Plugin Visibility Index · Attribution 01

Who actually supplies the sources in AI answers about WordPress plugins?

Across 60 recorded answers in ten plugin categories, a single answer attributes a median of four domains, one of which takes 43% of the citations. A booking-plugin vendor’s blog is the most-cited domain overall — in four of the five categories where it appears, it sells nothing.

What was measured?

Fifty prompts — ten WordPress plugin categories × five prompt shapes — run against Perplexity on 5 and 6 September 2026, plus ten repeat runs to measure variance. For each answer I recorded which products were named and in what order, and I counted every inline attribution mark in the answer body and which domain it pointed at.

That last number is the one this page is about. Call it attribution share: the proportion of an answer’s inline citations that point at a single domain. It is not the same as the source count the engine reports. Perplexity said “15 sources” on almost every answer; the visible body leaned on a median of four.

How concentrated is attribution inside one answer?

More than the “15 sources” label suggests. In the 28 answers where attribution was measurable, the top domain took a median 43% of all citations in the body. In 18 of 28 it took at least 40%; in five, at least half. One membership answer took 81% of its attribution from a single blog.

The practical reading for a plugin vendor: you are not competing for a slot in a list of fifteen. You are competing to be one of about four pages the answer actually leans on, and ideally the one it leans on hardest.

Which domains supply the most citations?

Ninety-five distinct domains produced 657 attribution marks. The distribution has a long tail and a very short head.

Two things stand out. The first is that wordpress.org — the canonical directory, with the install counts, the changelogs and the support forums — supplied 6.4% of attribution, and that a booking-plugin blog supplied more. The second is where those wordpress.org citations actually landed.

Why did the engine cite the isiZulu WordPress directory?

Of the 42 marks pointing at wordpress.org, only 17 went to the canonical domain. The other 25 went to locale mirrors — and not the ones you would guess:

Every one of these serves the same English plugin readme. I checked zul.wordpress.org/plugins/optinmonster/: the page is in English, it is indexable, and it declares itself canonical. That is not a bug — it is correct hreflang. The canonical English page lists the locale mirrors as rel="alternate" alternates, and each alternate self-canonicalises, exactly as the specification says it should.

The finding is what the retriever does with that. Answering English-language queries, it treated the isiZulu and Tamazight mirrors as separate sources and cited them 25 times — more often than the canonical page it was supposedly an alternate of. Hreflang tells a search engine which version to show a user. It did not consolidate retrieval here.

Why is a booking plugin the most-cited source about popups?

Booknetic sells appointment-booking software. Its blog was attributed 67 times — one citation in ten across the whole corpus — spread over five categories:

In the popup answer, 25 of that answer’s 67 citations pointed at one page: booknetic.com/blog/best-wordpress-popup-plugins. Booknetic does not appear in the answer as a product. It appears as the authority on which popup plugins are good.

This is not a trick. It is a vendor that publishes long, structured, comparison-shaped listicles about categories adjacent to its own, and those pages are what the retriever finds. Their sibling domain fs-code was the top-attributed source in the table-plugin answer by the same mechanism.

How much attribution goes to vendors rather than independent reviewers?

Of the 657 marks, 232 — 35% — pointed at a domain owned by a WordPress product vendor. Another 8% went to wordpress.org, WooCommerce docs, YouTube or LinkedIn. The remaining 56% went to independent blogs, review sites and hosting companies.

Vendor citation does not reliably buy the recommendation. In the OptinMonster vs Popup Maker answer, OptinMonster’s own domain was the most-cited source and the answer still opened by recommending the free rival. In the AffiliateWP vs SliceWP answer, 11 of 20 citations came from Awesome Motive properties — the network that owns AffiliateWP — and the answer led with SliceWP. But in the product-feed category, AdTribes’ own domain was the top source and AdTribes took first position.

Three cases, split two to one. That is an observation, not a rate, and it is written here as one.

Does the shape of the question change the answer?

In four of ten categories, the product that led the “what is the easiest way to do X without writing code” prompt did not lead any of that category’s discovery prompts.

Problem-shaped prompts also pull in products that appear nowhere in the category’s discovery answers. “Easiest way to publish sortable tables” surfaced TableKit and TableCrafter, neither of which appears in any “best table plugins” answer. “Easiest way to capture email leads” abandoned popup plugins almost entirely and answered with page builders and form plugins instead.

And AffiliateWP, the largest name in its category, was absent from the affiliate problem prompt altogether.

Do the engines cite staging sites?

Twice in this corpus, yes. dev.smackcoders.com supplied 13 citations — 48% of the attribution in the WPML vs Polylang answer. staging.wpmayor.com supplied six in the affiliate listicle.

Two instances is not a rate. It is enough to be worth thirty seconds of your time: if your staging or dev subdomain is reachable and indexable, it can be retrieved, and the citation goes to a URL you did not mean to publish. Check it with site:dev.yourdomain.com and a robots.txt you have actually read.

Are the engines still recommending products that no longer exist?

Restrict Content Pro was named in all three membership answers, twice in third position, once described as the “best lightweight / developer-friendly” option. In one answer the engine hedged the merge into a single label — “Restrict Content Pro / Kadence Memberships” — which is closer to right than the other two but still presents a sunset product as a live recommendation.

This matches the Run 01 finding, where BackupBuddy appeared the same way. Product names outlive products in a retrieval index, because the pages that named them are still online and still being read.

What can this data not tell you?

Seven of the ten categories showed a different leading product between differently-worded discovery prompts. That number mixes two effects I cannot separate here — the wording of the prompt, and run-to-run variance in the engine itself. Run 01 measured the variance directly by repeating identical prompts and found a mean overlap of 74%, with one category as low as 40%. So some of that seven is real prompt sensitivity and some is noise, and this corpus cannot say how much of each.

I am saying this because on the last index entry I nearly published a prompt-shape finding that turned out to be variance. The fix is not to stop reporting; it is to say which numbers are load-bearing. The concentration figures and the domain league table are averaged over 28 answers and are load-bearing. The single-answer observations — the 81% membership answer, the two staging domains, the three vendor-citation cases — are anecdotes, labelled as anecdotes.

The attribution count is also a count of visible marks in the answer body, not of pages retrieved. One table-plugin answer reported 15 sources and attributed none of them inline. If your page is retrieved but never marked, it influenced the answer and sent you nothing.

---

Want your category measured?

Send me your wordpress.org slug. I will run the five prompt shapes for your category and send back who leads, who supplies the citations, and how concentrated the attribution is. Free, no call, usually within a day.