How to check your AI visibility yourself in an hour, and when a manual check stops being enough

What can an hour of manual prompts tell you?

An hour of manual prompts tells you whether the engines name you at all, who they name instead, and which sources they lean on. It cannot give you share of answer with any precision, or tell you what changed since last month. We will say where fixing them starts to cost money.

This is the manual we would hand a marketing lead at a small company: ten prompts, two engines, three runs each, one tally sheet. It is slower than a free checker and more honest, because you read the raw answers and know exactly what you asked.

The reason to spend the hour is that buyers now start in the engines. In G2's survey of 1,076 B2B decision makers across North America, EMEA and APAC in March 2026, 51% said they start software research in an AI chatbot more often than in Google.¹ G2 runs a software review marketplace, so it has a stake in the answer.

A disclosure before the method. Sigzen AI sells AI visibility audits and a free check, so we have an interest in where you decide the manual hour runs out. We have tried to draw that line where the data puts it rather than where our pricing would like it.

Prior work

Others have published versions of this method. Ahrefs offers a free checker that reports how often AI answers mention a brand, with the topics and domains around them.² Prose, a marketing services agency, published a March 2026 audit guide that scores answers on presence, placement, accuracy and citation quality.³ LLM Pulse, which sells an AI visibility tool, published a ten-step guide in July 2026 that runs prompts in private sessions and logs them in a spreadsheet.⁴

What this post adds is the sampling logic behind three runs, prompts shaped by 2026 prompt-length data, and a named line where the spreadsheet stops being enough.

The one-hour method, step by step

The method is six steps and sixty answers: ten prompts, two engines, three runs of each. Set a timer, because the hour is the point. A check that takes a whole afternoon will not be repeated next week, and a check that is not repeated is an anecdote.

The one-hour manual check

1. Write ten prompts (10 minutes)

   Buyer questions in buyer words, no brand name in the prompt.

2. Set up the tally sheet (5 minutes)

   One row per answer: prompt, engine, run, mention, citation, rival named.

3. Open a fresh private window

   Logged out, memory off, a new window for every engine.

4. Run each prompt three times on two engines (40 minutes)

   Paste the answer into the sheet, then close the window and open a new one.

5. Total the sheet (5 minutes)

   Mentions out of sixty, citations out of sixty, the rival named most often.

6. Repeat next week, same day, same hour

   Keep the prompts fixed so the totals can be compared.

Two rules keep the hour honest. Paste the whole answer, not your summary, so next week you can check what you counted. And never edit a prompt mid-session: keep running a bad one, note the problem and replace it next week.

Sixty answers will not give you a precise share of answer, and later sections say why. They will show whether you appear, who appears instead, and which pages the engines quote.

Why do the prompts have to read like real prompts?

Because engines answer the question they are asked, and buyers in 2026 ask long, specific questions, not keywords. A prompt such as "accounting firm Canada" measures something no buyer types.

Google says the average AI Mode search is triple the length of a traditional search query, in US insights it published in May 2026 on AI Mode's opening year.⁵ In the same post, Google said AI Mode queries about planning had grown 80% faster than AI Mode queries overall over the previous six months.⁵ Both are company statements about its own product.

ChatGPT answers mostly come from the model's memory rather than from a live search. Ahrefs, which sells SEO and AI visibility tools, studied 66.6 million free, logged-out ChatGPT responses collected in August–September 2026 and found that 14.4% triggered a web search.⁶

The commercial share of those conversations is growing. Profound, which sells AI visibility tools, analysed 7.5 million US English ChatGPT conversations from 23 June 2025 to 22 June 2026.⁷ The share with commercial intent rose from 13.9% to 19.2% over that year.⁷

How buyers prompt in 2026

Longer questions, mostly answered from memory, more of them commercial

3×the length of a classic query, average AI Mode search (US, Google statement)Google, May 2026 ⁵

14.4%of free, logged-out ChatGPT responses triggered a web searchAhrefs, Aug–Sep 2026 ⁶

19.2%of US English ChatGPT conversations had commercial intentProfound, Jun 2025–Jun 2026 ⁷

