Plugin Visibility Index · Run 05
ChatGPT reads twenty-three sources and shows you three
Across ten WordPress plugin categories, ChatGPT retrieved 217 source URLs and displayed 36 of them — 17%. The median answer was built from 23 sources and credited 3. If you are judging your AI visibility by the citations on screen, you are looking at a sixth of the retrieval.
What was measured?
The same ten categories and the same prompt shape used in Run 03 — “best WordPress [category] plugins in 2026” — put to ChatGPT on 10 September 2026. For each answer I recorded two things: the domains ChatGPT displayed as citations, and the domains it actually retrieved, read from the conversation’s own stored source metadata rather than from the rendered page.
This makes ChatGPT the only engine here measured at retrieval rather than at display. Run 03 compared what Perplexity and Google AI Mode show a user, because that is all those two expose. So ChatGPT’s larger domain counts are partly a wider measurement window, not purely a wider index. Every comparison below is labelled with which of the two it uses.
How much does each engine show you?
Like for like — sources visible to a user, no backend data:
| Engine | Median sources shown |
|---|---|
| Perplexity | 10 |
| Google AI Mode | 8 |
| ChatGPT | 3.5 |
ChatGPT credits roughly a third as many sources as Perplexity. And it is the engine reading the most: a median of 23 retrieved per answer against the 10 Perplexity displays.
Where does the other 83% go?
| Category | Shown | Retrieved | Share shown |
|---|---|---|---|
| Booking | 6 | 6 | 100% |
| Affiliate | 5 | 25 | 20% |
| Membership | 4 | 20 | 20% |
| LMS | 4 | 24 | 17% |
| Caching | 4 | 27 | 15% |
| Translation | 3 | 20 | 15% |
| Tables | 3 | 22 | 14% |
| Backup | 3 | 26 | 12% |
| Product feeds | 2 | 22 | 9% |
| Popups | 2 | 25 | 8% |
Booking is the one exception and it is instructive: six retrieved, six shown. That answer barely searched. Everywhere else the engine read twenty-odd pages and named a handful.
The practical consequence for a plugin vendor is uncomfortable. You can be read and not credited. Your page can be in the set the answer was written from, shape what the model says about your category, and never appear as a link — so a monitoring tool counting visible citations records you as absent while your writing is doing work.
This is a third state, alongside the unreachable / unknown / unretrieved split in the diagnostic. Retrieved-but-not-displayed is invisible from outside, invisible to every tracker, and it is the majority case on this engine.
Does a third engine bring the picture together?
No. It pulls it further apart.
| Pair | Median source overlap |
|---|---|
| Perplexity & Google AI Mode | 15.5% |
| Google AI Mode & ChatGPT | 12% |
| Perplexity & ChatGPT | 8% |
Across all three, 250 distinct domains. Seventeen of them appear somewhere in all three corpora — 7%. Three engines answering the same ten questions are reading three largely separate slices of the web.
So “am I visible in AI search” has three answers, not one, and they do not average. A tool that blends engines into a single score is hiding the only number that tells you where to work.
A correction to Run 03
Run 03 found Perplexity citing locale mirrors of wordpress.org — ps., yor., ast., fur. — alongside the canonical, while Google AI Mode cited only wordpress.org. I wrote that this “looks like a Perplexity retrieval behaviour rather than something inherent to AI search.”
That was wrong, and ChatGPT is the counter-example.
| Engine | Categories with a locale mirror | Distinct mirror hosts |
|---|---|---|
| ChatGPT | 4 of 10 | 6 |
| Perplexity | 2 of 10 | 4 |
| Google AI Mode | 0 of 10 | 0 |
ChatGPT split across en-gb., en-ca., de., el., es. and it. mirrors — twice as many categories as Perplexity, and in every case the canonical wordpress.org was retrieved as well. The product-feed answer alone pulled five different locale mirrors of the same directory plus the canonical plus wordpress.com.
The corrected reading: this is a behaviour of retrieval-based engines generally, present on two of the three tested and absent only on Google. If you publish hreflang alternates, two of three engines will sometimes treat a translation of your page as a separate source rather than consolidating on your canonical. That is a wider problem than Run 03 described, and I would rather say so on a new page than quietly edit the old one.
What travels across all three engines?
Seventeen domains appear in all three corpora. Two of them are the FS Code network: fs-poster.com and booknetic.com.
Run 02 and Run 03 identified this booking-software publisher as the one thing in the data that crossed from Perplexity to Google. It crosses to ChatGPT too — booknetic.com in the LMS answer, and a second booking company, bookingpressplugin.com, retrieved in both the backup and affiliate answers. Booking-software vendors, cited in categories about backups, learning management and affiliate tracking.
Three engines, three largely disjoint indexes, and the same publishing pattern surfacing in all of them. It is not a backlink strategy and it is not schema. It is volume of long, structured, comparison-shaped writing about categories adjacent to what they sell. That is now the most replicated finding in this index.
What can this run not tell you?
One prompt per category, one run each, one day. Ten categories is a small base for a median and the per-category rows are single observations. The Perplexity and Google figures are carried over from 8 September, so the cross-engine comparison spans two days.
The biggest limitation is the one stated at the top. ChatGPT is measured at retrieval and the other two at display, so the 8% and 12% overlap figures are comparing a wide set against narrow ones. If ChatGPT’s displayed sources were compared instead, the overlaps would be computed from three or four domains per category — too few to mean anything. Neither framing is clean, and I have used the one with more data in it while saying so.
The 17% display share is solid — both halves come from the same engine on the same answers — but it describes ChatGPT on this date and should not be generalised to other engines or assumed stable.
The locale-mirror correction is also drawn from ChatGPT’s retrieval set. I can say those mirrors were retrieved. I cannot say they would have been displayed.
Which of the three can see you?
Send me your wordpress.org slug. I will run your category on all three engines and send back the full source lists — including which engines read you without crediting you. Free, no call, usually within a day.