Blog · Measurement

What belongs on an AI visibility dashboard

Published 16 September 2026 · Updated 8 October 2026 · 9 min read

What belongs on an AI visibility dashboard
Photograph: Cole Freeman / unsplash

The short answer: Six things earn their space. The prompt set you are measuring, your mention rate with the number of runs behind it, which sources the engines actually cited, whether the claim beside your name was true, the same figures split per engine, and a dated log of what you shipped. Everything else is decoration. A single blended visibility score with no run count under it is the one tile to delete first.

The dashboards clients have shown us during audits are mostly built to be looked at rather than argued with. We can only speak for the ones we have been shown, not for the category. The difference shows up the first time a number moves and somebody asks why.

The six tiles that earn their space

The prompt set itself. Not a count, the actual questions, visible and editable. If a reader cannot see which forty or fifty prompts produced the number, the number is unreadable. Prompt sets drift as buyers change vocabulary, and a dashboard that hides the list hides the drift too.

Mention rate, with n. The share of runs in which your brand was named, printed next to how many runs that was. Twelve per cent of 450 runs is a measurement. Twelve per cent of eight runs is a coin toss wearing a percentage sign.

Citation source mix. Which domains the engine leaned on when it answered. This is the tile that tells you where the work is. If nine of the top ten cited domains for your category are review sites and forums you do not control, your problem is corroboration, not copy, and the rented versus owned split decides how you spend the next quarter.

Claim accuracy. Of the runs that named you, how many described you correctly. Pricing, category, geography, what you actually sell. Being named alongside a wrong sentence is worse than silence, and we have not yet been shown a dashboard that tracks it, because it needs a human to read the answer.

Per-engine split, never averaged. Perplexity, ChatGPT, Claude, Copilot, and Google's two surfaces, AI Overviews and AI Mode, each as its own row. These are separate systems with separate retrieval, and nothing in their published documentation ties one engine's citations to another's, so there is no reason to expect one row to stand in for the rest. An average across five engines hides the only useful finding, which is that you are fine on one and absent on another.

A change log. Date, what shipped, which prompts it was meant to affect. Without it you have a line chart and a story, and the story is usually wrong.

The number most dashboards get wrong

Here is a check you can run this afternoon. Open any dashboard that shows you AI Overviews or AI Mode clicks sourced from Google Search Console, and ask where that figure comes from.

Google's own documentation on AI features says that sites appearing in AI features are included in overall search traffic in Search Console, reported in the Performance report within the Web search type. Since June 2026 Search Console has also carried a dedicated Generative AI performance report, rolled out to all sites worldwide on 31 August 2026, which reports impressions for AI Overviews and AI Mode on their own. So the question to ask of a tile is which report it came from.

That matters because the dedicated report exposes impressions, not a clicks-by-surface breakdown. A tile labelled "AI Mode clicks" is therefore still somebody's model unless its method is stated, and it should be labelled that way on the dashboard. We have seen such a tile presented as reported data in client decks more than once, and nobody in the room knew the difference until it was checked.

The same documentation is worth reading for a second reason. It states that a page must be indexed and eligible to be shown in Google Search with a snippet, and it lists nosnippet and max-snippet among the controls that limit what Google shows from a page. On Google's surfaces, then, those directives are a visibility control, and they belong on your dashboard as a checked fact rather than an assumption. Other engines document their own rules, and this one is Google's.

Split your bot log by user agent, not by the word "AI"

Server logs are the half of the picture that nobody can spin, and most dashboards flatten them into one tile called AI crawler traffic. That tile mixes unrelated events into one line.

OpenAI publishes four user agents separately, and they mean different things. GPTBot crawls content that may be used for training foundation models. OAI-SearchBot surfaces sites in ChatGPT's search features. ChatGPT-User is the one that fires when a person asks ChatGPT a question and the assistant goes and fetches a page. OpenAI's documentation notes that ChatGPT-User is not used for crawling the web automatically. The fourth, OAI-AdsBot, checks landing pages submitted as ads against OpenAI's policies and goes nowhere else, so it belongs in your log view but not in the same tile as the other three.

That last distinction is the whole value of the tile. A spike in GPTBot tells you a training crawl ran. A spike in ChatGPT-User on a specific URL tells you a human asked something and an engine went to your page to answer it. One is supply. The other is demand. Charting them on the same axis destroys the only signal worth having.