Prompt length: what the data says

Search-enabled ChatGPT prompts have nearly doubled in length in a year, while chat prompts without search have become shorter. Semrush, which sells AI visibility tools, analysed US clickstream data from October 2024 to February 2026.⁸ Comparing January–February 2025 with January–February 2026, the average search-enabled prompt grew from 4.7 to 8.7 words.⁸ Prompts that did not trigger search went the other way, from 24.9 to 13.5 words.⁸

US clickstream, Semrush

Average ChatGPT prompt length in words, early 2025 and early 2026

Search-enabled, Jan–Feb 20268.7

Search-enabled, Jan–Feb 20254.7

No search, Jan–Feb 202613.5

No search, Jan–Feb 202524.9

So write prompts of roughly eight to fifteen words, as a question or request with a constraint a buyer would add: a country, a budget, a use case. Leave your brand name out; a prompt that names you measures recognition, not visibility. Our guide to building a prompt set from real buyer questions covers where the questions come from and how to weight them.

Ten example prompts

Five for a direct-to-consumer skincare brand selling in the UAE, five for an accounting firm serving small businesses in Canada. Swap the category and the country for your own.

1. Which fragrance-free skincare brands ship within the UAE in a day or two?
2. Best moisturiser for humid weather that does not break me out, available in the UAE
3. Is there a UAE skincare brand with a refill programme and clean ingredients?
4. Sensitive skin routine under AED 300 that I can buy online in the UAE
5. Compare the top three vitamin C serums sold in the UAE for dark spots
6. Which accounting firms in Canada work with small online stores that also sell into the US?
7. Cloud accounting firm for a two-person consultancy in Canada, fixed monthly fee
8. Who can do GST and HST filing plus year-end for a small restaurant in Canada?
9. Best accounting firm for a Canadian startup that needs SR&ED claims handled
10. Compare accounting firms in Canada that specialise in ecommerce and inventory

Which two engines, and why a fresh private session?

Our view: run ChatGPT and Google AI Mode, because between them they cover the assistant buyers open on purpose and the one that sits inside the search box they already use. Gemini and Microsoft Copilot are good second-week additions.

The private session matters more than the engine choice. A logged-in account carries memory, chat history, saved preferences and a location, and an engine uses all of them when it writes an answer. If you have ever asked ChatGPT about your own company, a check from that account measures a conversation you have already had.

How to get a clean window

Open a private or incognito window, do not sign in, and close it after every answer. For ChatGPT, use the free logged-out experience and note that it is the tier Ahrefs studied when it found most responses answered without a web search.⁶ For AI Mode, open Google in the private window, switch to the AI Mode tab and type the prompt there.

Record what you cannot control: the private window still reveals your rough location through your network, and answers may vary by time of day. Write the date, time and country on every row, so week three is comparable with week one.

The tally sheet: mentions, citations and the rival named

One row per answer, three things counted, the rest pasted in for later. The discipline is in counting the same way every week. Two definitions prevent the most common mix-up.

Share of answer

The share of answers to a prompt set in which your brand appears. It is a sample estimate, so it moves week to week even when nothing about your brand has changed.

Mention vs citation

A mention names your brand in the answer text. A citation links to or attributes a page on your site as a source. An answer can mention you without citing you, and cite you without recommending you.

The columns

Prompt, engine, run number, date and time, country, mention (yes or no), position (which brand is named before you), citation (yes or no, with the URL), rival named, and the full answer text. Count citations carefully; engines differ in how many sources they attach. In Semrush's analysis of 126 million US AI search prompts from January–April 2026, ChatGPT cited an average of 15 sources per response and Gemini an average of 3.⁹

The rival column is the one small companies skip and the one that pays most. After three weeks, the name that appears most often there is your real competitor in AI answers, whatever your sales team thinks. Our guide to picking a rival set explains how to turn that column into a stable benchmark.

Illustrative. Fictional brands and invented, round totals.

Why is three runs the floor, and one run not a measurement?

Because an AI answer is written fresh each time, so each run is a draw, and a single draw cannot tell you the shape of the distribution. Ask once and your brand is either in or out. Ask three times and you can see whether a prompt is an always, a sometimes or a never, which is the information you act on.

One run per engine, which is how most people test, is a screenshot, not a measurement. It can show a wrong fact or an unexpected rival; it cannot say whether you went up or down, because the next run may differ for no reason.

