How I Measure AI Visibility Before I Trust a Dashboard

I would measure AI visibility as a distribution across engines, personas, and repeated runs, and I would buy a dashboard only after a manual pass had taught me which number I am allowed to act on. The Machine Relations Index has monitored 15,782 answer runs across six engines since May 10, 2026, and it now observes 22,179 cited domains. A single score on top of a system that answers the same prompt differently each session is incomplete in a way that makes it easy to misuse.

The metric set I would want a vendor to map their product onto is The Seven Metrics That Measure AI Search Visibility. That page is the deeper method. This essay is how I would use it as a founder: which layer I would collect first, how I would connect a number to an action, and what would make me pay for scale.

The same prompt is not a measurement

Run one prompt in ChatGPT three times and the answers can differ. Sometimes the brand is cited, sometimes named, sometimes missing, sometimes linked. Cassie Clark's guidance is to run the same prompt 10 times in a row and track inclusion. I would do that, and I would treat the need for ten runs as a description of the system. Models retrieve, synthesize, and generate on each pass. "Position X for keyword Y" does not survive the trip. The question that does survive is whether, across a real number of queries, personas, and sessions, an engine uses our material as evidence when a buyer asks something in the category.

That is share of citation. Share of voice counts mentions. Share of citation counts use as a source. I would report both and refuse to let the first one stand in for the second. The definitions I would put in the footnote live in the glossary. The stack the measurement has to point back through is on Machine Relations.

Layer one: every engine, then the persona

Presenc.ai's 2026 metric analysis found that Perplexity rarely refuses to name brands, Claude hedges often, Google AI Overviews leans on traditional ranking signals more than the others, and ChatGPT sits between them. An average across those four hides the engine where we are present and the engine where we are absent. I would keep four series.

For each engine I would record citation rate (a URL on our domain), mention rate (the name, with or without a link), and position. First-position citations earn 4 to 5 times the click-through of fifth-position citations, on Presenc's figures. Position inside an answer is a different fact from being somewhere in it.

On volume, Citare recommends 50 to 150 queries per buyer persona and 3 to 5 personas, which is 600 to 3,000 dispatches in a cycle. Ten queries on two engines is a sample of noise. I would not call it a visibility check.

Persona variance is large enough to change a pipeline story. Citare published a case in which one B2B brand surfaced at 38 percent when the prompt came from a CTO persona and at 4 percent from a CMO persona. Same brand, same engines, same week. The buyer's framing changed the recommendation. A system that does not segment by intent is hiding the asymmetry that decides whether the right buyer sees us.

The Index's 15,782 runs are the scale at which I trust a market-level rate. They are not a substitute for our 20 revenue queries. I would want both, labeled.

Layer two: tie the number to the input that moves it

Frameworks on the market sort metrics into mention rate, citation share, sentiment, prompt coverage, and traffic attribution. Semrush tracks 239 million prompts and sells an overview, prompt tracking, narrative drivers, competitor research, and perception tools. The volume is real. The test I would apply is whether a specific number on that screen traces to a specific action.

The three inputs I would require the measurement to point at:

1. A third-party editorial source that corroborates the claim. Muck Rack's May 2026 analysis found 84 percent of AI citations coming from earned editorial coverage. The visibility score sits downstream of that coverage.
2. Entity resolution. If the engine cannot tell us from three neighbors, the citations that do occur are hard to steer.
3. Extraction. Answer-first paragraphs, claims with sources, sections that can be lifted. This is what decides whether a citation is still there next session.

A dashboard that shows the number and cannot say which of the three is weak is a thermometer. I would pay for a diagnostic. Measurement works when it traces back through authority, entity, architecture, and distribution. A score with no input is a number I can recite and cannot manage.

Layer three: the attribution chain, kept as a chain

AI answers create value with no click to attribute. Semrush reports AI referral visitors worth 4.4 times organic search visitors. A large share of the value happens before the click. A buyer asks who to hire. The engine names three companies. The buyer searches the one that sounded credible. That branded query lands as organic or direct. The citation that caused it is invisible in the default report.