The volumes involved are not small. Cloudflare's crawler analysis published on 1 July 2025, covering May 2024 to May 2025, put GPTBot at 7.7 per cent of observed crawler share in May 2025 against Googlebot's 50 per cent, and reported ChatGPT-User requests up 2,825 per cent across that year. Those shares are what Cloudflare observed across its own network over that window, not a census of the web, and your logs may look different. Split by agent, verify against the IP ranges OpenAI publishes, and the tile becomes evidence instead of a vibe.

Every tile needs a denominator and a date

A visibility figure without a run count, a prompt count and a timestamp is not a metric. It is a claim. In our own audit runs, which put a prompt set of about forty five prompts through five engines with retrieval on and retrieval off, we have watched the same prompt return different sources minutes apart, so a tile built on one run per prompt per week carries a lot of sampling noise rendered in brand colours. We have not published a stability rate for that, and the figure would be specific to the prompt set anyway.

Three labels fix most of this. How many runs. Over what window. Retrieval on, off, or both. Add those and a stakeholder can tell the difference between a real movement and the engine having a different morning. How to measure AI visibility without fooling yourself goes through the sampling arithmetic in more detail.

Be careful with the headline share figure too. Share of voice in AI answers is useful for direction and poor for precision, because the denominator is a prompt set somebody chose, not a market.

What to leave off

Cut the blended visibility score, or demote it to a footnote. It compresses five engines, two retrieval modes and a mixed prompt set into one digit, and the compression is where the information went.

Cut branded prompts from the headline. Of course you appear when the question contains your name. Keep them in a separate view as a sanity check on tracking, not as evidence of reach.

Cut sentiment as a primary number. Tone is cheap to score and rarely actionable. Accuracy is expensive to score and tells you what to go and fix.

Cut competitor leaderboards that rank you against named rivals on an averaged score. They generate meetings, not fixes.

What a dashboard still cannot tell you

Two honest gaps, and both belong printed on the page rather than buried.

Attribution is mostly broken. In the client logs we have looked at, visits that followed an assistant answer frequently arrive with no referrer naming the assistant, so the mention and the signup cannot be joined up. We cannot see from outside what every surface sends, only what reached those logs. Treat an AI-sourced revenue figure as a model unless somebody can show you the join.

Prompt volume is not something we can see. We have not found an engine that publishes how often a given question is asked inside ChatGPT or AI Mode, and we have found no primary source for it, which is not the same as proving none exists. Estimates circulate, they are estimates, and a dashboard should say so.

If you are commissioning this work rather than building it, how to buy an AEO audit without being sold a dashboard covers the questions that separate evidence from interface. Our own audit ships one deliverable, a report, and the sample audit report shows how the evidence sits under each finding.

Honest limits

This post does not establish what the category as a whole does, only what we have seen in the dashboards clients brought to audits, which is a small and self selected set. The Search Console point rests on Google's published documentation, so it holds for Google's surfaces and says nothing about how other engines report. The referrer and prompt stability observations come from client logs and our own audit runs rather than a published study, and we give no rates for either because we have not measured them in a way an outsider could reproduce. Absence in our logs is absence from our logs, not proof that an engine never sends a referrer. Cloudflare's crawler shares are one network's view of its own traffic over one year, so treat them as direction rather than as your own numbers.

FAQ

Does Search Console show AI Overviews or AI Mode separately? Yes, for impressions. Alongside the overall Performance report, Google has run a dedicated Generative AI performance report since June 2026, available to all sites worldwide since 31 August 2026. It reports impressions rather than clicks, so a clicks-by-surface tile is still modelled rather than reported.

What is a dashboard visibility score, and can I trust it? It is normally one number compressing mention rate, citation share and sometimes sentiment across several engines and a prompt set somebody chose. Nothing standardises it, so two products can print different scores from the same week without either being wrong. Treat it as a direction of travel, ask for the formula and the run count behind it, and read the per-engine rows underneath before you act on the headline digit.

How many runs per prompt before a number means anything? Enough that one odd run cannot move the percentage visibly. In practice that means multiple runs per prompt per engine per window, and the dashboard printing the count so the reader can judge for themselves.

Should server logs be on the same dashboard as prompt results? Yes, and split by user agent. Prompt results tell you what the engine said. Logs tell you whether it could reach you at all, which is the cheaper problem to fix.

Is a weekly refresh often enough? For most sites, yes. Engines change under you, but a daily refresh on a small prompt set mostly surfaces noise, and teams start reacting to it.

Sources

Related reading

How to measure AI visibility without fooling yourselfShare of voice in AI answers, and its limitsHow to buy an AEO audit without being sold a dashboardRented visibility vs owned visibility in AI searchAttribution when there is no clickProving a change worked: before and after in AI search