Most UK businesses asking about an AI visibility tool do not yet know which prompts matter to them, which engines their customers use, or what a normal week looks like.
A paid dashboard will not answer those questions. It gives you a score on day one and a different score on day two, with no way to tell whether the change means anything.
A spreadsheet forces you to find out. It costs nothing but time, and after a month you will know exactly what you would be paying a tool to automate. This is how we build one, column by column, with the formulas and the point at which it stops being enough.
Why the spreadsheet comes first
The spreadsheet comes first because the hard part of AI visibility tracking is the design of the measurement, and no tool does that part for you. The current buyer's guides for this keyword are written by vendors. Writesonic's roundup ranks itself, WebFX's ranks its own OmniSEO first, and Frase's puts Frase at number one. None of them shows a formula.
WebFX goes further and says manual prompting "results in inaccurate tracking". That is true if you type one prompt into ChatGPT and write down what you see. It is not true if you record enough answers in a structured way, which is the measurement discipline the survey of 45 generative engine optimisation studies argues no single run can replace.
There is also a budget argument. Entry plans run from about $29 to $99 a month, and the tiers that track a competitive prompt set sit between $189 and $399. A month in a spreadsheet tells you whether you need 15 prompts or 150 before you pick one.
The columns the tracker needs
The tracker needs one row per answer, with twelve columns, and every later formula depends on that shape. Do not build it as a grid of prompts across the top and weeks down the side. You will have nowhere to put the second run of the same prompt.
Set up a sheet called log with these columns:
date: the day you ran the prompt, as a real date so you can group by week.engine: ChatGPT, Gemini, Perplexity, Google AI Overviews or Google AI Mode. One value per row, spelled the same way every time.prompt_id: a short code, P01, for the underlying question ("best accountants for small businesses in Northampton").paraphrase_id: A, B or C. Each prompt gets three wordings, because "recommend an accountant in Northampton" gets a different answer from "who are the top small business accountants in Northampton".run: 1, 2 or 3. The same paraphrase on the same engine on the same day, repeated.mentioned: Y or N. Your brand name appears anywhere in the answer.cited: Y or N. A link to your domain appears in the sources.position: your place among the businesses named, 1 for first. Blank if not mentioned.competitors: the other businesses named, separated by semicolons, in order.sentiment: positive, neutral or negative, judged by what the answer says about you.snippet: the sentence that mentions you, pasted in.week: a formula,=ISOWEEKNUM(A2), so the summary sheet can group by week.
Columns 6 to 8 are what the paid tools report. Keyword.com's documentation defines its headline metrics as the "% of runs where your brand appeared", the "average ranking position across runs" and the "% of runs where your brand appeared in the top 3", which are these columns with arithmetic on top.
Columns 9 and 11 are the ones you will thank yourself for later. The competitor list tells you who is winning the prompt. The snippet tells you what the engine thinks you do, which is often wrong and often fixable.
How many rows a week you need
You need enough rows per engine per week that a ten-point change is visible above the noise, and that means about 90 answers per engine. Ricardo Batista at cloro has done the sample size arithmetic: roughly 72 answers at a 25% visibility rate gives a 95% interval of about plus or minus 10 points, and about 300 answers narrows it to plus or minus 5.
So the starter shape is 10 prompts, 3 paraphrases, 3 runs, which is 90 answers per engine per week. Across four engines that is 360 rows. At about two minutes per answer, including pasting the snippet, you are looking at 12 hours a week, and that cost is the reason this method has a shelf life.
If 12 hours is too many, cut engines before you cut prompts. Batista's advice on a fixed budget is to spend it on more prompts, not more repeats, because the gap between prompts is bigger than the gap between runs of one prompt.
For most UK service businesses the two engines to keep are ChatGPT and Google AI Overviews, since AI Overviews sit on top of the searches people were already making.
The weekly routine
The routine is the same every week: same prompts, same paraphrases, same engines, same day, fresh browser session, and a hard rule that the prompt set does not change mid-month. Pick a day and keep it. We use Tuesday morning so that a Monday website change has had a day to settle and the run never collides with month-end.
Open each engine in a private window so your history does not shape the answer. Paste paraphrase A of prompt P01. Record the row. Open a new chat and paste it again for run 2, then run 3, then paraphrase B. Do not ask a follow-up; that is a different prompt and does not belong in the log.
The no-changes rule matters more than it sounds. Batista's piece puts it plainly: add prompts mid-window and you change the denominator and "manufacture a trend". Keep a second sheet called prompts with the id, the three wordings and a since date. If you must add or retire a prompt, do it on the first Tuesday of a month and note it.
"One run is a coin flip with the answer hidden." Kevin Indig, How to make prompt tracking more accurate, Search Engine Land
Indig's own protocol is five runs per prompt per platform, weekly, over a 40-prompt panel. Three runs across ten prompts gets you the same discipline at a size one person can hold.
The formulas for mention rate and citation rate
Mention rate is the number of answers in which you were mentioned divided by the number of answers you logged, for one engine in one week, and the formula is two COUNTIFS. On a summary sheet, put engines down the side and ISO week numbers across the top. In the cell for ChatGPT in week 44:
=COUNTIFS(log!B:B,"ChatGPT",log!L:L,44,log!F:F,"Y")
/COUNTIFS(log!B:B,"ChatGPT",log!L:L,44)
Citation rate is the same with column G:
=COUNTIFS(log!B:B,"ChatGPT",log!L:L,44,log!G:G,"Y")
/COUNTIFS(log!B:B,"ChatGPT",log!L:L,44)
COUNTIFS takes pairs of range and condition and counts the rows that satisfy all of them, which is why one formula can filter on engine, week and outcome at once. Average position when mentioned is =AVERAGEIFS(log!H:H,log!B:B,"ChatGPT",log!L:L,44), and AVERAGEIFS ignores the blank cells for rows where you were not mentioned. Top-three rate adds a fourth pair: log!H:H,"<=3".
