Jemeen · Melbourne, Australia

Jemeen’s AI visibility case study: issues, fixes and evidence

We are using Jemeen itself to test our approach to AI visibility. This case study follows the issues we encountered, the changes we made and the evidence for each result. The objective is to make useful Jemeen material discoverable when people ask AI systems how to get started with AI in a small business.

Why we are doing it

We want to use our own methods in our own business and show what happens. That means keeping the starting point, the work done and the business outcome separate. Publishing an article is a completed action; being cited in an independent answer is a measured outcome.

Our starting point

Our initial public HTTP checks found:

These are observations about website readiness. They do not establish whether Jemeen was indexed by every search engine, or how often any AI platform recommended it.

What we encountered and how we addressed it

Useful service information needed its own pages

Our homepage introduced the business, but did not answer the detailed questions a prospective customer might ask an AI assistant. We added a practical getting-started guide and a clear audit scope page, with visible links from the homepage. We verified that the published pages returned their content as ordinary HTML.

Crawl guidance and canonical information were missing

We added a robots.txt file, an XML sitemap, canonical URL metadata and consistent business information. Follow-up checks confirmed successful responses, valid sitemap XML and a crawl policy allowing the public pages. This corrected the missing website information; it did not establish search-engine indexing.

Our first URL setup sent conflicting signals

A deeper check exposed a mistake in our own implementation: the host redirected addresses ending in .html to clean URLs, while our canonical tags and sitemap still pointed to the old form. We aligned the page metadata, sitemap and guide links with the final URLs, and added a redirect-consistency check to our audit agent. This gives search systems a consistent preferred address for each page.

A successful page request did not prove AI retrieval worked

Our ordinary HTTP checks could read the pages, but initial direct retrieval tests in ChatGPT and Perplexity could not. We separated those tests so a successful browser or HTTP check could not be mistaken for a successful assistant fetch. We checked the crawl policy, page content and available hosting diagnostics, then tested the final clean URL directly.

Direct retrieval retests succeeded: ChatGPT returned the guide’s actual title and first numbered heading with a citation. After signed-in access was available, Perplexity, Claude, Kimi and Grok also returned the correct title and heading when given the guide’s URL. Perplexity automatically handled that request in its Computer mode, so that result is specific to that retrieval path. We cannot attribute recovery to a particular edit: the URL, account state, test time and provider conditions differed. A successful direct fetch does not establish organic discovery.

The host also inserted a Cloudflare browser-detection script into HTML responses. That is an observation, not proof that it caused the retrieval failures. We have not labelled it as a confirmed blocking rule.

We also ran an independent check using Google’s public inspection tool. It successfully crawled the guide and detected valid Article and Organization data. That confirmed access from Google’s test. A successful inspection is not confirmation that the page is indexed.

Public access did not mean Google had indexed the pages

After gaining access to the verified Search Console property, we inspected the homepage, getting-started guide, audit page and case study. Google’s stored records showed that none of these four URLs was indexed. The homepage had been crawled previously; the audit and case study were discovered but not indexed, while the guide was initially unknown to Google.

We submitted the sitemap once. Its first report showed a fetch error, but Google’s live inspection successfully retrieved it. A later check showed Success, with six pages discovered, without another submission or a website change. The initial error’s cause remains unknown. We also requested indexing for the four key pages, and Google confirmed that all four entered its priority crawl queue. At that stage, discovery and accepted requests were verified; completed indexing and search visibility still needed to be measured.

The visibility baseline needed a clear denominator

The first sample contained four unbranded answers across ChatGPT and Perplexity. A separate signed-in check then used two unbranded questions in each of Perplexity, Claude, Kimi and Grok: how to start using AI in a business, and who can help a Melbourne small business find practical AI opportunities. None of those eight answers mentioned Jemeen. These are small samples under different settings, not a controlled before-and-after comparison or a full-platform estimate. Missing tests remain unmeasured.

Finding the name and describing the offer were separate problems

In that earlier round, we also asked what Jemeen is and how it helps small businesses use AI. Perplexity and Claude identified the business and cited the Jemeen homepage. Kimi did not identify the business; Grok found a possible legal-entity match but did not connect it to Jemeen’s AI services. Both could read the guide when given its URL. This separates a brand-discovery gap from a page-access failure.

