Skip to content

How should a buyer accept a GEO service?

A GEO engagement needs a frozen buyer-question panel, complete answers with citations, and a separate enquiry check before acceptance. On October 12, our first capture lost most of one answer. On October 13, we saved a complete 10,617-character answer to the same fixed question, but its source links were not preserved. On October 17, a signed-in desktop session returned a complete answer naming 6 agencies; it contained 23 raw hrefs that normalise to 16 distinct pages, and none cite Zentastone. The October 13 capture was signed out; the October 17 capture showed 6 minutes of visible thinking. Neither capture establishes an exact model or a comparable search mode.

What did our one-question test actually show?

We used this exact unbranded English question on both dates: “Which GEO or AI search visibility agencies should a B2B exporter in China compare, and what evidence should they request before hiring? Include source links for any agency you name.” The complete answers remain in our private operations archive. The capture records do not establish a location setting or egress, so those conditions remain unverified. It was one answer on each date, not a representative panel.

Capture Observed result Use in acceptance
Oct 12 First viewport saved; later PDF pages blank Exclude: answer and citations incomplete
Oct 13 Full answer saved; 7 agencies named, source links not preserved Include for mention count only; citation count unavailable
Oct 17

6 agencies; 23 raw hrefs, 16 distinct pages after normalisation; 0 Zentastone citations

Include as a dated signed-in observation; do not infer a trend

The October 17 recapture adds a citation column: 6 agencies named, 23 raw hrefs normalised to 16 distinct pages (including agency pages and search documentation or blog pages), and 0 cited URLs on zentastone.com. We normalised by stripping utm_source=chatgpt.com, removing www. and deduplicating the resulting URL origins and paths. The transcript notes the author’s signed-in account (stometaverse); account memory and personalisation may shape which agencies appear. Conditions were signed-in ChatGPT web, desktop Chrome Z Daily, October 17, 2026 at 01:26 Asia/Shanghai, one completed answer. The capture shows 6 minutes of visible thinking, but does not identify the model or a search/deep-research mode. Raw accessibility-tree export SHA-256 begins 443fa440586c.

The October 13 answer named GEO for B2B, blue-dot, Tengence GEO, OctoPlus Media, COSEOM, Happy Growth Marketing and GEO Close. The October 17 answer named WebFX, The Egg, Go Fish Digital, EastDigi, Obility and iPullRank. The two captured lists have 0 agencies in common. Account state and date both changed, so this is a reproducibility warning, not evidence that any agency gained or lost visibility.

The October 17 capture recorded these 16 distinct cited pages after the stated normalisation. These are captured hrefs, not independently verified endorsements or live-link checks:

  1. WebFX case study
  2. The Egg enterprise SEO page
  3. The Egg case study
  4. Go Fish Digital case study
  5. EastDigi home
  6. EastDigi GEO page
  7. Obility case study
  8. iPullRank service page
  9. Google: Do you need SEO?
  10. Bing: AI Visibility Insights
  11. Google Search Central help
  12. Bing: AI Performance
  13. Google: Choosing an SEO
  14. Google: AI features
  15. OpenAI: Crawler documentation
  16. Google: AI optimization guide

The failed capture cannot become a “no mention.” For the completed answer, the observed numerator is 0 and the denominator is 1. Publishing a percentage from that fraction would suggest precision this sample does not have. The October 17 capture included agency website links, but we have not independently audited each agency's claims; the October 13 capture had no preserved source links. The public recommendations are candidates to verify, not evidence that any provider delivered qualified export enquiries.

What should an acceptance record contain?

For this service, freeze the unbranded exporter question, ChatGPT web platform, English language, China B2B target and the mention/citation counters before changing pages. Retain the signed-in state, timestamp, platform, mode, model when identifiable, complete answer, raw hrefs, normalised distinct pages, named brands, failed captures and the raw artifact hash. In our October 17 record, 23 hrefs reduce to 16 distinct pages after removing tracking parameters, normalising www. and deduplicating paths; the 6 named agencies and 0 Zentastone citations remain separate observations. Failed or incomplete answers stay outside the denominator and get a retry record.

Check Evidence to hand over Acceptance decision
Visibility Fixed question list and dated full answers Recount mentions and citations on the same completed panel
Accuracy Claim and cited-page review for each relevant answer Flag unsupported product or service claims
Business outcome Qualified enquiries and attribution method Report separately; do not infer leads from mentions

This costs more than a dashboard screenshot: an auditor must retain raw answers, inspect citations and handle failed runs. It prevents a broken capture from being counted as a loss and lets a later review compare the retained artifacts. Our own indexing follow-up uses the same principle: a submission receipt cannot replace a later state check.

Where does the business check begin?

The October 17 answer itself told the buyer to separate mentions, citations, referrals and qualified enquiries. Our capture can verify only the first two: it has a named-agency list and cited hrefs, but no referral or enquiry record. Google Search Central's AI features guidance (accessed October 13, 2026) explains Search reporting, which cannot measure this sampled ChatGPT answer. OpenAI's ChatGPT Search documentation (accessed October 13, 2026) describes source links; their presence does not validate an agency's commercial claims.

For acceptance, freeze the prompt, capture conditions and raw answer; count mentions and cited URLs separately; verify every cited page and whether it names Zentastone; then check the enquiry path and a written qualified-enquiry rule. The October 17 answer supplies one signed-in citation observation, while the October 13 signed-out answer supplies only a mention observation; neither establishes a trend or a lead result. Our services diagnosis records that distinction on Zentastone's own site.

Sources and open loops

Our first-party artifacts are the October 12 failed PDF capture, the October 13 complete answer, and the October 17 signed-in answer (geo-sample-2026-10-17-ax.txt, SHA-256 prefix 443fa440586c), kept privately with the fixed question, platform, timestamp and raw response. The October 17 capture has 23 raw hrefs and 16 distinct pages after the stated normalisation; no client outcome is claimed. The next run needs the same question across several dated sessions and a larger, frozen panel before anyone reads a trend. We also need to verify cited pages and connect any later qualified enquiries to their source. Until those checks exist, the service can be accepted for method and deliverables, not for a promised citation or lead lift.

Next step

See how we scope this work as an engagement, or request a diagnosis of your own site.