You can track AI visibility manually, and you should do it by hand at least once before paying for anything. The method is not secret: a fixed set of buyer questions that never contain your brand name, each run more than once, with only the sources marked inside the answer text counted as citations. What a paid tool buys you is scale, a steady cadence and a stored record of every answer. It does not buy you a measurement you could not make yourself.
Why doing it by hand first is worth it
A manual round teaches you what the numbers mean. Once you have read a few dozen answers yourself, you know what a mention looks like, how often a competitor you had never heard of appears, and how much the same question drifts between runs. That makes you a much harder customer to mislead later, because most of the ways an AI visibility number goes wrong are choices in the method, not limits of the software.
Two of those choices do most of the damage. The first is asking questions that contain your brand. Ask an assistant whether a named company is any good and it will name that company, so the question proves nothing. In one real account, a reported 13% visibility came entirely from a single brand-named question; the unprompted figure was zero. The second is counting every page the assistant retrieved as a citation. An assistant typically retrieves many pages and cites a few, and counting the retrieval pool inflates the result several times over.
A manual template you can copy
- Write the question set once. Full sentences with a situation in them (a budget, a constraint, a use case), not keyword fragments. No brand names, yours or a competitor's. Note the intent of each question: research or purchase. The guide on building a prompt set covers this step in detail.
- Freeze it. The question set is the denominator of every percentage. Adding easier questions later raises the average while nothing has improved.
- Run each question more than once per engine, in a fresh session, and paste the full answer into your sheet rather than a summary.
- Record every brand the answer names, not only the ones on your list, and merge spelling variants of the same brand before counting.
- Record citations only where the answer marks them in the text, and keep "we were cited" separate from "we were recommended". Your page can be the source for an answer that recommends a competitor.
- Leave failed runs blank. A timeout is not a non-mention, and writing it as zero turns a normal month into a fake decline.
date | engine | question | intent | run | brands named (in order) | we were named? | cited URLs (in-text only) | we were cited? | note
-----+--------+----------+--------+-----+-------------------------+----------------+---------------------------+----------------+-----
| | | | 1 | | yes / no / - | | yes / no / - |
| | | | 2 | | | | |
# "-" means the run failed or timed out: leave it out of the denominator.
When you summarise, report each engine on its own line before any average, and write the question count and the number of distinct brands next to your share. A small percentage among many named brands is a different situation from the same percentage among a handful. For how to keep this tracking honest from month to month, see measuring change over time.
What a paid tool actually adds
The method stays the same; the labour changes. A tool earns its fee in three places:
- Scale. Many questions across several engines, run repeatedly, is a lot of copying and pasting. By hand, people quietly shrink the set, and a thin sample produces a share figure that swings on the next run.
- Cadence. Change only means something when the same question set is run the same way on a schedule, with an alert only after a change shows up in two consecutive runs.
- Evidence. Stored answers let you go back to the exact response behind a number, and recompute a report weeks later so it still adds up. A spreadsheet of summaries cannot do that.
If a tool cannot tell you whether it excludes brand-named questions, which citations it counts and what it shows when a run fails, the manual sheet is the more trustworthy of the two.
When manual tracking is enough
A small question set, one or two engines that matter to your buyers, and a check every month or so is a reasonable manual workload. It stops being reasonable when you add engines, markets or a client who wants a trend line. The walkthrough on tracking brand mentions in ChatGPT shows the manual method on a single engine end to end. When you are ready to compare that against a tool, InsightWonder's public tools at tools.insightwonder.com are one place to start.