I would keep four proxies and call them proxies.

Server logs for GPTBot, ClaudeBot, PerplexityBot, and Bingbot on every property we own. Crawl frequency and depth are leading indicators. If an engine stops crawling, I would expect citations to thin, and I would look at robots.txt before I looked at the copy. Citare's guide is right that most brands have never checked crawler access.

Branded search in Google Search Console, set beside citation movement. The lag I would test is two to four weeks: citation lift on a query cluster, then branded searches for the company plus the category term. Correlation, not causation, and the strongest buyer-side signal I have found that does not require the click to be labeled.

Pages engines cite, whether those pages can convert, and whether conversion on them moves after a citation lift. Again a correlation. The pattern worth testing is that a page engines cite converts differently from a page with similar organic traffic that they do not cite. The engine's selection is a trust mark the buyer can carry into the visit.

None of this is a closed identity. Four correlated signals beat one dashboard number with no path back to a page.

The table I would not collapse

Metric	What I would let it mean	What it leaves out	When I would look at it
Mention rate	The name appeared	Whether the engine used us as evidence	Baseline: are we in the answer at all?
Citation rate	The engine linked our domain	Whether anyone acted afterward	Are we being used as evidence?
Share of citation	Us against the others on a defined pool	Why one competitor leads	Where we stand on that pool
Citation position	First, third, fifth in that response	Whether first place holds next session	Is the slot improving?
AI share of voice	Mention frequency against competitors	Whether mentions become trust	A board slide, dangerous as a strategy
Recommendation rate	Inclusion across repeat runs	The cause of inclusion or exclusion	Is presence stable?

I would run the full table, per engine, per persona, at least monthly. One cell, reported quarterly, is how a team congratulates itself on a number that moved for a reason it cannot name.

What each public guide hands me, and what it leaves on my desk

I went through the guides that were current when this measurement question got loud, and they agree on the instruments and stop before the cause.

Cassie Clark's useful instruction is the ten consecutive runs. I would do them on one revenue query and keep the ten answers. The spread is the finding. A single green cell in a dashboard is the thing those ten runs are there to humble.

Presenc separates the engines: Perplexity rarely refuses to name a brand, Claude hedges, Google AI Overviews lean on ranking signals, ChatGPT sits between them. The same analysis says a first-position citation earns 4 to 5 times the click-through of a fifth-position citation. I would keep position as its own column. I would not let an average "visibility score" mix a Perplexity mention with a Claude hedge and call it one market.

Citare's volume guidance is the one that makes the afternoon look small on purpose. Fifty to 150 queries per persona, three to five personas, 600 to 3,000 dispatches. And the case that justifies the tedium: one B2B brand at 38 percent surface rate for a CTO persona and 4 percent for a CMO persona, same engines, same week. If I only run the prompt the way I would ask it, I will measure my own curiosity and miss the buyer.

Semrush's 239 million tracked prompts, and the report that AI referral visitors are worth 4.4 times organic visitors, tell me the category of tooling is large and that the click, when it happens, can be valuable. Muck Rack's May 2026 analysis, 84 percent of cited links from earned media (its term for sources brands neither own nor pay for, with journalism at 27 percent), tells me where the cited links come from, which is a different question from why a score moves. AI Search Tools' 4 to 6 week window is the shortest experiment I would trust once I am changing pages rather than only observing them. BrightEdge sits with Semrush at the enterprise end. Otterly.AI and Peec AI sit at the end where a small team can automate the 20 queries after the spreadsheet has taught us what a row means.

The shared omission is the trace from the number back to one of three inputs: a third-party source, a resolved entity, or a page that can be extracted. I would supply that trace myself. For every one of the 20 queries where we are cited, the cited URL is the asset to protect. For every query where we are absent, the question is whether any crawlable, editorially credible third party corroborates us. The Index can tell me the market is measurable at 15,782 runs and 22,179 domains since May 10, 2026. It cannot tell me which of our pages those runs would have used. The bots in the log, GPTBot, ClaudeBot, PerplexityBot, and Bingbot, are the leading indicator I would watch while the 30-day repeat is still in progress. A crawler that stops showing up is a reason to read robots.txt before I rewrite the headline.