Recognition also needed an accuracy check. Claude’s answer gave the audit a three-business-day turnaround; the published audit scope says five business days after complete inputs. Perplexity described the agent walkthrough without its availability qualification. We made the audit timing and chat-integration status more prominent in our content and retained these errors in our evidence. A corrected source is a completed action; correction in future assistant answers still needs a fresh test.

What our fresh checks showed

Google now indexes all four key pages

We reinspected the same homepage, getting-started guide, audit page and case study in the verified Google Search Console property. All four were indexed. The earlier inspection had shown none of the four indexed. Google’s sitemap report also showed Success. We retained the actual crawl and canonical records and did not submit the pages again. These results do not establish that a particular edit caused indexing or that Google has processed the latest content revision.

No Jemeen mention in eight completed unbranded answers

In this fresh core check, we asked two unbranded questions in separate conversations across ChatGPT, Perplexity, Kimi and Grok. None of those eight answers mentioned Jemeen. This is a small observed sample, with model, account and search settings recorded. It is not a full-platform visibility estimate.

When asked what Jemeen is, Perplexity, Kimi and Grok correctly identified the business and cited its website; ChatGPT said it could not verify the business. Kimi and Grok’s successful name identification differs from their earlier failed matches, but changed settings and provider conditions mean we cannot attribute that difference to one website edit. That browser pass completed 16 of 20 planned answers. Claude’s four questions could not be completed because its browser session required sign-in and the authentication flow did not finish. They were unmeasured at that stage. We subsequently used the signed-in desktop app, with the different conditions recorded below.

Correct recognition still needed an accuracy check

Perplexity, Kimi and Grok correctly described the audit’s five-business-day report turnaround and said live implementation is separately scoped. However, Grok labelled old homepage dollar amounts as Australian prices, and Perplexity repeated those amounts without a currency. We checked the raw homepage against the live catalogue and found legacy price fallback text. The live catalogue returned the current AUD prices, while the page’s initial HTML exposed the older fallback amounts. We aligned those static fallback labels with the live AUD catalogue and retained the existing pricing script. The rendered homepage already showed the current AUD amounts. Source consistency is the repair; correction in future assistant answers still needs a fresh accuracy retest.

We also distinguished the authenticated Get Found workspace from public service information. A customer workspace is not a page to add to the public sitemap for search visibility. We published a separate Get Found on AI scope page explaining eligibility, approval, implementation and measurement using the current service contract. We linked it from the homepage and added its public URL to the sitemap. We also aligned the business’s verified company-profile links in the public structured data and qualified walkthrough availability against the service’s current release conditions. These changes improve the source information; they do not establish new indexing or assistant citations.

We have verified indexing progress and different branded-identification results in these samples. We have not verified an increase in unbranded recommendations, attributed enquiries, sales or revenue.

What the Claude desktop check added

We used fresh incognito conversations in the signed-in Claude Pro desktop app, with Opus 5.5 Medium and web search enabled. Neither of its two unbranded answers mentioned Jemeen. The broad getting-started answer gave practical workflow advice without a visible search invocation. The Melbourne help-selection answer searched the web and named government-backed programmes, the Victorian Chamber, a university programme and mentoring options.

When asked about Jemeen by name, Claude correctly identified the business, cited its homepage and returned the current AUD base prices. It also noted that much of the available information was the company’s own description. That is an assistant’s assessment of the sources it used, not independent verification of the business or a finding about its age. The answer did not qualify the current availability of every offering.

The original audit-scope question stalled on a request to connect Google Drive. We did not connect it or provide private files, and that attempt remains interrupted and unmeasured. A separate fresh conversation repeated the question with an explicit public-web-only instruction. That follow-up returned the audit scope, five business days after complete inputs, conditional walkthrough access and separately scoped live implementation, with citations to the audit page. It used the current AUD prices while flagging older amounts in a search result. We saved the completed answer and its cited sources.

