Choosing questions for DeskSEO AI visibility runs

DeskSEO's AI visibility feature asks an AI model the questions your customers ask and records whether your brand is mentioned in the answers. The mechanics are simple. The hard part, and the part that decides whether the numbers mean anything, is choosing the questions. This guide covers how to write them, how DeskSEO runs them, and how to read the results. For the general method behind any AI visibility measurement, see measuring AI visibility.

How a run works

1. You enter your brand: the full name, your domain, and optionally other names people use. Matching ignores accents. Your domain and long names (10+ characters or three or more words) match as a whole phrase, ignoring case. A short name that is also an everyday word, such as "Brightside" or the "example" part of example.com, only counts when the answer writes it capitalised as a name and not just because a sentence starts with it, so "look on the bright side" is not a mention.
2. You enter 3 to 10 questions, one per line. A counter under the box shows the number of questions and the price as you type; above 10 the button locks and tells you how many to remove, and nothing is charged. DeskSEO can suggest starter questions for your niche and market, which you then edit. A question is rejected only if it contains your full brand name or domain, not a single everyday word from it.
3. You see the price (questions × the price per question, excluding VAT) and pay it from your prepaid balance.
4. DeskSEO asks each question on its own, as an ordinary user would, with no instruction that mentions you. The model is Anthropic's Claude, and the run page says so. It answers from what it learned in training, without searching the web during the run, so the result reflects how well the model knows your brand rather than what your site says today.
5. Answers are stored word for word. For each question, DeskSEO records whether you were mentioned, quotes the sentence, and notes your position if the answer is a list of recommendations.

Each question takes a few seconds to tens of seconds; you can leave the page and come back. If the AI fails to answer every question in a run, the whole run is refunded to your balance automatically. If only some questions fail, the run page shows which ones; contact support if you want those refunded. Prices are on the pricing page.

Rule 1: never put your brand in the question

"Is Northfield Cycle Works good for e-bike repairs?" will produce an answer about Northfield Cycle Works because you asked about it. That tells you something about reputation, but nothing about visibility. Visibility is whether the AI brings you up on its own when a customer asks a general question. Watch for indirect leaks too: your domain, a product name only you use, your founder's name, or your slogan all count as mentioning the brand.

Rule 2: write the way customers ask

Customers do not ask "best e-bike service provider solutions". They ask "Where can I get my e-bike battery checked in Leeds?" or "My e-bike motor cuts out on hills, who can fix that?" Good sources for real phrasing:

• Questions from sales calls, emails, chat and reviews.
• Your Search Console queries, rewritten as full sentences; see Search Console in DeskSEO.
• People Also Ask boxes for your main topics.
• Forum and community threads where your customers ask for recommendations.

Rule 3: cover the different ways people choose

A set of five near-identical questions measures the same thing five times. Cover different angles instead:

Angle	Example for a hypothetical bike workshop in Leeds
Direct recommendation	"Which bike repair shops in Leeds are good for e-bikes?"
Problem-led	"My e-bike battery drains fast in winter. Where can I get it tested in Leeds?"
Specific service	"Who builds custom wheels in West Yorkshire?"
Comparison or alternative	"Is it better to take my e-bike back to the big chain store or to an independent workshop?"
Segment or situation	"I commute 20 miles a day by e-bike. How often should it be serviced, and where in Leeds?"

Mix broad and narrow questions. Small businesses rarely appear for the broadest questions but can appear for specific ones that match what they do well.

Rule 4: include the place or segment when it matters

For local businesses, a question without a location produces national or generic answers that will almost never name you. Include the city or region the way customers would. For businesses serving a segment (small agencies, vegan bakers, first-time buyers), include the segment. The suggestion feature asks for your niche and market to do exactly this.

Rule 5: keep the set stable

To compare runs over time, use the same questions each time. Save your final set somewhere you can copy it from, and change it only deliberately, for example once a quarter, noting the date. If you add a question, compare it only with runs that included it.

Questions for businesses that are not local

Online shops, software companies and publishers replace location with use case, segment and constraints. A project management tool for design studios might ask: "What project management software works well for a ten-person design studio that bills by the hour?", "Which tools let clients approve design proofs inside the project?", "What is a simpler alternative to the big enterprise project tools for a small agency?" Each describes a real buying situation without naming the product, and each tests whether the model connects the product to that situation.

A worked example

The hypothetical bike workshop runs the five questions in the table above. The result: mentioned in two of five answers. It appears for the custom wheel question, listed second, with an accurate description, and for the e-bike battery question, listed fourth, but described as "a general bike shop" without mention of its e-bike specialism. It is absent from the broad recommendation question, the comparison question and the commuter question; two larger chains and a well-known independent shop appear there instead.

The owner rewrites the e-bike service page with specific diagnostics, prices and turnaround times, adds a short guide on servicing intervals for daily commuters, and corrects an outdated directory description that called the shop "a general bike shop". Because the measured model answers from its training rather than a live search, those changes are not expected to show in the next run; they reach models as new versions are trained on fresher data, and they reach search-enabled assistants, such as ChatGPT search and Perplexity, much sooner. The owner keeps the same five questions, runs them every quarter, and checks the search-enabled assistants by hand in between. The page changes helped search traffic within weeks either way.

Reading the results

• Mention rate: the share of questions that mention you. This is the headline number.
• Position in lists: being mentioned first in a list of five is different from being fifth.
• The sentence: what the AI says about you. Is it accurate? Is it the strength you want to be known for?
• Competitors named: read the full answers. The businesses named where you are not show who the model associates with your category.

One run is one sample. The same question can get a different answer next time, and other assistants learn from other sources. Look at the trend across runs, measured before and after each round of work.

What to do with the results

• Not mentioned anywhere: the model may not know enough about you. Work on consistent descriptions and mentions across the web; see entity SEO.
• Mentioned with wrong details: find the outdated sources (old listings, articles) and fix them; make your own site the clearest statement of the facts.
• Mentioned for some angles and not others: build or improve pages that answer the missing angles clearly; see writing content AI assistants quote.

How often to run

Monthly is plenty for most businesses; quarterly is fine if little changes. Because the measured model answers from its training rather than from a live search, the effect of your own work usually shows only when the model is updated, which can take months. Running the same set several times in one day mostly measures randomness in the answers rather than anything you did, so spread runs out, compare like with like, and use manual checks in search-enabled assistants to see faster-moving changes; the ChatGPT search guide describes how.

A sample set to adapt

1. "Which [service] businesses in [city] are well regarded?"
2. "[Specific problem]. Who can fix that in [city]?"
3. "What should I look for when choosing a [service] for [situation]?"
4. "Is [common alternative] or an independent [service] better for [need]?"
5. "Recommend a [product or service] for [segment] with [constraint, such as budget]."

Replace the brackets with your own details, read each question aloud, and ask yourself whether a real customer would type it that way. If not, rewrite it.