What have the authority sites measured recently?

Semrush published “How to build an AI visibility dashboard” on 2026-10-12. Its positioning is clear: AI search visibility cannot rely on gut feeling; it has to become a set of metrics you review on a regular schedule. The article discusses how to organize those metrics so you can keep watching them.

Ahrefs’ “AI Search ROI Is Messy” from 2026-10-02 echoes this from the other direction: the return on AI search investment is inherently hard to calculate cleanly because the path is broken. That means a dashboard with only one overall score is almost guaranteed to mislead. What you need is a set of layered metrics, so each layer can be judged on its own.

Semrush also wrote about tracking ChatGPT traffic on 2026-05-06 and about AI sentiment analysis on 2026-06-15. Read together, these pieces suggest a complete dashboard should cover roughly four things: whether AI mentions you, whether it mentions you correctly, whether anyone comes, and whether the people who come turn into business. Competitors are the fifth, used to judge relative position. For the overall definitions, start with how to define AI visibility monitoring.

Three things that should not go on a dashboard

The first is a standalone composite score. Blending mentions, citations, traffic, and inquiries into a single 0-to-100 score looks professional, but in practice nobody knows what to change. When the score rises, you cannot tell whether your content improved or the sources simply fluctuated.

The second is raw citation counts. Our strict mentions rose from 3/72 at T0 to 6/72 at T1, which sounds like a doubling; but cited domains fell from 322 to 245, with an overlap of only 90. Against that backdrop, citation count is just a highly volatile observation, not a core KPI.

The third is screenshots without timestamps. AI answers change quickly, and a screenshot with no date, no prompt, and no model is useless three months later. Every field you put on the dashboard should answer “when, what was asked, and which model.” You can self-check against the vanity metrics in a monthly report.

One more technical detail: store manually recorded fields separately from automated report fields. Manually sampled answers carry the model name, time, and original text, and that data does not belong in the same table as GA4 numbers for year-over-year comparison. My approach is one table for the content layer and a separate table for the traffic layer, presented side by side only in the monthly report, with every field’s definition and collection time written out clearly. That way, if either side is wrong, it does not contaminate the other. Clear definitions matter more than the numbers themselves.

What does each of the five panels track?

The two panels most likely to get dropped are competitors and business. Drop the competitor layer and you will not know whether your change is an absolute gain or everyone is moving. Drop the business layer and the whole dashboard becomes content navel-gazing. These two are precisely the most labor-intensive, and they should come first.

The content and source layers can be sampled by hand once a month with fixed prompts. The traffic layer relies on GA4 and Search Console. The business layer relies on an inquiry log. None of them require a paid tool to get started. To see the monitoring method for each panel, read the AI visibility monitoring checklist.

What does our own spreadsheet look like?

Our panel covers 72 answers per round, with prompts fixed by topic, language, and phrasing. T0 is 2026-09-17 and T1 is 2026-10-12. Core fields include the prompt, the full answer, where the brand appears, source links, and the model and time. Strict mentions went from 3/72 to 6/72, vendor name-drops from 0 to 2, and cited domains from 322 to 245, with an overlap of 90.

Three more data sets live in the same table: an 80.9% AIO coverage rate across 500 Chinese queries (96.7% for tutorial terms, 62.9% for technical terms); a 7-page canary indexed 7/7 at D+14 with impressions on 2/7; and a site-wide indexation rate of 100% against 12 impressions and 0 clicks over 28 days. These three map to the content, source, and traffic layers respectively, and together they show exactly where the break is. The full record is in field research and method notes.

I deliberately avoided creating an overall score. The reason, as noted above, is that these data sets contradict each other, and forcing them into one number only hides the breaks. When mentions are rising while traffic is near zero, the correct conclusion is that the entry point is not connected; blend that into “a visibility score of 62” and nobody can tell what to do. For what a report should say, see what a monthly monitoring report should include.

How do you implement this without buying tools?

Use one shared spreadsheet with five tabs

Fill the content and source layers by hand, one row per answer, with fixed fields. Copy the key traffic numbers from GA4 and Search Console once a week. Have a sales colleague fill in the business layer as inquiries come. Sample the competitor layer once a quarter. For the template structure, go straight to the SEO and GEO monthly report template.

Put the sampling routine on a fixed schedule

Run the prompts on a fixed day each month, with one person accountable. This step depends on discipline, not tools. Without an owner, even a great dashboard falls into disuse within a month. I also put this first in the monitoring checklist.

Consider tools only once volume grows

When should you buy? My rule of thumb: more than 100 prompts, or needing to watch 3 or more models at once, or producing reports weekly. Before that, manual work is cheaper. For tool comparisons, see AI search visibility tools and how to choose them, but run a full quarter first before deciding.

A low-cost implementation checklist

Three steps, and you can have the empty spreadsheet ready within a week.

1. Build a five-tab spreadsheet: content, sources, traffic, business, and competitors, with fields defined as in the table above. Set up the frame first and do not fill in data yet.
2. Fix 20 prompts and a fixed execution day: put it on the calendar and name one owner.
3. Connect only free data sources: GA4, Search Console, and an inquiry log. Do not buy any tools yet; run a full quarter.

Then look again at quarter’s end: if a panel has not moved in three months, the problem is most likely not the data but the business. To turn this into a client-ready report, read how mentions, citations, and traffic are measured.

The questions executives ask most often

Do you have to buy tools to build an AI visibility dashboard?

No. A shared spreadsheet plus GA4, Search Console, and manual sampling covers most of what you need to decide. The value of tools is saving time, not providing definitions you did not already have.

What should you most avoid putting on a dashboard?

Avoid standalone composite scores and citation counts. Even when these numbers rise, they do not map to the business, and they easily lead people to the wrong decision.

How often should you update it?

The content layer once a month, the traffic layer once a week, and the business layer in real time as inquiries come in. Updating too often amplifies noise.

What if the metrics contradict each other?

Do not merge them into one score. Keep them separate, write one sentence of your own interpretation, and wait for the next round. The contradiction is itself information: it means the evidence at one of the layers in between is insufficient.