The public-web-only instruction changed the prompt, so we report that answer separately rather than calling it a matched pass of the original question. The desktop check adds two completed unbranded observations to the earlier eight browser observations, all without a Jemeen mention. The interfaces, account settings and collection times differ; this is descriptive evidence, not a controlled before-and-after result or a full-platform estimate.

What we changed to address the discovery gap

We strengthened the getting-started guide with a reusable first-pilot brief and a direct answer about choosing practical AI help in Melbourne. The guide separates learning to use tools, assessing workflow opportunities and implementing an agreed pilot. We also added a customer-question brief to the Get Found scope page. These are useful content changes that we can verify publicly. Any later indexing, citation, recommendation or enquiry needs its own evidence.

We added the existing MirrorMirror Assistant example and its product and store links beside Jemeen’s Melbourne help option in the guide. This makes the company’s development statement and the store’s publisher attribution available in that section. Any effect of this content change on unbranded recommendations remains unmeasured. We also removed a stale last-updated value from the guide’s structured data while preserving its original publication date.

Distribution and URL submission are separate from discovery results

We published the practical pilot-brief guide through Jemeen’s company LinkedIn post. That confirms distribution of the guide; the publication receipt does not measure readership, assistant citations, enquiries or sales.

We also submitted five public website URLs to IndexNow in one request. The service returned HTTP 202: the URLs were received and key validation was pending. This response does not confirm indexing, a ranking change or an AI recommendation. We retained the response and did not submit the same request again.

Customer evidence and directory nomination

Jemeen developed MirrorMirror Assistant for MetaliCoin. The product website describes its AI skincare and makeup routines, and the Google Play listing identifies MetaliCoin as the publisher. The development attribution is Jemeen’s company statement.

Our directory nomination exposed an unmarked required-logo check. The browser’s file-picker integration then failed before attaching the file. We used a manual logo attachment to clear that obstacle and completed the reviewed nomination. The National AI Centre directory acknowledged receipt for review.

This is a completed nomination step. A live directory listing, government endorsement, improved indexing and additional unbranded AI recommendations have not been established by that acknowledgement.

What the process taught us

Website readiness, search indexing, assistant retrieval and an organic recommendation are different checks. Completing one does not prove the next. Our workflow now records the issue, the action taken, the evidence and the remaining uncertainty before we call anything a success.

How we will measure visibility

We use a fixed set of questions across ChatGPT, Perplexity, Claude, Grok and Kimi. The main group does not mention Jemeen, including “How can I get started using AI in my business?” A separate group asks about Jemeen by name to check factual accuracy. We do not combine the two groups into one discovery score.

For each completed test, the record includes the date, exact question, interface, visible model label, search setting, answer and cited sources. Repeated tests use fresh conversations. Missing or blocked tests stay marked as unmeasured. A result from an API is not presented as a result from that provider’s consumer application.

What counts as a result?

We will report the number of completed tests alongside each rate, preserve the underlying evidence and note changes to settings or questions. Search answers vary by time, geography, model and personalisation. A before-and-after change alone does not prove that a particular edit caused it.

Verified progress and remaining work

The four key pages are indexed in Google. None of eight completed browser discovery answers or two later Claude desktop discovery answers mentioned Jemeen. Perplexity, Kimi, Grok and the Claude desktop app identified the business by name; ChatGPT could not verify it in the sampled answer. Claude’s separate public-web-only audit answer used the current prices and described timing, walkthrough access and implementation correctly, while the original audit question remains interrupted and unmeasured. We repaired the source information and expanded the practical guide, but broader repeated discovery and other platforms’ post-repair accuracy remain open. No increase in unbranded recommendations, attributed enquiries, sales or revenue is claimed.

What the platforms say

Google’s guidance describes the importance of accessible, useful web content for its AI search features. OpenAI, Perplexity and Anthropic document their crawlers. Allowing retrieval makes content available for consideration; it does not guarantee a citation or recommendation.

Jemeen Pty Ltd is based in Melbourne, Australia. Jeremy Samuel is its founder and Executive Director. If you are beginning your own AI pilot, start with one workflow and a measurable baseline.