What I Track Before I Buy Measurement Software

I would track five things before I buy AI visibility software: whether engines name the brand, whether they cite a source, whether the description is accurate, whether that pattern repeats across prompts and dates, and whether the sample is large enough to change a decision. A dashboard can show a presence score. I would buy the dashboard after I know which of those five the score is allowed to change.

The method I would read first is AI Visibility Measurement: What CEOs Should Track First. That page is the deeper method. This essay is how I would use it as a founder: what I would collect by hand, what I would refuse to collapse into one number, and what would make me sign a software contract.

Presence, citation, and accuracy are three jobs

Presence asks whether ChatGPT, Perplexity, Claude, Gemini, Google AI Mode, or AI Overviews mention the brand at all. Citation asks whether those systems use the site, earned media, research, or another third-party source as evidence. Accuracy asks whether the answer describes the company, the category, the claims, and the competitors correctly.

A brand can be mentioned without being trusted, trusted in prose without a citation, or cited and still misdescribed. I would keep the three apart. The founder who folds them into one visibility score loses the operating lesson, because each failure asks for a different next action.

The IAB's 2026 guidance on measuring visibility in the AI era pushes disclosure around AI-powered discovery measurement: platform coverage, prompt construction, data collection, attribution logic, sentiment logic, and accuracy classification. I would use that list as a purchasing checklist. If a vendor cannot say which engines, which prompts, how the run was collected, and how accuracy was classified, the chart is ahead of the method.

Machine Relations research on decision-grade measurement draws the same line in operating language. Directional measurement tells me what to investigate. Decision-grade measurement tells me what to change. I want both, and I want the label on the page I am looking at. The vocabulary for those labels is in the glossary. The order I would put under the measurement, from authority through the record I would keep, is on Machine Relations.

One prompt is a screenshot

AI answers move across sessions, wording, retrieval paths, engine changes, and time. The 2026 paper "Do not Measure Once: Measuring Visibility in AI Search" argues that brand visibility in generative search should be treated as a distribution. If a brand appears in one run and disappears in the next five, the lesson is instability.

Habit	What I would learn	What I would still be missing
One executive prompt	A single answer	Repeatability across queries
One engine	How one model answered today	Engine-by-engine variance
Brand-name prompt only	Whether the model can say the name	Whether buyers meet the brand in category research
A screenshot	That an answer existed	Source path, citation quality, and drift
One aggregate score	A number for a board slide	Which layer is broken

I would build the prompt set from the way the buyer asks: category questions, comparison questions, problem questions, and "who should I trust" questions. I would run it across the engines the buyer uses. I would repeat it. I would separate a brand mention from a source citation. I would record whether the cited source is owned, earned, third-party research, marketplace content, or a scraped page. That is the point at which the measurement can change work.

Citation rate, with a floor

The cleanest metric I would defend is citation rate: how often an answer engine cites the source, brand, or domain across a defined set of observed runs. Machine Relations Index v2 reports a source-segment citation rate only after a segment clears an evidence floor of at least 10 observed runs across at least 7 distinct run dates. Thinner segments stay in collecting status.

I would borrow that floor even when I am reading a vendor dashboard that does not publish one. One ChatGPT mention is a reason to look again. Two Perplexity citations of a competitor are a reason to look again. Neither is a reason to reallocate the quarter. I need the rate, the source type, the engine, the segment, and the confidence tier.

Metric	Question it answers	Decision I would allow it to inform
Brand presence rate	How often are we named?	Are we in the consideration set at all?
Citation rate	How often are our sources cited?	Does our material get used as evidence?
Source mix	What kind of pages get cited?	Do we need earned media, research, product documentation, or owned pages?
Accuracy rate	How often are the claims correct?	Is the description stable or distorted?
Prompt repeatability	Does the answer survive variants?	Is the visibility a pattern or a prompt artifact?
Confidence tier	Is the sample big enough?	Can I change spend from this, or only investigate?

A tool that gives me the first column and cannot explain the second and third is still useful. I would buy it as a collector. I would not let it set the budget.

Software shows the gap. It does not close it

Adobe's AI Visibility documentation describes brand visibility across AI-powered search and discovery. Amplitude's AI visibility documentation frames the category around visibility scores, competitor rankings, and recommendations for brand presence in AI-generated answers. Both are measurement products. A score does not create the source an engine will cite. A recommendation inside the product does not place the company in a publication the engine retrieves. A competitor ranking does not repair a weak description.

The Artificial Intelligence Visibility Index dataset on Zenodo covers 2,729 businesses across five generative AI search systems. Measurement at that scale is where the market is going. Scalable measurement still leaves the operating work: sources worth citing, the same claim repeated across the site, earned media, the founder profile, product pages, and third-party coverage, and a category a machine can parse.

I would treat the software as the record of that work. I would keep the work.

Five questions I would answer before I take a demo

Which buyer questions should we own?

I would begin with category, comparison, problem, and trust questions. The buyer starts by asking who solves the problem.

Which engines does this buyer use?

An enterprise sale may need ChatGPT, Perplexity, Gemini, Claude, Google AI Mode, and AI Overviews. A local or consumer company may weigh Google's AI surfaces differently. I would measure the engines the buyer uses, and I would say so in the scope.

Are we mentioned, cited, or recommended?

A mention is awareness. A citation is evidence. A recommendation is selection. I would report them as three series.

Which sources explain the answer?

Owned page, earned media, research page, review site, directory, forum. If the model names a competitor because a third party explains them more clearly than our pages explain us, that source path is the work.

Is the signal repeatable enough to act on?

