Build a weekly AI visibility report in Notion with Claude

Four things: presence, recommendation, who else was named and which sources were cited. In Linkeddit's definitions, presence is the share of usable answers that name your brand or cite your URL, and recommendation is the share where the brand was recommended, so recommendation can never exceed presence. On the test project, ChatGPT named the brand in 16 of 18 usable samples and recommended it in 15.

Share of voice and cited domains explain the number. In the same segment the brand held 20.5% of vendor visibility and Fathom 16.7%, and Matomo's own domain was the most cited at 10 of 183 domain citations. A report that stops at presence tells you that you are named but never who sits next to you.

Notion's template gallery shows what manual versions look like. One AI visibility template has you log each check by hand, with a column for who the engine named instead of you and which sources it pointed to. That works for a few prompts. It stops working when you track dozens of questions across five engines every week, which is why the numbers should arrive from a measured source and only the layout is yours.

Because a missing engine changes what the rest of the report means. In the latest run on the test project, 13 of 60 samples failed, the run is marked partial, and Perplexity returned 21 errors with no usable sample. Linkeddit reports that as unmeasured with a reason. A summary that quietly averages four engines and calls it overall visibility would be wrong.

Google AI Overviews shows a second kind of gap. Google displayed an Overview in 16 of 21 samples, so five samples were absent, not failures. Presence of 100% across 15 usable samples is true, but it describes the questions where Google showed an Overview at all. Put appearance rate in its own column so a reader sees both facts.

Tell Claude in the prompt to keep both. The sentence that matters is: report an engine with no usable samples as unmeasured, never as zero.

Fewer than most dashboards imply. The Keystone Ops AI visibility template on Notion's marketplace makes the point directly: ask an LLM the same question twice and the set of names in the answer often changes, so it flags one run as a coin toss, two or three as directional and four or more as a number worth acting on. Treat that as one practitioner's rule of thumb rather than a standard.

Apply it to the report by storing the denominator next to every rate. Claude's presence of 0.5 on the test project is 5 of 10 usable samples, so a week-over-week move from 0.5 to 0.6 is a single answer flipping. Ask Claude to quote both denominators whenever it describes a change, and to say plainly when the change is inside that noise.

Use one database, one row per engine per week, instead of a new free-text page each time. Notion's connector can create a database and its pages, and can update a page after the fact, so Claude can add last week's row and leave earlier rows untouched. That keeps filters and charts working, which a pile of narrative pages does not.

Peec AI published the pattern in one prompt: pull last week's visibility with the percentage change, sentiment by model, share of voice against the top three competitors and the five prompts that moved most, then save it as a brief to a Notion page. Linkeddit's version differs in two ways. It carries the measurement disclosure, so you can see the engine, the sample count and what was excluded, and it leaves the written analysis to your own Claude.

Keep a free-text page for the reading and the database for the numbers. The page can link to the row, and Claude can add a note to it later with update-page.