Plugin Visibility Index · Run 01 · 5 September 2026

Ask an AI engine the same question twice and a quarter of the answer changes

I asked Perplexity for the best WordPress plugin in ten categories, twice each, and recorded every product it named. Mean overlap between the two runs was 74%. In the worst category it was 40%. The largest backup plugin on wordpress.org — five million installations — was absent from all four backup answers.

What did the engine actually name?

One prompt per category — “best [category] in 2026” — the listicle shape that carries 21.88% of all AI citations. Sorted by how many products the answer named, fewest first.

Mean products named per answer: 5.5. Range: 2 to 9. The affiliate row shows a range because the same prompt returned two products on one run and five on another.

Why is the biggest plugin missing?

All-in-One WP Migration reports 5 million+ active installations on wordpress.org — the largest backup plugin in the directory by a wide margin. It was not named in any of the three backup answers I ran.

The affiliate category looked like a second example — AffiliateWP, the best-known product in the category, absent from a two-product answer. It was not. Re-running the same prompt named it. That correction is set out below, because it changes what the rest of this page is worth.

Meanwhile FlyingPress, Academy LMS and WP Table Builder — all far smaller — were named. The mechanism is not mysterious once you look at the sources: the engine is summarising listicles, and a plugin’s presence in those listicles is what decides the answer. Installations are invisible to it.

Are the engines recommending products that still exist?

Not always. Two answers recommended products under names that were retired:

A buyer following either recommendation goes looking for a product that cannot be bought under that name. If you have rebranded, merged or been acquired, there is a live question worth asking: is the engine still selling your old identity?

Does the same prompt give the same answer twice?

No — and this is the finding that changes how you should read every other number on this page, including mine.

I ran the identical affiliate prompt twice on the same day, same engine, same wording. Not a rephrasing. The same string.

AffiliateWP was absent from the first answer and present in the second. Nothing changed between them but time.

So I repeated it across all ten. Each identical prompt, run twice, same engine, same day. Overlap is the share of products appearing in both runs.

Mean overlap across the ten: 74%. Roughly a quarter of the products named in any given answer will not be there the next time you ask. The top recommendation changed in three of ten.

Three things follow, and the third is the one that costs money.

Volatility is a property of the category, not of AI in general. Translation and booking returned identical product sets both times. Backup dropped four of nine and swapped its top recommendation from UpdraftPlus to BlogVault. If you sell a backup plugin, your presence in that answer is closer to a coin toss than a ranking. If you sell a translation plugin, it is close to a fixed list — which is bad news if you are not on it, and good news if you are.

Being named is more stable than being named first. Translation had perfect overlap on which products appeared and still changed which one led — WPML first on one run, Polylang on the other. Booking did the same. A vendor tracking only “am I mentioned” would call both categories solved.

A single-run number is not a measurement. At 74% mean overlap, any share-of-voice figure taken from one pass carries roughly a quarter noise — and in the worst category, more than half. That is larger than most of the improvements a vendor would be trying to detect. If someone sells you an AI visibility score with no run count and no variance attached, the number cannot tell you whether your last three months of work did anything.

I nearly published a different conclusion. Between those two runs I had tested a reworded prompt, seen AffiliateWP appear, and concluded that phrasing was the cause — that the product was invisible to buyers who said “affiliate plugin” and visible to those who said “run an affiliate programme”. It was a clean story and it was wrong. Re-running the original wording produced AffiliateWP too. The variance was noise, not phrasing.

This is why the audit runs the set four times across four engines rather than once, and why the tables here will carry a run count from now on. It is also why the single-number “AI visibility score” that several tools sell should be treated with suspicion: a number with no variance attached is a number you cannot act on.

Where do the answers come from?

Almost none of the cited sources were vendor product pages. They were listicles — and a striking share were published by other plugin vendors: booknetic, wedevs, blogvault, teamupdraft, adtribes, optinmonster.

Two answers cited a vendor’s own domain in support of an answer that named that vendor first. optinmonster.com was a source for the popup answer led by OptinMonster. adtribes.io was a source for the product feed answer led by AdTribes. Whether that is causal or coincidental, one thing follows either way: the companies writing the category listicles are the companies winning the category answers.

That is why off-site placement is half of this work, and why optimising your own product page is the smaller half.

How was this measured?

How much should you trust run 01?

Less than you would trust run 06, and I would rather say that plainly than dress twelve answers up as a survey.

Four honest limits. This is one engine — Perplexity behaves differently from ChatGPT, which queries Bing, and from Google AI Mode. It is one run, and these systems are not deterministic; the same prompt tomorrow may name different products. It was run from a logged-in account, which may personalise results. And ten categories is a small sample chosen on judgement, not at random.

What survives those limits is the shape of the finding, not the decimals: a plugin can lead its category on installations and be absent from the answer. That happened twice in ten, and once it involved five million installations. No amount of sampling noise explains that away.

Run 02 adds ChatGPT, the remaining prompt shapes, and a second pass a fortnight later so variance can be measured rather than assumed. The numbers here will not be quietly edited — if run 02 contradicts run 01, both stay up.

---

Is your plugin in one of these categories?

Send me your wordpress.org slug and I will run the five prompts your buyers use, and send back what the engines actually said about you. Free, no call required, usually within a day. Or build your own prompt set with the prompt set builder.