If you are paying for a website and SEO/GEO, one screenshot of an AI recommendation cannot tell you whether the work is progressing. Was your company named in the question? Did someone provide its website or keep prompting until it appeared? Would a customer asking in another language see the same information?
For Vancouver and Metro Vancouver businesses, an agreed test method makes the report more useful. Keep evidence of the question, answer, supporting sources and subsequent customer actions. Each answers a different business question.
Separate clicks from business outcomes
Pew’s study of 900 U.S. adults’ March 2025 browsing found traditional-result clicks on 8% of Google visits with an AI summary, versus 15% without. Search results were reconstructed on April 7–17, 2025, so they cannot be confirmed as identical to what participants saw in March. This was a sample association, not a Vancouver traffic forecast.
Adobe’s April 2026 report found AI-referred traffic to U.S. retail sites converted to purchases 42% better than non-AI traffic in March 2026. Its retail visit and transaction data cannot promise the same result for local service businesses.
Different populations and measurements cannot establish that fewer clicks automatically produce more sales. Your audit needs your own evidence.
Fix the questions and run two separate tests
Choose one priority service. Remove identifying information from inquiry records before using them to identify buyer questions. These teaching examples are not customer quotations:
- Find a service: “Which Richmond companies build English and Chinese business websites?”
- Check fit: “My website’s service information is outdated. What kind of team should I hire?”
- Compare evidence: “What public projects should I check when choosing a local web company?”
- Prepare to act: “What should I prepare before a website project consultation?”
The first two explore discovery. The others may return advice rather than supplier names; a useful answer need not mention your company.
Run the unbranded questions without your name, domain or a follow-up asking why you were excluded. Separately, start a fresh conversation and name the company to check its location, scope, languages and contact path. If you supply a URL, label that as a supplied-source check, not organic discovery.
Write English and Chinese versions around equivalent needs, preserving their exact wording and version numbers. Do not keep adjusting questions until they favour your brand. Add new questions transparently without replacing less flattering historical results.
Keep the conditions, including unsuccessful tests
A manageable starting point is four unbranded questions × two languages × two platforms: 16 discovery checks per round. Keep branded fact checks in a separate group, outside that denominator. Report each language/platform combination separately, with four checks in each. This is a suggested workflow, not an industry benchmark. Separate ChatGPT search from Google AI Mode; record when a conventional Google result does not trigger an AI Overview.
Repeat the panel on agreed dates, using fresh conversations. Record the actual testing location, the area named in the question, login status, visible model or mode, web-search setting and controllable memory settings. A fresh conversation reduces earlier context; it does not establish that personalization is absent. Naming Vancouver in a prompt is not evidence that a test was physically run there.
Preserve the complete question, answer, links and timestamp. Keep errors, unavailable features and incomplete runs in the log rather than quietly deleting them from the denominator. Never enter identifiable customer inquiries, patient records or private information for these tests.
Distinguish absence, factual errors and missing sources
Score separate observations: was the company mentioned, was it presented as a possible provider, were its facts correct, and were sources linked? A mention may not be a recommendation. Accurate information is not proof that your website was cited.
For an incorrect address, service or booking link, pair the exact erroneous statement with the current approved page. Open the cited source: does it actually support the claim, describe the right business and remain current? Without a traceable link, record “source unverifiable” instead of guessing where the answer originated.
The published Artemis website project offers a practical inspection example: its service pages, practitioner pages and Jane booking entry point are separate items to verify. This article has not run an AI recommendation test for Artemis and makes no claim about its rankings or booking growth.
A compact audit sheet to copy
Complete one record per response, then summarize by test group, language and platform. The sheet should show what needs fixing, who owns it and how the repair will be checked:
| Field | What to record |
|---|---|
| Question | ID, version, exact wording and language |
| Conditions | Time, platform/mode, web-search setting, actual location, requested area, login and memory settings |
| Test type | Unbranded discovery / branded fact check / supplied URL |
| Return status | Answer / no AI feature / error or incomplete |
| Company presence | Absent / mentioned / listed as a possible provider |
| Factual accuracy | Accurate / incorrect / partly accurate / unverifiable; include the error and approved reference |
| Source status | None / supports claim / does not support claim / outdated / wrong same-name business; include URL and verification date, noting anything unverifiable |
| Follow-through | Repair, owner, completion date and retest result |
“We planned 16 unbranded discovery checks, completed 14, observed three mentions and recorded two errors” is auditable when accompanied by the language/platform breakdown. It does not describe all Vancouver customers or establish AI market share. Before comparing rounds, confirm that questions, modes and denominators match. If conditions changed, show both groups rather than drawing an uninterrupted growth line.
Assign the repair, then measure the next step
Website inaccuracies need business-owner confirmation and a content or web owner to correct them. Directory errors need the relevant account owner. If a source is correct but an answer is wrong, preserve the evidence and review the retest. One missing mention does not establish a need for a website rebuild.
OpenAI’s publisher FAQ says content intended for ChatGPT summaries and snippets should not block OAI-SearchBot. It also describes utm_source=chatgpt.com on referral links. Have the website maintainer check access and measurement; do not expose sensitive pages to improve visibility.
Google’s guidance requires no special AI schema and does not guarantee display. Our existing Google AI search checklist covers website foundations. This audit concerns observation and acceptance, not purchasing another set of special tags.
Finally, report AI-referred visits, qualified inquiries and bookings or sales separately. Agree on qualification before counting: a genuine need, within service scope, with a usable reply path. Exclude tests, spam and duplicates. Lost parameters, device changes and later direct visits can leave attribution unknown. Do not assign every unexplained inquiry to AI.
Focus Media’s SEO/GEO services and website design and development connect information repairs with the customer action path. To review whether your current reporting supports the next decision, submit your website and one priority service through the AI growth assessment. The first-stage checks and any additional materials will be agreed during the conversation.
