What I Build Before I Buy Visibility Software

I would not buy visibility software first. I would write the questions a buyer asks before they know our name, and I would build a source that can answer each one with a claim someone else can corroborate. Software can show whether ChatGPT, Perplexity, Gemini, Claude, or Google names the company. It cannot invent the evidence those systems need before they do.

The comparison is AI Visibility Tools for Startups: What to Track First. That page is the deeper resource. This essay is how I would use it as a founder: the order I would build, the moment I would start paying for measurement, and the result that would justify buying the dashboard earlier.

The counts that set the order

Muck Rack's May 2026 Generative Pulse analyzed 25 million citations across ChatGPT, Claude, and Gemini. Earned media was 84% of AI citations. Paid and advertorial content was 0.3%. Journalism alone was 27% of cited sources, across more than 20,000 outlets (Muck Rack). I would underprice a lot of things before I underpriced that. A startup with no third-party page is asking an engine to trust a self-description, in a corpus where the cited URL is usually not the brand's own site.

BuzzStream and Citation Labs analyzed 4 million AI citations across ChatGPT, Google AI Mode, Google AI Overviews, and Gemini. Original editorial accounted for 81% of the news citations. Press releases distributed through syndication accounted for 0.04% of the full dataset (Search Engine Journal). Different study, same direction. I would not let a wire invoice and an earned article share a line called "coverage" and then buy a tool to watch them, as if the tool could promote the wire into the 81%.

Machine Relations research on earned versus owned pages ties the same pattern to a lift: distributed earned media can generate up to 325% more AI citations than brand-owned content alone, set next to Muck Rack's non-paid citation data (earned versus owned). The site still matters. It is the record the third party can point at. It is not the whole proof. I would build both, in that order: the source-of-truth page, the corroboration on a domain we do not own, the same claim structured so a machine can lift it without guessing, and only then a measurement of whether those sources appear when the buyer asks.

What software can see, and what it cannot decide

A tool can show a mention rate, a citation share, a sentiment tag, and a competitor list. It cannot decide which claims deserve to exist. It cannot earn the quote. It cannot turn a positioning line into a sentence with a metric, a sample, and a date. It cannot make an answer understand why the company belongs on a shortlist. I watched the SEO version of this mistake a decade ago: the measurement layer arrived before the substance, and the dashboard confirmed an absence everyone could already explain.

Layer	What it can answer	What I would do myself
Measurement	Are we named or cited?	Keep the baseline once there is something to baseline
Entity clarity	Do the same facts show up everywhere?	Write them once and repeat them
Citation architecture	Can a machine lift the claim?	Publish the specific, sourced sentence
Earned authority	Does a third party corroborate it?	Earn the page on a domain engines already cite
The whole system	Do those pieces become a recommendation?	Connect them, instead of buying them as five logins

That last row is Machine Relations. GEO and AEO are tactics inside it for answer surfaces. Formatting helps, and it is not the foundation. The glossary is where I keep a vendor from renaming "earned authority" as "content score" so the missing third-party page disappears from the conversation.

Structure still has a measured effect, and I would not let it rescue a claim with no source. The Princeton and Georgia Tech GEO paper found that methods aimed at generative engines can improve source visibility, including citing sources, adding quotations, and including statistics (arXiv). A later structural study reported a 17.3% citation-rate improvement across six generative engines when structure was tuned for extraction (arXiv). I would take the 17.3% as a second-order lift. It assumes there is a credible claim to structure. "We help companies transform customer engagement with AI" gives a model nothing it can reuse. A sentence with an entity, a metric, a sample, and a period gives a journalist something to quote and a model something to attribute. That is the compounding I would buy with time, before I bought it with software.

Google's core-update guidance transfers as discipline rather than as a tactic: wait for a real pattern, compare the right periods, and make the page more useful instead of painting it (Google). I would give AI visibility the same restraint. A tool purchased on a volatile week will report the volatility. It will not create the missing source.

How I would know the source layer is real enough to measure

I would not declare the layer ready because a page exists. I would declare it ready when each of the 20 questions has a claim with a home, and the home is marked. Our site, a third-party URL, an analyst, a customer proof, or still nowhere. Muck Rack's May 2026 cut, 25 million citations, earned media at 84%, paid and advertorial at 0.3%, journalism at 27% across more than 20,000 outlets, is why a nowhere that I fill only with another owned post is still a nowhere for citation purposes. The site is the record. The cited URL, in that corpus, is usually someone else's page. BuzzStream and Citation Labs, 4 million citations, original editorial at 81% of news citations and syndicated releases at 0.04% of the whole set, is why I would not let a wire count as filling the cell. The earned-versus-owned research, up to 325% more AI citations from distributed earned media than from brand-owned content alone, is the direction I would expect, not a number I would promise a board.

