Dated SEO Field Note

AI Trend Prompt Hub SEO Case Study: What Worked and What Failed

This is a public record of a real prompt-site experiment, including the part where early search growth did not continue. It separates measured results from hypotheses and records why the strategy moved from publishing more URLs to reviewing fewer, stronger evidence assets.

Last verified: September 4, 2026

Verified State on September 4, 2026

These figures use finalized Google Search Console web data and a same-day technical audit. The two GSC windows are directly comparable 28-day periods; GA4 sessions and crawler requests are reported separately.

Published prompt pages

55

Inventory verified on 2026-09-04; ordinary page expansion is paused.

Sitemap health

61 / 61

Every unique sitemap URL returned HTTP 200 on 2026-09-04.

Current finalized GSC

0 clicks / 312 impressions

Web search, 2026-08-05 through 2026-09-01.

Preceding finalized GSC

166 clicks / 8,044 impressions

Web search, 2026-07-08 through 2026-08-04.

GA4 recorded 30 Organic Search sessions from 2026-08-28 through 2026-09-03. Google contributed 0; observed sources were Bing, DuckDuckGo, and Yahoo. These sessions are not added to GSC clicks because the systems measure different events.

Timeline

Growth, contraction, and the strategy change

May-June 2026

Build a focused, crawlable prompt library

The first phase established a single-domain prompt library, canonical URLs, static page data, internal discovery, and manual publishing gates. Early Search Console demand concentrated around outfit changes, same-face edits, personal photos, and Korean photobooth prompts rather than the full topic inventory.

July 8-August 4

Five pages produced most Google demand

The preceding finalized 28-day window recorded 166 clicks and 8,044 impressions. Dress change, clothes changer, same-face outfit change, pictures of yourself, and Korean photobooth pages produced most of that demand, so they became the historical winner set.

From July 24

Google visibility contracted across the site

Daily Search Console data showed the break beginning on July 24. The next finalized 28-day window fell to 0 clicks and 312 impressions. This is an observed visibility contraction, not proof of a manual penalty or a specific algorithmic cause.

August 17-September 4

A first-party benchmark did not guarantee indexing

One benchmark and four problem-led support pages added exact prompts, dates, source/result evidence, limitations, and internal links. As of September 4, all five remained discovered but not indexed. Submitting or discovering a URL does not guarantee indexing.

What The Checks Ruled Out

No obvious crawl or server failure

The full sitemap crawl returned 61 of 61 URLs with HTTP 200. The historical winner pages remained indexed, were crawled recently, and had matching Google-selected and declared canonicals. Robots, noindex directives, structured-data parsing, internal links, PM2, and the production release were healthy.

That evidence removes common technical explanations. It does not prove why Google's ranking systems reduced visibility. Template repetition, weak independent authority, and limited first-party evidence are priorities for review, not claimed causes.

  1. 1Separate finalized GSC clicks and impressions from GA4 sessions, crawler requests, and indexing status.
  2. 2Verify crawlability, canonical consistency, server health, and actual URL Inspection results before changing content.
  3. 3Retain historical winners while reviewing whether weak pages own a distinct query or evidence asset.
  4. 4Strengthen a page with reproducible source/result tests, exact methods, dates, limitations, and next fixes.
  5. 5Use review-needed for unknown or highly repetitive URLs; never convert missing data into zero performance.
  6. 6Observe a complete data window before another change; do not repeatedly resubmit unchanged sitemaps or URLs.

Pages Retained as the Evidence and Winner Set

Retention here means the page has historical demand or substantial first-party evidence. It does not mean that rankings recovered, every support page was indexed, or further testing is unnecessary.

AI Photo Editing Prompt Benchmark

The flagship first-party evidence asset: four dated, controlled prompt tests with source and result images.

AI Clothes Changer Prompt

Historical Google winner that also received 11 non-Google organic sessions from 2026-08-28 to 2026-09-03.

AI Dress Change Prompt

Historical winner for named garments and styling, distinct from broad clothes replacement.

Same Face Outfit Change Prompt

Historical winner focused on identity preservation during clothing-only edits.

Prompts for Pictures of Yourself

Historical winner for personal-photo generation and reference-image control.

Korean Photobooth Prompt

Historical winner for four-cut strips and photobooth layout intent.

Current Experiments

Truthful public reporting

Keep dated search windows, inventory, limitations, and corrections visible instead of presenting old numbers as a current baseline.

Human-led URL review

Review all 55 pages using demand, first-party evidence, similarity, template dependence, freshness, and maintenance cost. No report automatically merges or removes a URL.

Reproducible prompt evidence

Turn the benchmark and a small number of proven pages into repeatable tests with fixed inputs, model/version, date, scoring rules, failures, and comparable results.

Independent validation

Seek a small number of legitimate discussions or citations around real experiments. External publication is separate from this page and requires editorial approval.

What this page does not claim

  • No manual penalty or specific ranking-system cause has been verified.
  • No sitemap, internal-link, schema, or answer-block change guarantees indexing.
  • No GA4 session is presented as a Google Search Console click.
  • No local review recommendation changes a URL until a separate editorial and deployment decision.

Method and correction rules are documented in the editorial policy. The underlying prompt tests are available in the photo-editing benchmark.