for it.
  • A figure belongs to one phrasing. A topic's queries read together say more about demand for the arena than any single number.
  • Branded phrasings measure demand for the brand itself, so compare them with each other rather than with discovery questions.

Making the case for a prompt

A prompt earns its place when most of these are true:

  • Your audience genuinely asks it when choosing. Your sales and support teams would recognise it.
  • Its search phrasing has measurable monthly searches, or the other phrasings in its topic do.
  • Answers to it name providers, including your competitors.
  • You are missing or behind, so there is something to win - or you are ahead, and it is ground worth defending.

Reword or delete a prompt when its answers name nobody, when it asks the same thing as another prompt, or when nobody you sell to would ask it. Every active prompt is asked of every enabled assistant on every run, so pruning weak ones is a real saving. Deleted prompts keep their history, and you can restore them later.

Last updated

Was this page helpful?