Multiple prompts, multiple dates. If the sample sits under something like the Index floor of 10 runs across 7 dates, I would call it collecting. I would not take it to a board as a settled rate.

The stack I would keep, software or not

Layer	What I measure	What I would accept as good
Query set	Buyer questions by intent	Category, comparison, problem, and brand prompts
Engine coverage	The same set across surfaces	The engines this buyer actually uses
Presence	Brand named in the answer	Consistent inclusion on qualified category prompts
Citation	Source or domain cited	Owned and earned sources used as evidence
Accuracy	Claims match the record	Correct category, product, market, and differentiators
Source architecture	Pages behind the answer	Third-party coverage plus pages a machine can extract
Confidence	Repeated observations	Enough runs and dates to act, with thin samples labeled

SEO asks whether a page ranks. GEO asks whether AI systems can cite the content. AEO asks whether an answer can be selected directly. I would use Machine Relations as the frame that asks whether the whole record makes the brand legible, trusted, retrieved, cited, and recommended. Software sits inside that frame. The frame does not end at the login screen.

How I would spend a demo hour

I would take the IAB's 2026 disclosure list into the demo and spend the hour on it, not on the chart. Six items: which engines are covered, how the prompts were built, how the run was collected, what the attribution logic is, what the sentiment logic is, and how accuracy was classified. A vendor who can answer those six has a measurement I can interrogate. A vendor who can only show a competitive screenshot has a board prop. I would not let the prop win the budget.

I would also bring the floor I am willing to act on. MRI v2 withholds a source-segment rate until a segment has at least 10 observed runs across at least 7 distinct dates, and leaves thinner slices in collecting status. If the demo's headline segment cannot show me the run count and the dates, I would label it collecting in my notes even if the product uses a brighter word. The Zenodo index, 2,729 businesses across five generative search systems, tells me measurement at scale is already a published dataset, DOI 10.5281/zenodo.20076405. Scale in the market is not a reason to treat my thin slice as settled. It is a reason to expect the vendor to disclose the slice.

Adobe's documentation and Amplitude's documentation both describe scores, rankings, and recommendations for how a brand shows up in AI answers. I would use them as examples of the category, and I would ask which of my five questions the score is allowed to change. Presence, citation, accuracy, repeatability, confidence. A score that collapses the five will always look more decisive than the underlying runs. The 2026 paper on measuring visibility, arXiv:2604.07585, is the one-sentence rebuttal I would keep in the notebook: treat visibility as a distribution. One executive prompt, one engine, a brand-name query, a screenshot, and an aggregate number are the five habits I would retire before I signed. Each of them produces an artifact. None of them produces the source mix, which is the column that tells me whether the next dollar is earned media, research, documentation, or an owned page.

I would leave the demo with an export or with a no. The export I want is the prompt, the date, the engine, the answer text, the cited URLs, and the definition behind the score. If the contract ending means that export ends, the company has rented its memory. Directional measurement can tell me what to investigate. Decision-grade measurement can tell me what to change. I would write which one I was looking at on the first page of the notes, because the IAB list and the Index floor exist so that distinction can be checked by someone who was not in the demo.

The decision

I would spend the first measurement cycle on a prompt set, a repeat cadence, and a notebook that separates presence, citation, accuracy, source type, and confidence. I would buy software only after I can point to which column is weak.

Scope: the buyer questions I would be willing to fund work against, the engines those buyers use, and a horizon long enough to clear a floor I can defend. I would use 10 observed runs across 7 distinct dates as that floor, because that is the bar MRI v2 already publishes for a source-segment rate. Under the floor, the status is collecting.

If the weak column is engine coverage, I would buy the tool with the best prompt and engine breadth. If it is source attribution, I would buy the tool that exposes citations and the answer evidence. If it is accuracy, I would buy the tool that tracks a distorted claim. If the weak column is authority, I would put the money into earned media, research, and pages worth citing, and I would keep the software in a measurement role.

What would change my mind: a vendor that stores the exact prompts, dates, engines, answer text, citations, and the definition behind every score, and that lets me export them when the contract ends, could earn the budget earlier. I would still want the five questions answered in the export. A cleaner chart, by itself, would not move me. If my own repeated runs, past the 10-by-7 floor, contradicted a vendor's headline score on the same prompts, I would trust the runs I can re-read and I would drop the score.

The IAB list and the Zenodo scale of 2,729 businesses tell me this category will keep selling measurement. I would buy the measurement that changes what I build.

FAQ

What would I track on day one?

Presence, citation, and accuracy, on buyer questions rather than on the brand name alone, across the engines the buyer uses. The arXiv paper on measuring visibility argues for a distribution across runs. I would keep the first week's screenshots and refuse to treat any one of them as the rate.

What sample would I require before I change spend?

I would borrow the MRI v2 floor: at least 10 observed runs across at least 7 distinct run dates for a segment I am willing to call published. Below that I would keep collecting. The Zenodo index covering 2,729 businesses shows the market can measure at scale. Scale does not rescue a thin slice of my own category.

How would I choose among software products?

I would match the product to the weak column. Adobe and Amplitude both document visibility scores and competitive views. I would buy breadth when coverage is the gap, citation evidence when attribution is the gap, and claim-level accuracy when the description is the gap. I would not buy any of them to create the source the engine cites.

Where do GEO and AEO sit in this decision?

They are layers I would measure, not substitutes for the record. GEO is whether systems can cite the content. AEO is whether a direct answer can be selected. I would fund either one only after the prompt set shows that extractability, rather than missing third-party evidence, is the constraint.