One prompt, and a very large wrapper around it. Measured on 12 October 2026 across 100 random pages drawn from the 16,762 prompt URLs in the Faymas sitemap: the median page ships 563,156 bytes of HTML and carries a prompt of 195 words (1,384 characters). That prompt is 0.25% of the bytes and about a fifth of the words a reader can actually see. Every one of the 100 pages returned HTTP 200, every one carried exactly 12 related-prompt cards, and every one of the 1,300 view counters we read said zero.
This guide analyses published prompt text. No generation tests were run.
How we measured it
The method is four steps and anyone can repeat it. First, fetch the sitemap and keep every URL under /prompt/. Second, take a seeded random sample so the draw is reproducible. Third, download each page with a plain HTTP client and no JavaScript. Fourth, read the numbers out of the response.
Two details matter, because both changed our results. The prompt text was located in the rendered HTML by anchoring the page's meta description and reading forward to the "Show more" control; that worked on 84 of the 100 pages, so prompt-length figures cover 84. Fields such as the model label, the category and the author were read from the visible text rather than from the data payload, because a payload field cannot be pinned to the page's own prompt with certainty — a prompt page also carries a dozen related items, each with its own fields.
563 KB in, 195 words out
Here is the whole page, measured. Bytes are the size of the HTML document as served; "visible words" strips scripts, styles and tags and counts what a reader can see.
| Layer | Median (100 pages) | Range |
|---|---|---|
| Whole HTML document | 563,156 bytes | 527,616 – 615,648 |
| Words a reader can see | 911 words | 674 – 1,960 |
| The page's own prompt | 195 words / 1,384 characters | 22 – 1,302 words (84 pages) |
| Prompt as share of bytes | 0.25% | — |
| Prompt as share of visible words | 21.7% | 3.3% – 66.4% |
| Related-prompt cards | 12 | 12 – 12 |
Every page in the sample landed between 515 KB and 601 KB. That is a strikingly narrow band for 100 different pages with different text and different images, and it is what you would expect from one template filling variable slots.
The prompt itself is not uniform. Five of the 84 were under 60 words; the shortest was 22 words and asked for "a futuristic energy reactor". Forty of the 84 ran past 200 words, and the longest ran to 1,302. The deciles run 71, 94, 129, 155, 196, 238, 342, 519, 679 words — a long right tail rather than a tidy average.
The view counter has never moved
This was the least expected result. Each page carries 13 view-count fields in its data payload — one for the prompt itself and the rest for the related items. Across 100 pages that is 1,300 fields, and not one of them held a number above zero.
Likes and copies do move. Taking the counter fields nearest each page's prompt string, the median prompt had 2 likes (24 of the 100 had none) and a median of 46.5 copies, with a maximum of 1,101. One copy count — 143 — appeared on 14 separate pages, which is a sign that the same items recur across unrelated pages.
Two limits on that paragraph. We measured the server response; a counter incremented in the browser after load would not appear. And the like and copy figures sit next to the prompt string we could identify, so treat them as best-effort rather than exact. The zero view counts need no such caveat: zero appeared everywhere we looked.
Three model labels, one category, eleven regular authors
The metadata on these pages is narrow. Of 100 pages, 78 were labelled "GPT Image", 19 "Nanobanana" and 3 "Midjourney". Categories were more lopsided still: 85 Photography, 7 Social, 4 Illustration, 3 Wallpapers, 1 Presentation.
| Field (read from visible text) | Values across 100 pages |
|---|---|
| Ai Model | GPT Image 78 · Nanobanana 19 · Midjourney 3 |
| Prompt Category | Photography 85 · Social 7 · Illustration 4 · Wallpapers 3 · Presentation 1 |
| Distinct authors | 54 for 100 pages |
| Authors with 3 or more pages | 11 authors, covering 48 pages |
| Most prolific author in the sample | info_x_nik, 10 pages |
So the supply is not 100 independent creators. Eleven accounts produced nearly half the sample, and a single account produced a tenth of it. That matters if you read a catalogue size as a measure of a community; it is closer to a measure of a handful of productive accounts.
Half the prompts ask for a photo you do not have
Of the 84 prompts we could read in full, 50 instruct the reader to work from an uploaded or reference image — "use my uploaded face as the exact identity reference", "use photo 1 as face and hair reference", "keep the person exactly as shown in the reference image". Those prompts are not standalone. They are the second half of a workflow whose first half is a photograph the page does not hand you in a usable form.
Fifteen of the 84 contained pasted JSON — blocks with keys like subject, setting, composition, negative_prompt and retouching_analysis. That is an image-analysis dump, not a sentence a person typed for another person to read. Both patterns point the same way: a large share of what sits in the prompt field was produced by a tool working on an image, then published.
What the page is for
Put the numbers together and the unit page reads as a delivery wrapper rather than a document. It spends 563 KB to deliver 1,384 characters, repeats the same 12-card grid on every page, offers three model labels where the studio advertises more, and shows a view counter that has never recorded a view. The part that varies — the prompt — is a fifth of what you can read and a quarter of one percent of what you download.
None of this makes the catalogue fake. All 100 pages answered 200, all 100 served their prompt text without JavaScript, and no two of the 84 prompts we read were identical. The pages exist and they work. But "16,762 prompts" is a count of URLs, and a URL here is a wrapper with one prompt inside it — not a measure of distinct demand, not a community, and not a signal that anyone looked at the page.
The last point is the one worth watching. If the view counter is genuinely dead, then the only live popularity signals on a Faymas prompt page are likes and copies, and both of those are counts of clicks by people who were already on the site.
What we could not check
Sixteen of the 100 pages are missing from the prompt-length figures because the meta-description anchor did not appear verbatim in the rendered text — HTML entity escaping is the likely cause. Those pages did render a prompt; we simply could not extract it mechanically, so we left them out rather than guess. We did not use an archive service, we made no traffic estimate, and we make no claim about how any search engine treats these pages. Everything above is a measurement of the response as served on 12 October 2026 and nothing more.
FAQ
What is actually on a Faymas prompt page?
One prompt inside a large template. Across 100 pages sampled on 12 October 2026 the median prompt was 195 words (1,384 characters), the median HTML document was 563,156 bytes, and every page carried exactly 12 related-prompt cards.
How big is a Faymas prompt page?
Median 563,156 bytes, with every page in our 100-page sample falling between 527,616 and 615,648 bytes. The prompt inside it is a median 0.25% of those bytes.
Does the prompt text appear without JavaScript?
Yes. We read every prompt out of the raw HTTP response with no browser involved, and all 84 prompts we located were present in the rendered text of the served HTML.
How many prompts does faymas.in actually have?
Its sitemap listed 16,762 URLs under /prompt/ on 12 October 2026. In our sample of 84 readable prompts, no two were identical, but we sampled 100 pages and cannot speak to the other 16,662.
Why does the view count show zero?
Every one of the 1,300 view-count fields we read across 100 pages held zero. We measured the server response only, so we cannot tell whether views are counted in the browser after the page loads.
Can I check this myself?
Yes, in three commands: curl -s https://faymas.in/sitemap.xml and count the loc entries under /prompt/; curl -s any prompt URL and grep for viewsCount to see the counters; and strip tags from the response to count the words a reader can see. Each takes under a minute.