Synthesis — bupa.com.hk

Where the sources corroborate each other

• AI crawler access is open — two sources prove it. Live checks show 0% blocking across GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended and Bytespider. Brand Radar independently confirms it: Bupa's own domain is cited in 29 AI responses. If production AI crawlers were hard-blocked, own-domain citation share would be near zero, not 21%. Together these eliminate bot access as a cause and redirect the diagnosis to retrieval preference, content, or indexation.
• Both demand channels fail on the same product area: VHIS. Ahrefs: 「自願醫保」(10,000 volume, the largest tracked head term) ranks 15th with zero traffic. Brand Radar: VHIS comparison questions are the top AI demand cluster (「香港自愿医保哪家好？」2,100 impressions; 「自願醫保幾錢？」800), yet Bupa's own VHIS comparison page earns only 4 responses. Organic and AI underperformance converge on one asset family.
• The traffic decline has a two-sided AI mechanism. Ahrefs shows click suppression at rank 1 (「危疾」33% CTR, 「危疾保」18%) and paid spend up 645% backfilling organic losses. Brand Radar shows where the remaining clicks go: third-party pages (79% of mentions sourced externally; moneyhk101's single VHIS compare page matches Bupa's entire domain at 29 responses). The paid-cost line in Ahrefs is the financial symptom of the citation gap in Brand Radar.
• Migration residue is a shared hypothesis. Ahrefs: 95% backlink attrition and 76% referring-domain loss, a pattern that co-occurs with migrations or restructures. Live checks: /contact/ and /services/ return 404 for every agent — consistent with URLs moved without redirects. Neither source proves a migration; together they justify pulling Wayback/server logs in week 1.

Where sources contradict or complicate each other

• "Heavily cited in AI search" (Ahrefs) vs "own site barely cited" (Brand Radar). Both are true once disaggregated: Ahrefs counts brand presence in overviews (215 AIO citations, #1 for 危疾/危疾保/懷孕保險/自願醫保門診); Brand Radar counts sources. The resolved pattern: Bupa wins citations on direct brand queries (客服, about pages) and some informational sub-topics, but loses on comparison/evaluative queries. The models treat Bupa as a known entity but not a preferred source.
• Brand Radar internal flag: top-10 own pages sum to 33 responses vs 29 reported total — minor double-counting; does not change the <25% own-citation conclusion.
• Platform volume anomaly: Copilot triggers 54 questions but only 660 modelled impressions vs Google AIO's 8,750 from 24 questions. Modelled volume may undercount Bing-surface AI; do not allocate effort on this model without analytics referral data.

Narratives that only exist across sources

1. Bupa is in the answers but is not the source. Bowtie alone supplies 48 responses to Bupa's 29 — on questions about Bupa's own market. Access is proven open (live checks + existence of own-domain citations), so the gap is retrieval/content/indexation, not blocking, and not authority (DR doesn't explain why Bupa's own VHIS page loses to aggregators on Bupa-adjacent questions).
2. The VHIS convergence: the single most valuable topic fails in classic SERPs (pos 15, zero traffic) and AI retrieval (4 responses) simultaneously. One fix targets both.
3. The authority gap is a symptom, not the root cause. DR 48 vs 63–84 is real, but Bupa holds 417 keywords in positions 1–3 despite it. The failures concentrate in head commercial terms and AI source-citation — exactly where template/indexation/retrieval issues, not links, are the differentiators. The 95% link attrition is a cleanup/risk item (what happened?), not an acquisition brief.

Data gaps visible only in combination

• No crawl or indexation data for the VHIS comparison page (or any money page). Brand Radar shows it's under-cited; live checks never probed it; Ahrefs shows the keyword failure. Indexation status, crawl depth, internal links, and template health are all unknown — this is the week-1 verification.
• Live checks probed only 3 paths; the two 404s can't be confirmed as real pages without the sitemap/URL structure.
• Single-datacenter-IP caveat stands; server logs or GSC needed to confirm production crawler treatment and the migration timeline.
• Brand Radar tracks Bupa alone — no competitor SOV baseline; SOV 1.0 is an artefact.

Ranked priorities

1. GEO source-citation gap on VHIS comparison demand (Brand Radar + Ahrefs + live checks).
2. Onsite verification: indexation/crawlability of the VHIS comparison page and money templates; resolve /contact/ and /services/ 404s.
3. Commercial head terms: 自願醫保 pos 15, 醫療保險 pos 6 at ~8% CTR — same asset family as 1–2.
4. Link attrition: investigate cause; frame as risk cleanup. Not the headline.

Featured finding

Bupa's own VHIS comparison page is invisible as an AI source on the highest-demand questions in its market — and the door is technically open, so this is fixable on-site.

Brand Radar shows the top AI question in the category, 「香港自愿医保哪家好？」, generates 2,100 impressions in Google AI Overviews, with the VHIS cluster adding thousands more — yet Bupa's directly relevant page (bupa.com.hk/en/medical-insurance/vhis-info/voluntary-health-insurance-schemes-comparison-2026/) earns only 4 responses, while moneyhk101.com's equivalent page earns 29 and Bowtie's domain earns 48. Ahrefs shows the same failure in classic search: 「自願醫保」(10,000 volume) ranks 15th with zero traffic, and paid spend has risen 645% to cover the shortfall. Live checks confirm 0% AI-crawler blocking, and Brand Radar's 29 own-domain citations prove production crawlers reach the site — so the cause is not access and not links; it is that AI retrieval prefers third-party comparison pages over Bupa's own asset.

The mechanism: Bupa is mentioned in the answers (entity recognised) but the citation — and the click — routes to aggregators and a direct competitor, where Bupa's message is filtered and competitors are promoted alongside. This co-occurs with zero organic traffic on the head term and the paid-spend backfill.

The client can verify this in week 1: run the top five VHIS questions in Google AI Overviews and inspect whose pages are cited. What remains unknown — whether the page is indexed by AI crawlers, its crawl depth, internal linking, and content comprehensiveness — is exactly what a crawl plus GSC access resolves, and it is the first deliverable.

This matters more than anything else in the data because it sits at the intersection of the two largest demand pools (the 10,000-volume head term and the top AI question cluster), it explains the paid-cost trajectory, it is demonstrably not a backlink problem, and every month it stays unfixed, Bowtie and the aggregators capture demand that originates with Bupa's own brand.