The two-to-four-week branded-search lag is a correlation I would test in Search Console, not a law I would brief as causation. Citation lift on a cluster, then branded queries for the company plus the category. If it appears, I have a buyer-side echo of a citation that never sent a labeled click. If it does not appear, the citation can still be real and the echo can be elsewhere. I would write down which one happened. Four proxies, none of them an identity, are still a better decision than a quarterly share-of-voice tile with no path to a page.

The decision

I would start with 20 buyer queries, the questions a prospect asks while deciding who to hire. I would run each through ChatGPT, Perplexity, Gemini, and Claude. I would record mention, citation, position, and whether the answer recommends us. That baseline is an afternoon. I would repeat the same 20 in 30 days. The comparison tells me more about our revenue queries than a tour of someone else's 239 million prompts.

Then I would trace backward. For every citation, I would name the source: our site, a third-party article, a news mention. That source is what I would protect. For every absence, I would ask whether any crawlable, third-party, editorially credible source corroborates us on that topic. If none does, that topic is the next earned-media target, which is what Muck Rack's 84 percent says the score is made of.

I would buy automated monitoring after the manual pass, when I can tell whether the tool saves the afternoon or adds a series I cannot collect by hand. Semrush and BrightEdge are the enterprise ends of the market. Otterly.AI, Peec AI, and a spreadsheet are the ends I would use while the query list is still 20. A manual pass, done this way, carries most of the decision. One earlier accounting put that share near 80 percent. The subscription buys repetition at Citare's 600 to 3,000 dispatches, which I would not do by hand.

Scope: the 20 queries, four engines, a 30-day repeat, crawler logs, and Search Console branded queries. The Index's 15,782 runs and 22,179 domains stay the market context. They do not replace the 20.

What would change my mind: a tool that stores the prompt, the persona, the engine, the date, the answer, and the cited URL, and that shows me which of the three inputs moved when the score moved. I would pay for that earlier. A cleaner average across Perplexity's willingness to name brands and Claude's habit of hedging would not qualify, because Presenc's whole point is that those behaviors differ. If a 30-day repeat on the 20 queries was pure noise, inside the band of ten-run variance Cassie Clark's method is designed to expose, I would lengthen the window toward the 4 to 6 weeks AI Search Tools recommends for a content experiment. I would not shorten it to please a weekly report.

FAQ

How often would I measure?

Monthly for a decision about spend. Weekly while an earned-media push is in market and I need to see whether new citations land. The AI Search Tools guide recommends 4 to 6 week windows for a controlled content experiment. Anything shorter than 30 days, on a system that disagrees with itself across ten runs, is a weather report.

What is the difference I would refuse to blur?

Share of voice counts mentions. Share of citation counts linked use as evidence. One is exposure. The other is trust. Semrush's 4.4 times referral value and Presenc's 4 to 5 times click-through for first position both sit downstream of being used, not of being vaguely present.

Which tool would I buy first?

None, until the 20 queries have been run by hand on ChatGPT, Perplexity, Gemini, and Claude. Then I would buy the tool that removes that labor or adds persona scale I cannot fake, in the Citare range of 50 to 150 queries across 3 to 5 personas. I would judge Semrush, with 239 million prompts tracked, on whether our 20 are in the product in a form I can audit.

Can I do this without a subscription?

Yes. The queries, a spreadsheet, the four bots in the logs, and branded-query trends in Search Console. The 38 percent versus 4 percent persona case is the reason I would tag the prompt with a buyer even in the spreadsheet. The paid product is how I would get to hundreds of dispatches. It is not how I would learn what a dispatch means.

What would I do with a number I cannot trace?

I would leave it out of the decision. Muck Rack's 84 percent says the score follows earned coverage. If the report cannot tell me whether the weak input is that coverage, the entity, or the extraction, the number has not earned a budget change. The Index can tell me the market is measurable at 15,782 runs. Our afternoon with 20 queries tells me whether we are.