Our own AI visibility score is 0

We sell AI visibility measurement, so the first brand we measured was our own. On 19 September 2026 we put eight buyer questions to Perplexity through its API and analysed the answers with the same code our clients’ reports run on. Bungad was named in 0 of 8 answers. Mention rate 0%. Citation share 0.0%. No average position, because we never appeared. Visibility score: 0. We ran it twice to be sure, and kept both records; every figure on this page traces to one of them.

Nobody else in this category publishes their own audited number, and we understand why: ours is the worst one possible. We are publishing it anyway, for three reasons. It proves the measurement is real rather than a marketing graphic. It shows you exactly what a client report contains, including the parts that sting. And it starts a public before-and-after that we will update every month on this page.

The numbers

The score is weighted: mention rate counts for half, citation share for three tenths, and position among named brands for the remaining fifth. All three inputs were zero, so there is no rounding to argue about.

The eight questions

These are fixed. We will ask exactly these again next month, so the comparison means something. Every one of them returned an answer that did not contain the word Bungad.

Question four is the one that hurts. “A monthly AEO subscription for small businesses” is a description of our entire product, and the engine still had six other names to reach for before it would have needed ours.

Who got named instead

Across the same eight answers, these brands were named. The count is the number of answers each appeared in, out of eight. Both stored runs produced this table identically, naming the same brands in the same answers: questions 2, 5 and 8 carried every mention, and the other five answers named nobody in our set at all.

Worth noticing: even the leaders in this category are named in a minority of answers. Nobody owns it yet. That is the whole opportunity — and the reason a score of 0 is a starting line rather than a verdict.

The domains Perplexity cited in run A

This is the most useful output of any audit, ours included: the sources the engine drew on while answering questions about our category. The figures below come from the stored record 2026-09-19T04-22-26-292Z.json — the run we call A — which contains 90 citations across 74 distinct domains. Not one of them was bungad.com.

Only domains cited more than once are listed. Everything else sits in a 65-way tie at a single citation each, and picking a few names out of that tie to pad a top-ten would be presenting noise as a ranking.

Read that list as a map of who the engine trusts here. Two things stand out. First, a competitor’s own domain — peec.ai, cited four times — is the second most-cited source of all, ahead of Ahrefs and business.com, and scrunch.com appears twice. A vendor site can be treated as a source in this category if it answers the question directly. Second, everything else above the tie is publishers and directories rather than vendors. Earning an accurate presence on the sources an engine already quotes moves a score faster than another post on your own blog, because you inherit a trust you have not had to build.

The citation list moves between runs. The score did not.

We ran the same eight questions twice, about a minute apart, and stored both: run A above and run B (2026-09-19T04-23-37-261Z.json). Every headline number was identical — 0 of 8 named, 0% mention rate, 0.0% citation share, no average position, score 0 — and both runs named exactly the same competitors in exactly the same answers. The citation counts did not hold still:

So treat an individual citation count as an estimate, not a measurement. What is stable is the shape: 90 citations both times, 71 of the distinct domains shared between the two runs, and eight of the nine repeat-cited domains appearing in both. That is the level at which the list is worth acting on — which publishers and directories the engine reaches for, not whether HubSpot scored four or five.

What we are doing about our own zero

The same four things we would put in a client’s plan, in the same order, against the same questions.

1. A direct answer page for each question we lose. Eight questions, eight pages, each titled with the question and answering it in the first two sentences. This page is one of them.
2. Our specifics in plain text on our own domain. $99, $299 and $799 per month, the engines we actually cover, what we do not track. Text on a page, not a pricing image and not a PDF — because peec.ai being the second most-cited source in run A shows a vendor page can be quoted when it states facts plainly.
3. An accurate presence on the domains above. Business directories and category roundups first, since those are what the engine reached for when it did not reach for a vendor.
4. Re-run the same eight questions monthly and publish the new number here, up or down. A score that only ever gets published when it improves is not a measurement.

One positioning correction sits alongside this run rather than inside it: in the Philippines, “AEO” means Authorized Economic Operator, a Bureau of Customs accreditation programme for importers and exporters. Competing for that acronym locally means competing with the Bureau of Customs, so we use “AI search visibility” here instead. To be precise about provenance: no Philippine question is among the eight above, so this run does not measure that effect. Adding one is on the list for the next audit, and we will publish whatever it returns.

How to check this yourself

The run is a script in our repository, not a spreadsheet we maintain by hand. It asks the eight questions above through the configured engine APIs, analyses each answer with the same functions that produce client reports, and writes a complete JSON record of the run — every answer, every citation, the metrics — so a number on this page traces back to a file that exists. From web/:

npm run self-audit

Each run lands in web/data/self-audit/, named for the moment it started. The two behind this page are 2026-09-19T04-22-26-292Z.json (run A, the citation table above) and 2026-09-19T04-23-37-261Z.json (run B, the comparison column). Both record complete: true with 8 of 8 answers analysed and none failed.

Three honest limits. The engine is non-deterministic, so a re-run reproduces the picture and not the bytes — as the two runs above show, the score held at 0 while individual citation counts moved by one or two. Citation counts are therefore an estimate; the score and the brand tallies are the measurement. And eight questions on one engine is a small sample — enough to be certain about a zero, not enough to read meaning into a one-point move later. That is why the question set is fixed, the cadence is monthly, and every run is kept.

Why publish this at all

Because the alternative is asking you to trust a number we will not show about ourselves. Every claim on this site is checkable: the questions are listed, the method is a command you can read, and the result is the worst one available. If our measurement flattered anyone, it would have flattered us first.

Next update: October 2026, same eight questions, published here whichever way it moves.