Two rules about these cells. Show the answer count beside every rate, with =COUNTIFS(log!B:B,"ChatGPT",log!L:L,44), because 40% from 90 answers is a reading and 40% from 5 answers is an anecdote. And never pool engines into one score, which is also Indig's rule: a pooled number hides a drop on ChatGPT behind a rise on AI Overviews, and the fix for each is different.
For competitors, a third sheet with =COUNTIF(log!I:I,"*"&A2&"*") beside each competitor name counts the answers that named them. Divide by your own answer count and you have their mention rate on your prompts.
How to read a change
A change is real when it is larger than the interval around both readings, and at 90 answers per engine that interval is about plus or minus 10 points. A move from 30% to 36% on ChatGPT between two weeks is nothing. A move from 30% to 55% that holds for two weeks is something.
The engines themselves move a great deal with no action from you. SISTRIX tracked 82,619 prompts for 17 weeks across six countries, including the UK, and found ChatGPT Search replaced 74% of the domains it cited from one week to the next. Google AI Mode replaced 56%. AI Overviews were far steadier. If your ChatGPT citation rate jumps about, that is the engine, not your marketing.
This is why the weekly rate is the unit, and why Batista suggests judging any change you made (a new page, a rewritten service description) no earlier than two weeks after it went live. Note the date of every change on the prompts sheet so that when a line does move you can see what preceded it.
When the spreadsheet stops being enough
The spreadsheet stops being enough when any one of three things happens: the prompt set passes about 25, you need more than two engines every week, or a second person has to read the results. Each breaks the method in a different way.
Prompt count is the first to go. At 25 prompts, 3 paraphrases and 3 runs on two engines, you are at 450 rows a week, about 15 hours. Nobody sustains that beside a day job, and the moment you skip a week the series has a hole in it. Paid tools run daily, which is also the only way to reach the 300-answer samples that narrow the interval to 5 points.
Engines are the second. ChatGPT and AI Overviews can be run by hand. AI Mode, Gemini, Perplexity and Copilot in the same morning is a different job, and each adds another 90 rows.
Team is the third. A spreadsheet one person fills in is a record. A spreadsheet three people fill in is an argument about whether "Growthcart Ltd" in a footnote counts as a mention. Tools make those judgements consistently, if not always correctly, and keep the answer text so a colleague can check.
At that point the spreadsheet has done its job. You know your prompts, which engines matter, and your baseline. You are buying a tool to continue a measurement, not to start one.
What the paid tools cost
The paid tools cost from $29 to $99 a month at entry level, and the plans that track a hundred prompts across several engines cost $189 to $399. Prices are in US dollars as published, because none of these vendors publishes a sterling price. All checked October 2026.
| Tool | Entry plan | What it includes | Next tier | Trial |
|---|---|---|---|---|
| Otterly.AI | $29/month (or $25 billed annually) | 15 prompts, 4 engines (ChatGPT, AI Overviews, Perplexity, Copilot); AI Mode, Gemini and Claude are paid add-ons | Standard $189/month, 100 prompts; Premium $489/month, 400 prompts | Free, no card |
| Semrush AI Visibility Toolkit | $99/month | 25 prompts, 1 domain, Google AI Mode and ChatGPT; extra 50 prompts $60/month, extra domain $99/month | Bundled in Semrush One plans with 50 to 200 prompts | None |
| Peec AI | Starter, reported at $95/month | 50 prompts, choose 3 of 6 engines, unlimited users | Pro, reported $245/month, 150 prompts; Advanced, reported $495/month, 350 prompts | Free trial, length not stated |
| Profound | Free 7-day trial: 50 prompts on ChatGPT, Gemini, AI Overviews | Agency Growth plan (400 credits per client workspace), price not published | Enterprise, custom, up to 9 engines | 7 days |
Peec's pricing page lists prompts and models but its prices did not render for us; the dollar figures are those Writesonic reported after checking the page in July 2026, and Peec's own comparison page gave Starter at €89 in January. Profound's page no longer shows a self-serve price; the same sources quoted $99 and $399 tiers earlier in 2026. Check both before you sign.
Set against those numbers, the spreadsheet is not really free. At 12 hours a week it is several hundred pounds of someone's time. What you get for it is the ability to read any of those pricing pages and know which line you need.
What to do with the first month
Treat the first four weeks as a baseline, not a report, which is the four-week plan Batista sets out. Freeze the prompts, log every row, and do not touch the website in response to anything you see until week three. Then read the per-engine rates with their answer counts, look at who the competitor column keeps naming, and read the snippets to see what the engines think you do.
The decision rule at the end of the month is simple. If your rates sit under 20% on prompts your customers plausibly type, the fix is content and citations, not software. If the spreadsheet already takes more than a morning a week, you have outgrown it and can buy a tool with a shopping list. If neither applies, run it another month.
If you would rather we built the prompt set and ran the first baseline for you, that is what our AI visibility service starts with. We run it for professional services providers and healthcare providers alike, since both are now compared inside AI answers. Get in touch and tell us which engines your customers use.