The tool's job, once those cells have started to fill, is narrow. Mention, citation, whether the description matches the claim, which competitor's URL was used instead. Sentiment and a leaderboard are optional. A content score that rises when we add headings is the 17.3% structural study being mis-sold. That study found a 17.3% citation-rate improvement across six engines when structure was tuned for extraction. The Princeton and Georgia Tech paper found gains from quotations, statistics, and cited sources, and methods up to a large visibility change on its benchmark. Both assume a claim worth structuring. "We transform customer engagement with AI" has no metric, no sample, and no period. I would not pay a tool to watch that sentence fail. I would rewrite it until it has an entity, a number, a scope, and a date, and then I would watch.

Google's core-update note, compare the right periods and skip cosmetic edits, is the rule for the second run of the same 20 questions. A new prompt is not a result. A branded prompt is a separate test of whether the engine can repeat a name it was handed, and I would keep it out of the 20. B2B research happening inside engines before the prospect reaches a vendor site is why the 20 are category questions. Absence there is a shortlist. The dashboard printout of that absence is not a new fact. Stanford-sized adoption pressure is not in this decision. The fact in front of me is a corpus that cites earned pages, and a company that has or has not earned one.

I would buy software when exporting the 20 by hand is the bottleneck, not when I am embarrassed by a low score. Reliability, a saved answer, a saved URL, and a visible failed run matter more than a single index. If the demo cannot show me the sentence behind a point, I would keep the spreadsheet. The spreadsheet already contains the only fields the decision uses.

What I would do

I would start with 20 buyer questions that do not contain our name. Category questions, the kind a person asks before they know who to call. For each question I would write the claim we need the answer to be able to support, and I would mark where that claim lives today: our site, earned media, an analyst, a customer proof, or nowhere. The nowhere column is the work. I would publish the source-of-truth page, earn the interview, place the quote, build the comparison we can defend, and strip the vague sentence until a machine can extract what remains.

Then I would measure. The same 20 questions, across the engines, with a record of whether we are named, which sources are cited, and whether the description is accurate. If the answer cites a competitor's earned page and ignores our post, the next move is corroboration, not another schema plugin. B2B buyer research now happens inside these engines before many prospects ever reach a vendor site (Machine Relations research). Absence from that answer is a shortlist we were never on. A low score in a tool is just the printout.

Scope: a startup that can say something specific and true. If the proof does not exist yet, the honest build is the proof, not the monitor. I would also exclude branded navigational queries from the 20. Those measure whether the engine can repeat a name it was given. They are a useful separate check. They are not the visibility problem the tool is usually hired to solve.

What would change my mind: a category where Muck Rack's mix inverted, and owned pages or syndicated releases carried the citations the way earned editorial does in the May 2026 study, at 84% earned and 0.3% paid. If our buyer questions were being answered from our own site alone, I would buy measurement immediately and I would stop treating third-party corroboration as the first build. The 4 million citation study, with syndication at 0.04% and original editorial at 81% of news citations, is the evidence I have. While it holds, the software comes after the source. A dashboard on an empty layer will tell me the layer is empty, which I can see by reading five answers for free.

FAQ

What is the visibility I would try to earn?

The probability that an answer names the company, cites a real source, and describes it accurately on a category question. Earned authority, a consistent entity, an extractable claim, and a measurement across engines are the pieces. More posts, by themselves, do not create it. Muck Rack's 25 million citations put earned media at 84%.

When would I pay for the software?

After the 20 questions and the proof map exist, and after the nowhere column has at least started to fill. The tool can track whether the engines cite us. It cannot place the quote or invent the metric. Buying it first repeats the SEO mistake: a precise view of a missing substance.

What would I build in the first pass?

A source-of-truth page, one third-party corroboration on a domain these engines already retrieve, and the same claim in a form that includes a number, a scope, and a date. The structural study's 17.3% across six engines is the lift I would expect from making that claim extractable. The GEO paper's gains from quotations, statistics, and cited sources are the same idea on an earlier benchmark. I would not chase either number with a page that has nothing to quote.

Where do GEO and AEO fit?

They are tactics for getting a page cited and for becoming the direct answer. They sit inside the broader system: earned authority, entity clarity, the architecture of the claim, distribution, and measurement. I would not hire a GEO engagement to compensate for a nowhere column. I would hire it to structure a claim that already has a source.

How would I test this without a vendor?

I would run the 20 questions on ChatGPT, Perplexity, Claude, Gemini, and Google AI Mode. I would record the name, the citation, the sources, and whether the description matches the proof map. Then I would fix the missing source and run the same 20 again. Google's advice on core updates, to compare the right periods and skip cosmetic edits, is the rule for that second run. A new prompt is not a new result.