Why more prompts beat more runs

MaxAEO, which sells an AI visibility tool, published a July 2026 piece on sample size that treats repeat runs of one prompt as clustered observations rather than independent ones.¹⁰ Runs of the same prompt resemble each other, so a fourth run adds less than a new prompt would, and three is where another run stops paying for a small set.

Dmitrij Żatuchin's July 2026 preprint breaks answer variation into its components, including the model and the language of the query.¹¹ Our measurement spec works through the arithmetic, including why repeat runs widen the margin of error rather than shrinking it.

Our view: treat sixty answers as a screen, not a score. Report counts, "named in 14 of 60", and never a percentage share, because a percentage claims a precision the sample does not have. Our note on how often to re-run checks explains why weekly runs and monthly totals suit a set this small.

What a manual check cannot do

A manual check cannot set the country or language, keep a history, attach a margin of error, or keep your own account out of the answer. None is a reason to skip the hour; each is a reason to know what it is not telling you.

Four limits, in order of how much they mislead

Country and language. The engines retrieve sources in the language of the question and shape answers to the location they infer, and a private window in one country cannot ask as a buyer in another. Żatuchin's preprint found the language of the query to be a large component of answer variation in the region it studied.¹¹ If you sell in Canada and the UAE, an hour in one of them measures one market.

History. Your sheet holds what you pasted, but the engines change under you: models update, the search step retrieves different pages, a rival publishes a comparison. Without a fixed method and day, you cannot tell your change from theirs.

Interval. Sixty answers carry a wide margin of error, and the sheet will not tell you how wide. Two weeks that differ by five mentions may not differ at all.

Your own account. Even in a private window, your network location and the time of day leak in. A logged-in check adds memory and history, and is the most contaminated of all after a day of asking the engine about your own category.

Manual check against monitored measurement

Manual hour

• Ten prompts, two engines, three runs
• Your country, your language
• Totals, no interval
• History in your sheet alone
• Private window, still your network

Monitored measurement

• Many prompts, five engines, repeated runs
• Country and language set per run
• Share of answer with an interval
• Trend, alerts and rival tracking
• Clean sessions through a data provider

When does a tool or a service pay?

A tool or a service pays at one of three points: past about 30 prompts, past one country, or when the number goes into a board report. Below all three, the manual hour is the better buy, because you read every answer yourself and learn the category.

Thirty prompts is where the hour breaks. At three runs on two engines, thirty prompts is 180 answers, and the pasting alone takes an afternoon. A second country doubles everything and adds the location problem the private window cannot solve. A board report needs an interval and a trend, and a sheet of totals has neither.

When the manual hour stops being enough

Prompts →

Over 30 prompts, one countryA tool's entry tier pays; the sheet is the bottleneck.

Over 30 prompts, several countriesAn audit or a service; each market is its own sample.

Up to 30 prompts, one countryThe manual hour works; keep it weekly and fixed.

Up to 30 prompts, several countriesA tool with country control, or a small audit.

Countries →

Entry tiers cost less than many buyers assume. In our dataset of published GEO prices, seen on 5–6 October 2026, the median monthly US-dollar entry price across the 16 tools with such a tier was $79.¹² Entry tiers cap prompts and countries, and a tool measures without fixing anything.

A service pays when no one internal will act on the sheet, or when the number carries weight. Our audit runs 150 buyer prompts across five engines, each prompt on three separate days, and reports share of answer with confidence intervals, at our published price.¹³ Our view: buy that when a decision rides on the number, and keep the manual hour running alongside it, because reading answers yourself is the cheapest education in your category.

What to do this week

Run the hour once, this week, before you buy anything. Write the ten prompts, open the private window, log the sixty answers and read them. Then book the same hour next week and run the same prompts.

After three weeks you will know three things: whether the engines name you, which rival they prefer, and which pages they quote. If the rival column has one name in it, read that rival's cited pages before writing anything new. If the citation column is empty, start with your own pages.

Our view: the manual hour is worth keeping even after you buy a tool, because people who read sixty raw answers a week understand their category better than people who read a dashboard. Outgrow it on prompts, countries or stakes, not on patience.