We Re-Tested Our Own Case Study Three Months Later
AI-visibility case studies usually end at the trophy. Ours re-measures. Three months after FOIHK (Family Office Institute Hong Kong) reached #1 in two AI engines, we re-ran the same questions in the same wording across 7 engines and archived 49 dated screenshots. What held: brand identity (7 of 7 engines correct, all with citations — including Doubao, which had the name wrong in July) and the top-5 recommendation podium, which strengthened (Google AI moved FOIHK from #3 to #1, with Copilot #2, Gemini #3 and Perplexity #5 all new). What moved: the two #1 positions on the "I am looking for…" phrasing, now led by government- and university-backed bodies; and full-name recognition, which became engine-specific — ChatGPT confuses FOIHK with a similarly named body while three other engines improved. Diagnosis: engine-side drift on an unchanged site. The lesson: positions in AI answers are measurements with timestamps, not property — which is exactly why Visora ships case studies with dated re-tests.
In this guide
- The same ruler, measured twice: same questions, same wording, 7 engines, 49 dated screenshots
- What held — identity 7/7 with citations; the top-5 podium strengthened to four engines (Google AI #3 to #1)
- What moved — two #1s on "I am looking for…" vanished to government/university bodies; full-name confusion shifted engines
- Engine-side drift on an unchanged site: you cannot patch an engine, only keep measuring it
- FAQ: why rankings move without site changes, why two phrasings get two podiums, re-measure cadence
In July we published the FOIHK case study: from 0 appearances in 15 AI answers to #1 in two engines. Most case studies end there. Ours did not. On the morning of 21 October 2026 we re-asked the same questions — same wording, same order — across 7 engines, and kept 49 dated screenshots.
The Same Ruler, Measured Twice
A re-test only counts if it is comparable, otherwise it is just another sample. The three baseline questions were re-asked verbatim; engines expanded from five to seven (adding Perplexity, plus a Google Search control), with four more query families — English full name, Chinese full name, associations phrasing, networking phrasing. Every answer was read line by line: mentioned or not, ranked where, and who is ahead now. Every screenshot is archived in the case portfolio next to the July before-and-after pairs.
The result fits in one sentence: on the same morning, FOIHK was promoted to #1 by one engine and dropped from #1 by two others.
What Held
Identity: 7 of 7, every answer cited
Asked "What is FOIHK in Hong Kong", all seven engines — ChatGPT, Copilot, Gemini, Google AI, Doubao, Perplexity, Google Search — described FOIHK correctly, each with citations. Doubao, which spent July calling it "FOB HK", now cites foihk.org directly.
The top-5 podium got stronger, not weaker
For "Can you recommend me the top 5 family office institutes in Hong Kong": Google AI moved FOIHK from #3 to #1, Copilot listed it #2 (not listed at all in July), Gemini #3 (also new), Perplexity #5. Four engines now carry FOIHK on the same question that carried one in July.
Google organic: #1 on everything we tested
Brand queries, category queries, English and Chinese full names — foihk.org sat at #1 organic across the board. What the GEO work built, conventional SEO is still holding up.
What Moved
The two #1s on "I am looking for…" quietly vanished
In July, Copilot and ChatGPT both ranked FOIHK #1 for "I am looking for good family office institute in Hong Kong". In October, neither listed it — both now lead with the government-backed HKAWL and HKU SPACE. Same client, same site, zero changes on our side in the interim.
Full-name confusion moved to new engines
ChatGPT now answers the English full-name question with the lookalike FOI (a different organisation, foi.asia); Google AI merges FOIHK and FOI into one entity, and on the Chinese full name it leads with rival FOAHK. Meanwhile Copilot, Doubao and Gemini all improved on the same queries — recognition has become engine-specific, not uniform.
Diagnosis: engine-side drift, not a site problem
The FOIHK site was untouched for the whole period — canonical tags healthy, entity statements intact, hosted and reachable. When engines move in different directions on the same unchanged site, the movement is on the engine side: model updates and answer-policy shifts. You cannot patch an engine; you can only keep measuring it.
One Word Swaps the Podium: “Association” vs “Institute”
Same morning, same client — swap one category word in the query and the engines return two entirely different podiums. Ask for a "good family office association" and the answers fill with trade bodies: FOAHK ranks #1 on ChatGPT, Copilot, Doubao and Perplexity, while FOIHK surfaces on only two engines (Gemini at #3, Google AI as a secondary mention). Ask for the "top 5 family office institute" and the same engines pivot to education and research bodies — FOIHK on four: Google AI #1, Copilot #2, Gemini #3, Perplexity #5.
"Association" is FOAHK’s home turf
Engines map the word to trade associations — member bodies, advocacy, policy dialogue. FOIHK is not a member body, so competing for that word means fighting the category definition itself, not a competitor. Two-engine visibility is the honest ceiling there.
"Institute" is FOIHK’s winnable turf
The word matches the entity type — education, research, training — so the evidence engines retrieve (foihk.org guides, research, certification) supports the recommendation. When the category word matches what you actually are, the same engines that ignore you elsewhere put you first.
Chinese has the same mirrored pair: 協會 (association) and 學會 (institute/society). Asked the Chinese full name — 香港家族辦公室學會 — Google AI led its answer with the 協會 (FOAHK). One character class, a different entity. This is why Visora starts content work with a category-vocabulary audit: measure which word belongs to whom, then target the word you can own.
What This Milestone Actually Records
Not "we won once", but "we keep measuring". A position in an AI answer is rented, not owned — and the rent is paid in freshness, third-party citations, and measurement cadence. When two phrasings of one intent produce two different podiums on the same day, the one-screenshot case study is obsolete. A trustworthy case study is one that dares to print its next re-test date.
FOIHK’s next quarterly re-test is already on the calendar. The same ruler will measure again — and whatever it says, we will publish it.
FAQ
Our site did not change — why did our AI rankings move?
Because the engines did. AI answers come from models that get updated and retrieval policies that get tuned — continuously. In this re-check the same unchanged site gained a #1 on Google AI and lost two #1s on Copilot and ChatGPT on the same morning. Treat AI positions as measurements with a timestamp, not as properties you own.
Why did two phrasings of the same question get different results?
"Top 5 family office institute" and "I am looking for a good family office institute" express the same intent, but engines retrieve different evidence for each: list-style phrasings pull comparison and directory pages, conversational phrasings pull government and academic bodies. On the same day, FOIHK held a four-engine podium on the first and missed the second entirely. Target the phrasing your buyers actually type — then measure both.
How often should a business re-measure AI visibility?
Quarterly for the full ruler — same questions, same wording, dated screenshots — and monthly for a lightweight scan. Retrieval-based engines (Perplexity, Google AI Overviews) move in days; trained-model memory moves in months. This case moved measurably within one quarter, which is exactly the cadence that catches it.
Where Does Your Brand Stand in AI Answers Today?
Free 60-second scan across 9 engines — raw answers, side by side.
Last updated: